 I'm not going to take much time, but I'll be a little bit here to reduce the topic, Vishal, and take it forward from there. There is lunch after the event. The kitchen is free to stay back and talk to people over here and enjoy lunch over here. Vishal does have a half-stop at 1 o'clock, so clearly he should sleep at the time and make sure that he doesn't have the banger tradition of everything. This is for your livestream. Okay, it's really weird to be wearing a microphone when there's no speaker in this room. Is it on mute? Does it work? Does it work? Okay. What about that one? Just make sure it's working as well. Just press the mute button on top. Okay. So, people have been to Hasgeek before. Who's not experienced Hasgeek before? It's your first time. Okay, so just give a brief introduction. Hasgeek is a company that gets people in technology to talk about the work they're doing. It's peer review for practitioners. So typically a peer-reviewed publication is meant for scientific research. And what we do instead is to say that the practice of technology also needs review because often your theory of how you're supposed to do things differs from the practice of how you're actually doing things. So we try to get people in the tech industry to just talk about what they're doing and why they're doing it the way they're doing it for feedback from others who are doing similar things in the industry. And with the hope that by doing this, we lead to better practice of technology because otherwise we tend to get stuck in our organizational silos. That your only feedback comes from inside the organization because usually the technology work in an organization is not visible outside the organization. And that leads to bad quality technology developing and not being identified until it's too late to do anything about it. So that's roughly it. And one of the reasons Vishal is here today is because of something else that happened a few years ago. If you remember, we had something called Save the Internet in 2015 where there was this fairly massive grassroots campaign trying to get the telecom regulated to enforce net neutrality in this country. And one of the things that came up during the campaign was this question of saying, what is neutrality? Why does it matter? And if everybody says net neutrality, what does it actually mean? So I was part of the campaign on the campaigning side of the operation where I helped build a website and get a lot of people to participate in it. But getting attention from a regulator is one thing. Helping them make the right choice is an entirely different matter and that's where Vishal came in. So I had no idea who Vishal was until that point. But a bunch of us discovered that Vishal actually understood net neutrality and was able to define it in a manner that a regulator could understand. So it's not just about saying we need good things in life, but it's about saying that this is precisely what it means. So Vishal ended up convincing Try to enforce net neutrality regulations and now we have the world's strongest regulations. So a huge part of the credit goes to Vishal for that. The other thing that Vishal has done that you may be aware of is something called cricket info. If you've seen the cricket info website, it's been around since the mid 90s. It's a place where you can see cricket scores online. So for a lot of people growing up, we have been listening to cricket on radio or on TV if you can't actually be there in person. And when the internet became generally available in the 90s, you're looking for cricket scores on the internet. So one of the first places you could get that was from cricket info. Vishal was a co-founder of cricket info as well. He's now a professor at Columbia University in New York and works on networks. And I think from there you can take it forward. Okay. For the introduction, I don't know how much of that was deserved, but you guys did a great job in running the campaign. And now, yeah, Indians have the strongest net neutrality regulations in the world. What I'm going to talk about today is how we got involved, how our research started in this topic. This was around 2006-2007 when we started working on net neutrality. So this is joint work with my current student, Nilufer, who's still working on the aspects of it. My former student, Richard Ma, and my collaborators, Dhamming and John, and Dan Rubinstein. Dan is also the co-advisor of Nilu and Richard. So I should give full credit of this work to Richard. He was the one who pushed us in this direction. I had no idea of this whole area. But he was the one who was really driven and he had the right ideas. I'm just here delivering what he worked on. So this is a conversation that happened between a very prominent economist. At that time, he was at UC Berkeley. And Dave Clark is one of the foundational architects of the internet. He designed most of the early protocols that were implemented on the internet. And if you take nothing away from this talk, I want you to take this conversation. This is really the gist of everything. So this happened around 2000. So the economist doesn't want his name publicly taken. He feels slightly embarrassed about it. And Dave and I said, this is fine, perfectly fine. So he'll remain anonymous. But this happened one evening, Dave and the economist they were talking. And the economist said the internet is about routing money. Routing packets are side effects. And then he told Dave that you guys really screwed up the money routing protocol. So Dave said, well, we did not design any money routing protocol. The economist said that's what I said. So the whole issue of network neutrality and these fights between telcos and content providers is because the internet was originally designed as a research network. But it was meant to connect research institutions, educational institutions. It was never designed to be the biggest driver of commerce in the world. But slowly that's what happened. And people had not thought through how these sort of commercial interactions will take place. And people still thought sort of, even when it became commercial, people still thought of bandwidth as sort of the unit or the unit of currency on the internet. And that's not really true. I'll try to make it clear as we go along. So this really is the gist of it. The internet was not designed for commerce. It became the biggest driver of commerce. And all the sort of the disputes that have happened because of this fact. So now, I'll tell you how we got started working on it. My area of research has been networking. And we were sort of analyzing this protocol called BitTorrent. I'm sure all of you are familiar with BitTorrent. BitTorrent is this protocol which is a peer-to-peer protocol. And from a technical or a research point of view, it's like a fantastic protocol. It scales really well. It has all the nice properties. But what had started happening in the US was ISP started dropping these BitTorrent packets based on port number. Then users started doing dynamic port selection. So they would run BitTorrent on all sorts of random ports. Then ISP started doing deep packet inspection to see if this is a BitTorrent packet or not, and they'll drop it. Then users started to encrypt the whole communication. Then ISP started doing behavioral analysis. They would look at packet sizes, inter-arrival times of packets, and figure out that this was started. That actually I would say is one of the first applications of machine learning for networking. They were trying to figure out just from these statistics of transfers whether this is BitTorrent or not. And then the first major case involving net neutrality was Comcast in Texas. There was this case where they were actually actively throttling BitTorrent traffic. They were just dropping those packets right away. They were not even allowing these connections to happen. And then somebody filed a case where you can't really do that. That was really the start of this net neutrality thing. And we were sort of traditional engineers or computer scientists trying to understand the behavior and the performance of the protocol. And we were sort of mystified why, why is all this happening. Then we sort of quickly realized that the problem was rooted in economics, not engineering. So then we thought if it's economics, what are the economists saying? Maybe they have answers to these questions. So we started looking at papers written by economists trying to understand the internet. And we saw a bunch of papers and they all said that the internet is like a two-sided market. One in the middle is the internet and on the two ends are the customers and the providers of content. And being sort of computer scientists or networking people, we knew that that was far away from the truth. Reality is a lot more complex. Internet is not one single entity. There are thousands of ISPs that connect to each other, talk to each other to make this internet happen. This is just another view of all the autonomous systems that make up the internet. Just collapsing it into one single entity didn't really make sense. There were some papers trying to analyze net neutrality written by economists, but they didn't really make sense to us. So what we started doing was we started developing our own model. And we said the conceptual internet platform has three kinds of players. In the middle are the ISPs. I'll talk about what kind of ISPs are there. Then you have the users or the eyeballs. And then you have the content providers that consume the content that is generated by the content providers, the eyeballs, consume that content. And then this whole net neutrality debate was happening. And actually the faculty, the Columbia Law School Timbukh, who really first came up with this concept of net neutrality. And we looked at the sort of folk definition of net neutrality. What did it mean? And it said that all packets should be treated equally. And again, as networking people, it didn't make sense to us because all packets in the network are not treated the same. But we sort of understood what the spirit of the definition was. So we thought, okay, let's go a little bit deeper into it and try to define what it actually should be. Move on from the folk definition to something more concrete. And what we realized that this failure to route the money, which I had spoken earlier, was sort of the key difficulty, which made it difficult to price packets based on their values. So before we get on to the details, if you think about the value of a packet. The value of a packet which carries some data, which is some frame of a Netflix movie, is very different from the packet that contains Google search results with ads. In terms of bits, they might be the same. But in terms of monetary value, they are really different. And all sort of the exchanges that were happening between ISPs was just trying to value these packets by bits. Just like everything that you buy, you just wait, whether it's gold or aloe. But there's a difference. And also, there are all these peering disputes that were happening amongst ISPs. There's this small list of the disputes and why the dispute was happening, what was the reason, what ended up happening. So there are lots of these fights happening. And we wanted to understand why these fights are happening. This is sort of going back to the genesis of this net neutrality debate. So one thing is that Netflix and YouTube, the video providers were the biggest traffic consumers on the internet and they were really causing most of these disputes. So we'll start with ISPs are sort of the building blocks of the internet. The internet comes together because these ISPs exist. And there are three types of ISPs that I'll consider in the fast part of this talk. One are the so-called eyeball ISPs. They are the ones that directly connect to consumers. So they are the one, your residential ISPs or the ISPs in your office. So users directly connect to these eyeball ISPs. Second kind of ISPs are these content ISPs. So they connect to content providers. And I include content delivery networks in this class of content ISPs. And the third class of ISPs are the transit ISPs. So these are the tier one global ISPs and they connect sort of these content and eyeball ISPs to each other. So now I'll get into this theory of cooperative games and how we use cooperative games to model this interaction and come up with some sort of understanding of the physics behind what's happening. So cooperative games are games. So this is different from sort of the non-cooperative game theory that a lot of us are familiar with because of NASH. This is cooperative game theory which was pioneered by Lloyd Shapley. He was a fellow grad student of NASH at the same time. In fact, if you read NASH's landmark paper in the acknowledgement section, he credits Lloyd Shapley for helping him sort of famous stuff. And Shapley also got the Nobel Prize in economics a few years ago. So in cooperative games what you have, you have a bunch of users. We call them players. So let's say they are N players. They get together in some sort of cooperation and create some value. And we also look at coalitions. Different coalitions generate different values. So Kiran and I, we can get together or form a business and it has some value B and the coalition is Kiran and me, then Kiran and me and I don't can get together form another coalition and create another value. So cooperative game theory analyzes this coalition formation given value allocation. So I'll explain what that is, but the solution of this cooperative game is a value allocation. So you can think of this solution as an n dimensional vector. So the total value that is generated by this cooperation, you know some business with that we create the value. How is allocated to the individual players? Everyone will get some share. So together that's an n dimensional vector. Given a solution, we are interested in the properties of a solution. One property is stability. Do players want to deviate from the solution or not? Do they want to be part of this coalition or do they want to leave the coalition? You know maybe some players are better off not being part of the coalition. They are putting in more than they are getting back. Second is related but different which is fairness. Which is, is the allocation in this n dimensional vector somehow reflective of the amount of contribution that these players have made in this value generation. So these are two important aspects of cooperative game theory. A subclass of these cooperative games are these so called convex games. The convex games, you know the sort of the mathematical definition is given in terms of the value of a set. But you can imagine sort of the fork definition of convex game is that the hole is bigger than the sum of the parts. And networks are sort of naturally convex. There are different sort of laws which give the value of a network. So Metcalf had this law which said that if you have a network of n nodes then its value is n squared. Or Lisco gave another law which says the value of n nodes the network is n log n. So regardless of what the precise actual form of the function may be there is general agreement that you know networks are convex games. You know the bigger the network the more is its value. Everybody agrees you know network effect is the number that we frequently use and understand. So you know you can believe that networks are convex so they are convex games. Internet can be thought of as a convex game. Now I will describe this notion of shapely value and core of a convex game. So as I said you know every solution of a game can be thought of as an n dimensional vector. So think of any solution of the game as a point in an n dimensional space. Some of these points lead to unstable solutions where individual players have no incentive to remain in the coalition and they leave the coalition. Now for every convex game there is always this thing called core which is set of stable solutions. So every point in that core the solution is stable. No player has an incentive to leave the coalition given that value allocation. The sort of centroid the center of gravity of this core is called the shapely value. So that you can think of as the most stable of the solution. So here you see the way I have drawn it the solutions which are on the edge which are stable. If you perturb them a little bit then they can become unstable. But since the shapely value is right in the middle of it it is the most stable of solutions. So that's sort of the geometric interpretation of the shapely value. So the solution basically is you know you get together n players you create some value. The solution is how is that value now distributed amongst the players. So mathematically think of it as if you have n players a solution is an n dimensional vector. Any arbitrary? The sum is equal to that. Sum should be equal or less. If you have some greater than that there is something wrong with the solution. But the sum should be equal to or less than that value. That's how you should interpret this. It's an n dimensional vector. So suppose the three of us get together we create a value 100. And 10, 10, it is a solution 33, 33, 34 is a solution, 0, 100, 0 is a solution. Any of them is a solution. 0, 100, 0 is clearly an unstable solution. 33, 33, 34 seems fair. But I'll come to exactly what's fair and what's stable. Does that come? So the stability of the shapely value I'll explain it in the context of a two dimensional game. So here you have two players. Player one individually can generate some value a. Player two individually can generate some value b. When they get together they generate c which is greater than a plus b. Now the question is how do we allocate c amongst the two players? So you can draw it in this sort of two dimensional plane. Any solution has to lie here on this line. That's the line which is a plus b equal to c. If you look at this red line, this is the core of this game. Because here player one, any point here player one is going to get more than a. Player two is going to get more than b. So they have an incentive to stay here. Any solution which is outside this one of the players has an incentive to leave. Because they individually will get more than what they were getting. So this red value is simply the central point. That's the most stable of solutions. This is in two dimension you can sort of imagine that this extends to n dimensions. So shapely in 71 had given this sort of geometric definition. But he had also given an axiomatic characterization of the solution. Where you don't have the sort of the mathematical or the geometric interpretation. But the solution is given in terms of some axioms that the solution should satisfy. So shapely had given this shapely value in terms of these four axioms. Efficiency, symmetry, dummy and additivity. Meyerson in 77 had efficiency in symmetry but replaced these two axioms by fairness. And Young in 85 again retained efficiency in symmetry and replaced fairness with strong monotonicity. Regardless of which sort of set of axiom that you use, the solution that you obtain. If a solution satisfies those axioms is the unique shapely value. Once you have a solution which satisfies everything, there's a single value that satisfies them and it's the shapely value. So what I'll do is I'll explain efficiency, symmetry and fairness. So I'll give the shapely value in the context of Meyerson definition. And again all of them received the Nobel Prize in economics a few years ago. Separately, independently. Something happened here. It's somehow not playing. Okay, so let me try to explain it without the animations. So efficiency simply says that for a solution to be efficient, the entire value should go to the players. So that the n-dimensional vector, it should sum up to exactly v. So that's an efficient solution. There's no loss of value in the distribution. Symmetry says identical players who have sort of identical contributions to the game should get the same value. So the green player should get the same as all of the green players. The red players should all get the same value. The blue players should get the same value. The balance contribution axiom, the fairness axiom that I'll try once if it plays otherwise somehow it's not playing. So let me try to explain the fairness or balance contribution axiom. That is defined pair-wise. It's defined for every pair of players. What it says is that suppose I take this blue player and this green player. If the blue player was not in the game, the green player would get some value. Let's say it was getting two units. Now when the blue player joins the game, then it'll contribute to the system and the total value of the system goes up. And the share of the green player will also now go up. So suppose it goes from 2 to 3. What it says that the blue player got green, an additional value of one unit. That means that the green player is also bringing a value of one unit to the blue player. So if now the green player leaves the game, whatever the blue player was getting, now it should get one less than what it was getting. So what you do for me, I do for you. That's really the balance contribution axiom. The value they contribute are they just different players? They're just different players. So what makes green signal to the second player? So whatever they are doing in the game might pair ending up. Okay, so two people are watching Netflix and three people are using Netflix. So a better example is that green are both eyeball ISP serving the same term. That's a better way to think about it. Or both greens are content providers providing the same sort of content. That's the way to think about it. So that's the balance contribution axiom that what you do for me, I do for you. Okay, so those two slides were a problem and those were the one sort of... But hopefully you got the idea of balance contribution. So now let's get back to our original problem of trying to understand how we should share profits. Now again, I should say that we're going to use the Shapley value as sort of an ideal to aim to. In the real world, you don't have some sort of politburo deciding who should get what, who should get what. In the real world, you sort of organically come up with shares. Our thinking here or our whole idea here is look at what the ideal Shapley value solution is and look at what the real world solution comes to and see if they are close. If they are close, then that means the system is stable. You know, you're naturally arriving at a spare and stable point. If they are far away, that means the system is unstable and we need to fix the system. So that's the idea. So now let's look at the baseline case. In the baseline case, we have one content provider ISP, one eyeball side ISP. The total profit they are generating, you know, let's, for simplicity, just think of the revenue. Let's say it's V. You can apply the axioms and you come up with that the content side share and the eyeball side share is going to be the same and it should be half V. This is a simple case. We have, you know, one eyeball side, one content side, we have a bunch of users, content provider and how should we share profit here. Now let's make it a little bit more interesting. Now we still have one content side ISP but we have two eyeball side ISP. We don't have additional users but we just have, you know, users now have a choice. They have two eyeball ISP that they can choose to connect to. So the axiomatic solution says symmetry, the same profit for the symmetric eyeball ISP, they're playing the same role. So 5B1 equals 5B2 equals 5B. Efficiency says the summation of individual ISP profits equals V. So 5C1 plus 2 5B equals V. Now what should be the share of 5C1 in this scenario? Earlier it was half V. Now when we have two eyeball ISPs, what should be the new share? Still the same. So 5C1 is half V and these two get quarter V each. That's one solution. Anyone has any other thoughts? So the profit of C1 becomes less than or more than? Less than. Less than. Okay. So the total profit is still V. So you're saying because there are two eyeball ISPs, the profit share of C should decrease. So it should be less than. The law of diminishing returns. Does that catch up? There's no sort of diminishing returns here. The total profit is still V. This is sort of, you can think this is a real-world example. We're just going to apply the sharpening value axiom here. So when you apply the balance contribution axiom, balance contribution axiom says what I do for you, you do for me. Now in this scenario, if the content side ISP goes away, then the whole game collapses. Nobody's going to get anything. So the profit share of B1 without the content side ISP is zero. Because there's no more game. So it's increment because of the content side ISP is 5B1, whatever 5B1 we arrive at. For the eyeball side, for the content side ISP, if there was only one, there's still a game and we had to look that that it share would be half B. Without that, now that we have additional, it share is going to be 5C1. So this is basically that balance contribution axiom for a specific case. It makes sense. So now if we solve it, the unique solution is the share of the content side actually increases to 2 thirds and the share of the individual eyeball side decreases to 1 over 6, 1 over 6. So the total that they are getting is 1 third. Earlier it was half and half, now it's 2 third, 1 third. Why is that? Because the other side, they are competing with each other. Here on the content side, it's a monopoly. Without the content side, the game is over. So it's a monopoly. So it has sort of veto power on the game. So its sort of leverage increases. Now you can sort of extend it to N symmetric eyeball ISPs and in the theorem that we proved, the Shapley value profit sharing is the eyeball side gets 1 over N times N plus 1 and the content side gets N over N plus 1 times N. So the more competition you have on the other side, the bigger is your share. And the more competition you have on your own side, the less is your share. This is sort of a natural thing that, you know, it's intuitive. All we have done here is sort of make it explicit, make it what the numbers are. Now what, you know, the single content ISP has increased leverage. This is sort of the marginal profit loss of one content ISP. If you already have N eyeball ISP then losing 1 doesn't really affect the content side on that one. It's the marginal losses minus 1 over N squared. Only when it becomes completely symmetric. You have only one content and one eyeball ISP. That's when you have the content that ISP is in. Otherwise it will hold all the time. Now you can extend it to N content side, the M content side ISPs and N eyeball size ISPs and you get the result of the shafted value share. Again, these sort of results follow the same thing. The more competition you have on your own side, the less is your share. The more competition is there on the other side, the more is your share. Or eyeball it. Please go back to that question that I see now. This one? Yeah. This is the balance contribution axiom. This is the value that I bring to you. You are bringing the same value to me. Now let's consider the content ISP and one eyeball ISP. What is the value that the eyeball ISP is bringing to the content ISP? If it was not there, the share was half P. When it's there, the share is 5C1. We have to figure out what 5C1 is. What is the contribution of B1 to C1? What is the contribution of C1 to B1? It's getting 5B1 right now. If C1 is not there, then there is no gain. If you think of it with respect to the new one that's being added. It's a fair one. But if you think of it with respect to the new one that's being added. Then they are getting half B1. Yes. But then with respect to that, the addition of B1 would now mean that shares decrease. Shares for B1 and B2 decrease. Correct. If you had considered that in the equation, you would get something different. No. This is the unique solution. You can work it out. This is with multiple eyeball and content ISPs. Now what you can do is you can throw in these transit ISPs into the topology. And the results that you get, the formulas are not hairier and uglier. But the basic behavior is the same. You want more competition on the other side of the network. Less competition on your own side to increase your share. So now let's go to... So I said this is work where we had done back in 2007, 2008. And the common ISP business practices were either you had two forms of bilateral settlements. So either you had a customer provider relationship where eyeball ISP would buy a bandwidth from a provider ISP and pay money to that provider ISP. Or you had the $0 pairing between transit ISPs where they would just exchange traffic. And you know traffic was roughly symmetric in both ways and don't pay any money to exchange traffic. So these are the forms of bilateral settlements. Either you had this customer provider relationship where you're buying connectivity or buying bandwidth from an upstream ISP or sort of these pairs just exchange without paying money here. With? This example. So well, so the content ISP let's say this is Akamai. It's paying a transit ISP to carry its traffic on the internet. You know Akamai in the US is paying let's say AT&T. AT&T is pairing with Tata here. And Tata, you know your last mile ISP, whoever your ISP is buying bandwidth from Tata. That's sort of the framework. We go back here because there's some problem happening with the animation. But as I said earlier what we want to do is you know the theory says this is the right way to do you know profit allocation, the Shapley value which will give you the most stable solution. In practice in reality what you have are these bilateral settlements. This is like a multilateral settlement where some oracle decides okay this is your rightful share. This is what you should get. These are bilateral settlements, you know these ISPs are interacting with each other. It's sort of a distributed solution. So what we want to compare is how far is this distributed solution from the theoretical ideal solution. If it's somewhere near then you know we are all set. Nothing more needs to be done. The system is stable and it works well. If they're far away then we have a problem. So the implications are and we sort of plugged in real world numbers and so on what things look like. So the implications are when the content side revenue is roughly the same as the eyeball side revenue then these two settlement customer provider relationship and $0 appearing can achieve a solution which is close to the Shapley value solution. So then things are fine. But if the content side revenue is much higher than the eyeball side revenue then what you need is a reverse customer provider relationship to make the system stable. There needs to be a net flow of money towards the eyeballs to make the system stable. So we predicted at that time that new settlements are needed to make the system stable. Another thing is if you have the level of competition on the two sides is different. If you have a lot of competition on the content side but no competition on the eyeball side or vice versa then also it makes the system unstable and you will have paid bearing. So we had made this prediction around 2008 or so and the internet had also evolved where a lot of commerce was happening on the internet or because of the internet but the money was not directly going to pay for the bandwidth. The example I said earlier, ISPs were sort of paying each other even if there was a customer provider relationship on by weight, not by value. So Google was paying the same amount as Netflix although the volumes were hugely different and Columbia was paying the same as any other commercial entity. There was a lot of value being generated outside the internet and payments were being made directly. They were not being paid for the bandwidth. So there was this huge asymmetry that was creeping in onto the internet. So Netflix and Comcast had this deal earlier where instead of having Comcast exchange traffic with Netflix on sort of these public bearing points, they had a private paid bearing arrangement where there was a private bearing between Netflix and Comcast and all traffic from Netflix to Comcast went through this private arrangement and a bunch of other ISPs followed suit. So this is a chart that Netflix publishes of the download speed by ISP. So here you see AT&T, Verizon, Comcast and there's one more, the fourth one, AT&T DSL, AT&T 5.5. So all of them what happened was the quality of download of Netflix started going down. What had happened was these ISPs were getting more and more customers more people were watching Netflix. So there was more volume of traffic going towards the ISPs. So the public bearing points, they were facing a lot of congestion. It was very easy for these ISPs to just add additional ports at the bearing points and relieve the congestion but they chose not to do it. Why not? Because all of these ISPs in the US, they are like monopolies. They don't really have competition for customers in their local geographical area. So the US is islands of monopolies connected to each other. So what these ISPs figured that they could just choose not to do anything until Comcast came and signed paid bearing deals with them because your customers of Comcast are not going to leave because Netflix is not working well. They have no option but to stay with Comcast. They have other options of streaming. So they could afford to sort of the customers kind of suffer until the paid bearing happened and as soon as the paid bearing happened everyone's performance started increasing. So what was really happening was now instead of the eyeball ISPs buying bandwidth from the upstream providers it was the upstream providers buying viewership from the eyeball ISPs. This viewership, these eyeballs were sort of captive because the level of competition was very low. So the competition model in the US is I would say it's broken because the competition is facilities based. So most US internet access is through wired broadband. In India most of it is through cellular but in the US it's mostly through wired broadband and the model in the US is it's facilities based in the sense that everyone has to dig their own last mile. So everyone has their own last mile to every customer and what that means is that the first mover has a huge advantage. So it's inefficient first of all because if I have multiple last miles dug to my home I'm only going to use one. The other three, all the investment that they have made is wasted. The model in the UK or Nordic countries is service based where the last mile actually is either publicly owned or owned by some third party. So any ISP that wants to give you service can just lease that last mile. So the level of competition there is much higher. The prices are much cheaper. You know the quality is much better because the ISPs are now competing with each other for customers. They have access to all the customers. Unlike the US model where the ISPs were using the customers as leverage against content providers. You know if you have this model then you know ISPs compete on service. So this is sort of natural variation. Are they not related at all? They're not related at all. You can see Google Piper going up and down. So you know as sort of Google got more and more customers they kept on increasing their capacity because they were not really competing in the ISPs. They were there just as a proof of concept born to disrupt the market. Everywhere that Google went, these other ISPs started off every different space. Before that they were not ready. So after the private trading, Google going down is just... It's just natural variation. It has nothing to do with it. So the FCC definition of net neutrality was there should be no blocking, throttling of paid prioritization or fast lanes of any content by ISPs. Now even this definition is gone. We had started with that definition of fork definition that all packets should be treated equally. But what we realized through our work that you know it's really about competition. Competition is the key here. So this definition was wrong because it did not consider the issue of pricing at all. So now I'll talk a little bit about zero rating and differential pricing. So zero rating, you know all of us know what zero rating is. It's a relationship between ISPs and content providers where ISPs will not charge their users for accessing specific websites or applications. Other websites or applications which are not part of the zero rating arrangement customers have to pay or their bandwidth at least is charged against the quota. So there are several examples of zero rating. Free basics, I'm sure everyone remembers internet.org in India. And just I guess a little bit old, 80 countries offer zero rating or types of services very notably India and Canada have banned it. This was the 2015 sort of regulation that were enacted in India. So the big issue with zero rating is this issue of consumer surplus. How many of you are aware of this notion of consumer surplus? So consumer surplus is this concept in economics where consumer surplus is the difference between what a consumer is willing to pay for a commodity and what the consumer has to pay. There's a difference between the utility provided by that commodity and the price that a user has to pay. So consumers naturally choose commodity that gives them the most surplus. So you could have a commodity which is priced at 50 or let's say it's priced at 50 and what you're willing to pay for it is 60. So the surplus that you get is 10. There could be another commodity which is priced at 80 higher than 50 but you're willing to pay 100 for it. So the surplus that you're getting is 20. So even though it's priced higher because the surplus in your mind is higher you'll go for that commodity. That's sort of this concept in economics. Consumers choose commodity that has the most surplus. So the willingness to pay is a property of the content. How good the content is. What are the ISPs doing it? If it's a video content are they streaming it at the right quality level and all that. And FCC's definition, no blocking, throttling or paid prioritization keeps this willingness intact. So willingness is what content the content provider produced and are the ISPs faithfully delivering it to the consumers or not. So ISPs should not change the willingness to pay. But FCC was silent on what the consumer has to pay. So zero rating distorts what the consumer has to pay and so distorts the consumer surplus and hence the market. Because if Amazon Prime and Netflix have the same utility for me in my mind they're providing the same service but one is zero rated and the other is not then the price that I have to pay changes. So I'm paying some price to Amazon, I'm paying some price to Netflix. On top of that I'm paying some price to the ISP for the bandwidth. If the ISP changes that price then they're changing what I have to pay. So that changes the consumer surplus. Once the consumer surplus is changed then it distorts the market. Then the ISP is dictating what content you should use. It's not letting you decide. It's still letting you decide but it's pushing you in one direction if it's doing zero rating for specific websites or specific applications. That's sort of the broader idea. So what happened in the real world so in the US in November of 2015 T-Mobile which is a big cellular operator they introduced this program called Bingeon. In Bingeon you had they had some partner sites like Netflix, Hulu, HBO which had videos which were zero rated and then there were non-partner sites where the videos were not zero rated like YouTube and other video providers and any video provider that they didn't know about. What they did was that all videos were throttled to 1.5 megabits per second regardless of whether you were a partner or a non-partner. So in that sense they were being fair to all videos in terms of quality or in terms of pricing they were not fair. Some were priceless, some were priced more. So there were two separate studies on the impact of Bingeon. One was done by T-Mobile and one was done by a consulting firm engaged by T-Mobile. And T-Mobile claimed that Bingeon benefits everybody according to both the studies. What did the studies reveal? So the consulting firm study said that partners showed an average viewing time increase of 50% but the viewership of the most prominent non-partner, YouTube, that also increased by 16%. So they're saying, you know, this benefits everybody. And this makes sense, right? If I'm using part of my data to watch Netflix and now that data becomes free what I was using to watch Netflix, then I can use that data to watch YouTube. So it makes sense that YouTube also increases. And T-Mobile said that, you know, their analysis was a little bit different but they said there are 79% benefit for partners and 33% benefit for non-partners. So you can claim that it benefits everybody. But if I'm in the business and I see my competitor grow by 79% and I'm growing by 33%, I'm not going to say it's fair. I'm saying that this has distorted the market. It's no longer a level-paying thing. So consumer surplus isn't just theory. These sort of large-scale experiments conducted by T-Mobile confirm that it happens in practice. When you do zero-rate something, consumers will move to that even though no change has been made on the quality of the content. So they did not change anything in the two, three months where this binge on was introduced. So whatever utility I had for YouTube before, I'm assuming I still have it after. Whatever utility I had for Netflix, I still have it after. The only thing that changed was how much I was paying for bandwidth for that. And that caused this distortion. Yeah, so the quality was not different. Outside the binge on, yeah, there would have been nothing. So you can say that the same sort of quality was same for both. And the content did not change. The selection of movies on YouTube versus Netflix did not change. I forget what, but it was a fairly large sample to give you a statistical confidence. I can look up those numbers, but fairly large numbers did not change. So what we have done, you know, this work Neelu has been doing is we know that consumer surplus sort of causes this distortion. So what you've done is we have come up with a model for this interaction between ISPs and content providers where first we define a model for each pair of ISP and content provider where the content provider has the option to zero rate its content or not. When it zero rates its content, it'll get additional viewers. It has to pay that money to the ISP for that bandwidth. And it'll always do it when the revenue gained from the users is more than what it has to pay the ISP. So we made this model of where we can analyze this interaction. Every ISP content provider pair individually makes this decision whether you should zero rate or not. And we came up with this theorem where we could compute the user or market share of different ISPs and content providers under this sort of scenario where they decide whether they want to do zero rating or not. And you know, they come up, they come to some sort of equilibrium, sort of like Nash equilibrium. But we also showed that in some scenarios, what happens is some content providers are forced to do zero rating. Even when they are paying more to the ISP, then the value that they're getting out of zero rating, the additional customer, so the total revenue actually falls. But they're forced to do zero rating because their competitor is doing zero rating. So just to keep up with the competitor, they are forced to do zero rating. And the competitor can afford to do zero rating because the competitor might be a richer content provider. So that is what we call zero rating pressure. Some content providers, their decision on an absolute basis, you know, my revenue is actually not going to increase if I do zero rating. But I'm forced to do it because if I don't do it, my revenue is going to fall even further. Then what we did was, we did this analysis of something called a Herfindal index. This Herfindal index is again a concept from economics and markets where this index indicates how competitive a market is. The ideal market, the Herfindal index goes towards zero. That means all is a very competitive market. There are lots of players providing competition to each other. If the Herfindal index goes to one, that means the market is going towards a monopoly. So if you have, you know, we can set up the system and we can analyze it with or without zero rating. We can get different market shares of the different ISPs and content providers. And then we can look at what the Herfindal index looks like and whether it causes this market distortion or not. So, you know, I'm going to skip most of the details here of, you know, how we did the modeling and, you know, what kind of decisions go into choosing whether an ISP should do zero rating or not. You know, the kind of questions the ISP has to look at to do zero rating is whether I'm zero rating with somebody else or not. What is the data price per user? How popular are the ISPs? How many users do I have? How popular are the content providers are? How many users does Netflix have? All sorts of things that go into the decision of the ISP. The content provider also has, you know, a bunch of questions which influence whether it should do the zero rating or not. So they will go through their own analysis, come at the decision whether they want to do zero rating. And what we did was this discrete choice modeling. There's an axiom that was developed by Luce, Luce's choice axiom, which says the probability of choosing some commodity I from a set S can be computed using this formula where WIs are the weights. So what we really, what we did was we extended this axiom where instead of this commodity I, you have a pair, idea that you pick. So you pick a pair of content provider and ISP. So the idea is you have a choice of ISPs, you have a choice of content providers, some of them have zero rating relationships, some of them not. What is my probability of picking one over the other? And the weights, you know, the weights that you have here that are computed by whether there's zero rating or not. What's the consumer surplus? How popular, you know, individual ISP and content provider are? So users can choose using the generalized Luce's axiom, we model the consumer surplus, we model the stickiness, you know, sometimes consumer surplus is not enough. Sometimes, you know, whether you're paying more or less for Netflix, you like the content, you're not going to shift. So we model all sorts of things. And, you know, eventually we come with this theorem which gives the market shares of individual ISP content provider pairs based on, you know, the parameters of the model. You plug in the parameters if zero rating is there. Theta ij equals one means there is a zero rating relationship between content provider i and ISPj. If theta ij equals zero, that means there's no zero rating. So you can plug in those numbers, you know, this sort of ugly looking formula, but it allows us to get these numbers. I'm just going to present the results rather than go through the sort of mathematical details here. Here's a good figure. So here we are looking at a monopolistic ISP. So this is a monopoly ISP and the X axis here is the per unit price of bandwidth. How much does the ISP charge? The Y axis here is the utility or the revenue of the content provider. The dashed line represent the content provider's utility or revenue when zero rating is available. The flat line is when it's not available. So when zero rating is not available, then let's say you have two content providers, one is doing better than the other, then their profit is not going to change with the ISP price. It doesn't matter what the ISP price is because they don't have to pay the ISP. It's the users that they are paying there. It's sort of the users problem. But if they decide to do zero rating, then they are paying the ISP and then you can see the behavior. So first, when the ISP is charging really low, the share of both the green and the blue, their profit increases. Why does it increase? Because the way we model it, there are a bunch of users who will not view the content unless it's zero rated. They don't want to spend their bandwidth on these content providers. But if it's become zero rated, they don't have to pay for it. Then they get these additional users. So initially, their sort of profit increases. Then if the ISP becomes more expensive, then they have to pay the ISP more. So their profit starts decreasing. Now you see here, the profit of the ISP has started decreasing. Of this content provider is below when it was not doing zero rating. But it's sort of forced to do zero rating because this one is doing zero rating. And now in this region, both of them are sort of have this zero rating pressure on each other. It's sort of like prisoner's dilemma where both of them are actually below what they were earning without zero rating. But because zero rating is available and the other one is doing, they're sort of forced into this zero rating pressure. One is that both C21 and CB2 have zero rating available. Second case, both C21 and CB2 do not have zero rating. Zero rating is not an option in the market. So you're not looking at the case that one of them is opted for CB rating and zero rating. So here there are scenarios where one of them. So here I forgot to say that. Star indicates that both of them are doing zero rating. So here I'm labeling what equilibrium the system comes to. Star indicates both of them are doing zero rating. This square is that the more profitable content provider is doing zero rating. The other one is not. Diamond is that the less profitable content provider is doing. And this dot is neither of them are doing zero rating. Oh, sorry. This is the more profitable. This is the less profitable. So what happens here is that initially, both of them do zero rating because their profit is increasing. But if the ISP becomes too expensive, then you know this guy, the less profitable content provider cannot afford it at this point. It's revenue. Now it's getting sort of negative. So it stops doing zero rating. When it stops doing zero rating, the other one's profit share increases. So it continues to do zero rating for a while. But at this point, even for this one, it's unaffordable. So it stops doing zero rating and then sort of the system comes back through. The equilibrium without zero rating. So it's a dynamic game? It's a dynamic decision making that's being made by each player, depending on what it knows of the other player. So it's sort of like the national equilibrium. But because it's sort of a pair-wise decision, it's slightly different. So we define it as zero rating equilibrium. Here is a complementary duopoly. So you have two ISPs, two content providers. And you know, the equilibrium that you reach, the game theory equilibrium. Again, we have labeled it using different. But the idea really, what you saw the phenomenon earlier, that same phenomenon kind of repeats. Now I'll show you what happens with the Harfin Dal Index. This is sort of, I would say, the punchline of this analysis. So the impact of zero rating, we have this Harfin Dal Index. So it is sort of the sum of the squares of market shares of all firms in the market. When it grows to one, the system moves from sort of a collaborative state to a monopolistic content provider state. And when it goes to zero, then it's sort of a perfect competition. So lack of combination causes market distortion and welfare loss due to monopoly. So here's what happens with the Harfin Dal Index. So Q1 and Q2, you can think of them as representing the profitability of a content provider. So one is sort of a rich content provider that can afford to do stuff. One is a poor content provider. So this is the Harfin Dal Index of the system when there is no zero rating. So, you know, values above zero, you see that it's sort of perfectly competitive. The market share is not dependent on the profitability, but the quality of the content. But as soon as you introduce zero rating, this Harfin Dal Index shows distortion, especially in regions where the difference between Q1 and Q2 is significant. So if there's imbalance in terms of the profitability of the content providers, Harfin Dal Index gets really distorted and goes, starts going towards one. This is where, you know, the sort of the richer player, just through its dominance in terms of money, can kill the poorer player. So there, you know, the statement, that sort of the intuitive statement that zero rating affects consumer surplus, which affects market share, which sort of distorts the market. What we did here was develop the formal model and showed it explicitly that, you know, this is what happens. You had evidence of it from the team of a study and there are also studies in South Africa and all where WhatsApp was zero rated and suddenly WhatsApp share just shot up. So zero rating, you know, causes this market distortion and it affects the equilibrium. So this is sort of the conclusion of the work and, you know, redefining network neutrality from the original definition of all packets should be treated equally. This is a definition that the internet should provide a platform that does not provide a competitive advantage to specific content apps or services, either through pricing or quality of service. The quality of service is roughly, you know, all packets should be treated equally. Pricing says then they should also be priced similarly. For a specific content class, if you price them differently amongst content providers and that distorts market and that's not providing a neutral platform. So this is a definition that, you know, saved the internet campaign went with. This is a definition eventually that TRA I also implemented in its regulation. So before I end, let's see what has happened post that differential pricing ruling in India, what has happened in the last three years. So the internet penetration has accelerated broadband speeds have improved. India now has the cheapest data prices anywhere in the world. People are saying that, you know, without free basics, no one will pick up on the internet and, you know, it will be very expensive for users. Now it's not all because of this differential pricing. You can say maybe even the primary reason is due to Geo. Geo came in but these differential pricing ruling meant that when Geo started offering free data or cheap data, it was data for all of the internet. They gave free data, you can use it anywhere on the internet and, you know, the competitors had to follow suit. They had to lower their prices and, you know, it was beneficial both for the content providers and the users. So all sorts of content providers, Netflix, Amazon Prime, Hostile, all of them sort of grew. Now people don't even think about streaming stuff on the phone because, you know, you have such high daily limits at such low rates. But if these rules were not there, then, you know, Geo would not have given free data for the entire internet. They would have said, okay, use this Geo data for Geo movies, Geo TV, Geo Music. Airtel would have offered their own Airtel versions, you know, Airtel Wink, you can use it for Airtel movies. What would have happened, what you would have ended up with a very vulcanized sort of confusing marketplace for competitors. The price pressure would not have fallen. All of these ISPs would, you know, either would have become content providers and give free data only for their own content or, you know, they would have tie-ups with other specific content providers. So for consumers, it would be bad because now you would have to make a choice of ISP and content provider. But because of this ex ante regulation which stopped it ahead of time before it came into the market, it has resulted in a much better and a much vibrant, much more vibrant internet marketplace in India. So I think the campaign did good and the regulation that was passed really did help the whole internet scene in India. So now I will stop here and take questions. So the definition is it's some of the squares of the market shares of all forms in the market. So, you know, I've shown one theorem which allowed us to explicitly calculate the market share with or without zero rating and, you know, all sorts of parameters in the model. So we can do that to compute the market share and then get the profit. So here, in this case, the content providers. Yeah, content providers. Yeah. So India, the situation is slightly different in terms that most of the broadband access is through mobile. And, you know, in India also it's facilities. In a sense that every provider has their own cell towers, they have their own spectrum and everything. You know, a better model would be if they had facilities sharing. Even spectrum sharing, you know, instead of doing auctions, maybe a better model would be that locally they can do auctions. In a sense, suppose here in the Nagar, GEO has bought all spectrum, but it doesn't have users. Whereas Airtel has a lot of users, but it doesn't have spectrum. You know, you should have the ability that Airtel can lease that spectrum from zero and offer it to them. It's happening. So that's good. What, is it dynamic or? No. It's a tower. It's a tower there. So Airtel doesn't have a lease in essence. So that part, so it's going to be about that. But you know, you should just share towers. Yeah. That's full time. I think it was too peaceful to understand how it would be that way. And I think my question was a bit of a meta-alumni. So if we were to use this framework of competition, and considering the value of packets to be equal to the commercial monetary value of the packets, then don't we kind of make networks to become police and serial to market culture? So my question is a bit of, do you think, do you think they might be meddling, trying to borrow from outside of this rainbow? Because I think even in the US right now, I understand a lot of neck-and-neck movements looking at, for example, making connection with the, how racial justice is connected to net neutrality, for example. I guess my point is not just like the, I think there's a value there. There's an economic value that is aside from the commercial value of packets. So how do we account for that within these definitions and if it's useful? Especially the time where network architecture is increasingly defined by what then-form lobbies and companies want, I guess, I was like, try once to look at machine learning for themselves, which is, I don't know, I don't understand what you mean. So I wonder if you see- So right off the bat, I would say that racial justice is beyond my favorite. I cannot- But the idea here is that if you have these net neutrality rules in place and you think of internet as, you know, a platform where different content providers get to compete based on ideas, not based on sort of pricing innovation, then that's best for competition and that eventually is best for users. So that, you know, the more competition you have, the more is the downward pressure on prices. The more it's better, the better it is for users. Both in terms of how much they have to pay and how much innovation is driving in the content provider market. You know, if I have competition, then I have to innovate to, you know, keep my customers. I think we are basically the same as well. I am just wondering if you could sort of think outside of the framework of competition theory. Again, I said I'm sure it is, but it's sort of beyond my 11 options. I understand. I had a few questions about the implications of the redefinition. The quality of service aspect. Do you think that definition that allows private by actually 1xp 1xp? So good question. So, you know, private pairing arrangements does affect your quality. So all these net neutrality regulations have been dealing with the last mile. Then the last mile, you should not speed up content and last mile, you should not discriminate. But private pairing does affect, and you know, we should be thinking about how we should extend this idea and can it really be extended to private or feed pairing between us? It does increase the quality of service. It does increase the quality of service. But you know, all these regulations have not gone into the pairing world. Yeah. Could be a better solution, yeah. Similarly, edge caches won't be allowed in this country because they sort of improve the quality of service. So this is actually more radical than the recognition of the country? Yeah, if you go beyond the last mile. But law and argument will be made that everybody has access to these edge caches. Okay, one more question regarding how you analyze private pairing arrangements. Yeah. Were these like contract agreements available? No, but you know where so we did not have the details of what the contracts is, how much they are paying for the private pairing. All we know is that they exist. It's like a binary in our formulation, whether they exist or not. And if they exist, they are sort of they are driving the solution in a certain direction. That's it. Do you know if there's any way to gather this data? Private pairing? I think they are sort of private business agreements. I'm not sure if the RAI has the authority to look into it and regulate it somehow, but I'm not sure. The model assumes throughout the competitors are equally capable? Not necessarily. Okay, the model accommodates. accommodates. For instance, the last one that I showed, Q1, Q2, the different prices that the content providers are charging, that's because they are able to command these different prices. Maybe it's because the quality of the content is better or they just have market dominance, they are an established player. So the model does accommodate. So what does the model change? Let's say if the content provider, right, there's a heavy monopoly on the content provider side compared to not so much of a monopoly on the high-quality aspects. It does. We can plug in those different numbers. But it's still the same percent. Zero rating. Zero rating, yeah. In fact, in that scenario, it's even worse because that monopoly content provider can quickly drive out the sort of new entrant through share money power. So this is regarding the model sharing the local last mile sharing with our provider. But isn't that also deserts bigger ISPs like in India we had an issue where GEO and ATL both are disconnecting their service. For example, in Tamil Nadu, ATL owns larger shares in Kerala, GEO like that. The quality of service between them, just because they're competitors, they try to dig the others in a bad way. But by sharing, giving that kind of sharing model, we allow the ISPs kind of dictate to the users that if you use this service in this area, you won't get good coverage of that. Instead of that, if you go for a public sharing what you can do that, that may be a better way for India to go where last mile it's always a public. Last mile being public is always the better option. If you can't do that, at least have this in the US, there was something called local loop unbundling. Basically, if I'm a private ISP, I have my last mile. By regulation, I have to lease my last mile to competitor if they want. That's called local loop unbundling. That used to be the case when you had dial-up access, but once broadband came, but that's sort of the next best option. The best option is really the public option where the last mile is a public utility, and ISPs can have access to it, because they are providing a service. So that, in case, is a better model then? That's a better one. It's been more than a year or so since, I think, that it's been suspended in the US. Well, the halfway definition that they have, even that they have gotten it. So what you see is that there's a lot of vertical integration happening in the US. These large ISPs are buying content providers. Verizon has bought Yahoo and others. You see a lot of AT&T has bought Dish. So you see a lot of vertical integration happening and I'm sure zero rating is standing. Zero rate their own content. So it has to be played out. Even when that happens, I said, consumers will not see any impact immediately, and it will affect the competitiveness of the market as time goes on. You may see a photo of this. Netflix is using your architect, and it's not going to be this way. Yeah, yeah, yeah, because they want to. And then modern-school enterprises don't make this happen. So if you don't have competition, then the government has to impose this. If you have a very competitive market, then you really don't need these regulations. A competitive market where users can quickly or easily switch between providers and you need some regulations to align the incentives for the consumers. That's sort of the stereotype, but I would say TRA has an excellent job in terms of this. It's to the government to define how it is here. You have to From the government's perspective, if they don't interfere the market will be in one of this. From the government's perspective, they are the people whose market is here. Going back to that 2014 in India where the pre-basics of corporate was more targeted to the people in the village sites, in the rural sites, where maybe they don't have much money to pay it and view it all the stops. So, I mean, just just thinking if you didn't enter, what would have been the India case as of now? It still would have been much better than if pre-basics was allowed to come anywhere. It sounded good to think of the high-level mission but if you looked into the details it was fairly nefarious. If you didn't enter, today, how people think maybe we allowed not going with people for 60 years, things would have been different. Just thinking in that point, majority would have moved in that saying that I think we need to go now. Pre-basics exists in a lot of places in the world and none of those places those advertised benefits have been observed. There are lots of privacy and security implications with pre-basics. I can talk to you offline. I was just thinking that way we would have been forced to move because majority would have seen there's a benefit that way because I want the same way as what you do that we would just start going in that way. What are the big problems with pre-basics was that all the traffic was posted on free-basics. First went to Facebook servers it got decrypted there and then went to those sites. So do you think in this scenario any other social network or Google or Twitter would become part of pre-basics? No. Then they would have to do their own sort of free-basics arrangement and then again you have this vulcanized kind of market. This is not good for anyone. Clearly a play to get user data was not some charity in the world. Even in the years we still have struggled in India to tell everyone how much privacy is important because I mean not being in the world but other things where we try to tell that the price is important but still the method is still not understanding that everyone is asking to authenticate in different means which is not needed for that. So I mean people still thought privacy is important so I can't support free-basics. This is a good happen there. I was just telling you if that did happen it would have been forced into problematic and so forth. In almost all I kind of I'm not sure. I was just reading through all our kind of not much developed countries the concept of privacy is not that the people concern I'm the majority. But that I would say is a separate talk. Thank you everyone.