 The idea of, most of what I do is sort of connected to this idea of machine learning for artists. So I'm interested in applications of machine learning for art, and it's something that I'm not going to talk about very much today, but it's kind of tangentially connected, and I'll kind of make a few interweave it a little bit towards the end. What this talk is going to be mostly about is this notion of decentralized artificial intelligence, or decentralized AI, which is something I've been thinking and writing about a lot recently. For me, I've been really invested in machine learning for a long time, and decentralization stuff is a lot newer for me, but I'm quite excited and also nervous about the intersection of these two things. There's a little bit of like a sort of double buzzword syndrome happening here, where the two things are like very sort of hot topics. And so it can be very hard to, if you're interested in how these things intersect, they'd be very hard to disentangle things that are actually interesting about it from this sort of like hype train that you see in both of them. So I'm going to do my best to kind of like try to focus on the old things. So let's get into it. So in some sense, this is kind of two topics in one, and I'm going to explore their intersection towards the end. I'm going to talk mostly about like kind of start out talking about machine learning and then talk about what happens when you add this notion of decentralization to it. And just to kind of be clear about a little bit like what I'm specifically talking about. By decentralization, I mean technologies like blockchains and cryptocurrencies and tokens and dowels and all this kind of stuff that we see a lot of these days and they're trying to disrupt this and that of the industry that it can. And then on the other side, you have machine learning or AI, deep learning and applications that it has through natural language processing and computer vision and so on. And these two things, I'm going to talk about like why they're important to each other. So that's actually the next slide. First of all, like why is this decentralized, why is the notion of decentralization relevant to AI? So for the most part, if you read like popular media and press and things like that about about these learning and machine learning, you will mostly get a lot of news about like things that sort of interface with human beings, right? So robots and chatbots and you know, recommender systems and self-driving cars and you know, Siri and Alexa and Katana and all this kind of stuff, right? But and this soaks up all of the press with AI and it should make sense why, right? Because it's the thing that we find most exciting. Like people want to know what AIs look like in front of them, how they interact with them, right? But there's kind of a problem with this number that it doesn't necessarily focus on those applications of machine learning which are actually very, very, very, very close, you know, very, very now happening now that are realistic. I mean, of course, like we see basic, basic versions of these already in production but they are kind of really, really simple versions of it, right? So like, you know, Siri and Alexa have very limited intelligence. They can tell you the weather but they can't, you know, like they can't necessarily do very much for your day-to-day activities, right? They're just kind of like advanced research, right? And the reason for that is twofold. One is that there's this, before you do, before it does any of the actual interesting work, it has to first understand what you'll, it has to first understand what you're saying to it or how you're interacting with it and that, you know, what we call human-computer interaction. And that's really actually quite difficult because humans are very, very inconsistent and capricious. We have many, many languages. We have slang and idioms and, you know, culture is very, very sort of complicated and hard for computers to understand. What computers understand much more readily is our computer protocols, right? Things that are very rigid and consistent and work the same way all the time. And so for that reason, applications of machine learning that are relevant to computer-to-computer interaction are actually much more close. And a lot of them, of course, are already in, are already actually in production. So if any of you use social media, assume that most of you do, you know that these days basically all content that people are exposed to on the internet on a daily basis is completely and entirely selected using machine learning algorithms. And that's an extremely huge shift in the way that we consume news and, you know, current events and things like that from just even five years ago, I think, I would say. So, and then there's this also, why are AI and blockchains relevant to each other? So for AI, what it gives to the decentralization space is the ability to carry out complex behavior, right? Because most of what we, you know, starting with Bitcoin and on out, most of the applications that we really think about with blockchains tend to be very simple. So exchanging money or, you know, kind of like, well, they're all variations that for the most part right now they're all basically financial applications that's growing a little bit, but they're still pretty simple. And what AI can give to decentralization is complex behavior. So the ability to adapt to situations as they change and to just generally act more intelligent, right? And on the other side, what decentralization technology gives to AI is this notion of autonomy, right? So if you think about artificial intelligence, right? You know, we usually associate it with things that we can control, right? But like, we can't control other intelligent beings, like you don't turn an animal on with a button, right? It's autonomous, right? So autonomy is actually, to me, I think a very, very crucial and necessary component of any intelligent system. And with blockchain and other kinds of decentralization technology, not just blockchain, we have the ability possibly to create systems that exist beyond the ability for us to turn them on and off. They're truly autonomous and sovereign and they have their own computational resources and their own value resources and so on. And so AI can benefit from that, you might say, once that. So the first part of this, I wanna get into like more about machine learning and deep learning and then we'll kind of like work our way back to decentralization stuff. So we'll kind of start talking with this and then loop back to decentralization. So machine learning is this field, kind of a subset of artificial intelligence which has existed for a few decades now and it basically is concerned with learning functional, learning functional models based on data combined with a learning algorithm. So like a really simple version of this is a system which will look at an image of a number and then tell us what number is in that image, right? So like image classification is a very, very core task in computer vision. And most of the time, we're gonna take a very high level view of this. We're not gonna talk in too much detail about how these work today, but just everything that we'll talk about depends on understanding this in the following sense that what you have is a function which takes in some input which can represent some real world object, let's say and outputs something that is meaningful or interesting to us. So what it is or a description of it or perhaps another image or lots of sort of more complicated, possible input outputs. And let's kind of consider this in the context of computer vision which is where it's had the most success in the last decade or so. This is how computer vision worked in the 1990s and 2000s. And even for me, like when I first got interested in machine learning, which is maybe 10 years ago, computer vision basically looked like this. What you, if you were interested in image classification like detecting objects inside of images, you would have to do all this sort of pre-processing where you would apply some well-known heuristics or algorithms for feature extraction on the image. Like for example, this is what you're looking at are called hog features. So histogram of oriented gradients. This is just kind of like divides the image into a bunch of segments and then extracts the dominant sort of motion vector in it and then constructs a histogram of those. And then it becomes a sort of small representation of the image upon which you can plug some image or some shallow machine learning model to it. And this is basically how machine learning worked for a really long time. And the problem with this is that it's, well, first of all, this, it worked pretty well, right? Like we could do like basic computer vision for a long time and it was pretty interpretable and it was simple enough to implement on commodity, like even on the laptop to do this 10 years ago, right? However, it didn't work that well, not so well that, you know, to inspire the kind of excitement that we see around machine learning today. And it was also like not very general, right? And, you know, because for example, these features that we're looking at are really only relevant to computer vision and they're not relevant to audio or to text. They're not even relevant to other tasks within computer vision, right? So if you wanted to teach this AI, let's call it, to do something totally different, you have to start over. So it's not very general. And generality is an important component of AI, right? We want to create AIs that can respond and that can be deployed in a very, very wide variety of situations. Now what's happened now is that we've replaced all of this with very, very deep learning algorithms, right? And this is kind of what we mean by deep learning. Generally it means like neural networks with many layers that do multiple rounds of processing, taking raw input data, like an image, you know, it's pixels or it's audio samples or text or whatever, and learning an efficient representation that is good enough to do some sort of image, some sort of a task at the end of it. It's trained end to end, has very good performance, the code is very homogenous, so it's kind of easy to work with. And it generalizes very well. So you can use the same approach for completely different tasks usually. All you have to really define is your objective. And then the machine learning takes care of the rest. So it's much more general and it's been much more successful with a lot of tasks that are important to people. And that has inspired a great deal of like incredible applications that we've seen. This is just like a small selection of really, really sort of recent applications that we've become quite good at, right? Speech recognition is done by deep neural networks and so is computer vision. So self-driving cars are all using what are called convolutional neural networks. We're not gonna talk about how this works so much, but that's kind of the workhorse of most of this. And also things as different from these two as machines that play Go or play other kinds of games or interact with the environment, right? This is like just some more applications. As you can see, there's quite a bit of diversity in the applications of deep learning. So extracting roads inside of satellite imagery, identifying galaxies, generating speech. So that's kind of like what we're looking at as wave nets. There's also been a huge explosion in medical diagnostic applications. So this is from a paper that shows how to detect breast cancer inside of basically in the histology images. And generally speaking, like a lot of this kind of stuff has now done better by neural networks than it's done by even doctors, right? So we're seeing like a huge explosion of these kinds of applications. There's a captioning system, right? So like things that can take images and describe them in natural language. And lots of others. And then my personal favorite is of course like artistic applications. And I mentioned this like really quickly. We're not gonna talk about this today, but this is kind of what I do most of the time. So generating images and sounds and text that's kind of more creative or maybe like less useful than some of the other things that we looked at. And for this reason, because of all of these disparate applications, there's been a huge increase in interest in this field. A lot of investment, a lot of hype, of course, a lot of press, a lot of people are calling this like the new golden age of AI, where the first one was in the 1950s and 60s when the humans first became excited about this idea that we could create AI. If you ever saw science fiction stuff from the 1960s, you know that it was very sort of like AI utopia and so forth. And of course, the investment numbers have been really staggering. So every just five or 10 years ago, major tech companies might have maybe like one or two people on the payroll who had some experience in machine learning. Now they have entire departments. So and you get, you know, coalitions like this from the CEO of Google, Sunder Pichai who says machine learning is a core transformative way by which we're rethinking everything we're doing, right? This is the largest, maybe not the largest, but like basically the most important company on earth, saying machine learning is everything that we do now. And if you look at their products, it basically is. This is kind of the way things have changed over the last few years. I've been really also like excited about developments in how machine learning is done. It's a lot more open than it used to be. So like when I first became interested in it, the only way that you could really like learn about machine learning was by either being a graduate student or by buying expensive textbooks and basically learning it on your own. Now there's like actually quite a big online community that is not really concerned with credentials or university degrees or things like that. There's a lot of people discussing this stuff online. There's all the paper publishing is completely open access. So archive.org is basically at this point in deep learning like universally selected as a publishing source. People publish in advance of conferences and journals and things like that. They share their research early and often. They talk about it online and on like the Reddit machine learning channel on Twitter. You have, you know, things like Kaggle where people where like amateur data scientists can compete and try to earn money actually independently doing machine learning tasks. You don't have to be hired by anyone. You can just submit algorithms to try to win competition. A lot of this stuff is like, it is I think actually like a really good development. And of course, a lot of this has also been supported by the appearance of these open source framework or doing machine learning. So you might be familiar with like TensorFlow and Torch. And this one Tiano is actually being discontinued now because of the appearance of Tiano and Torch. These are the TensorFlow and Torch but of course like, you know, this is completely different because before, you know, like all the machine learning stuff would be in like MATLAB or something like that which you would have to buy a very expensive license for or have your university or your work do it for you. And so this is kind of like much more open, much easier to enter. Not an easy topic to learn about but it's much more accessible than you think. That's kind of a nice phenomenon. So let's talk about how machine learning is done and start to connect it back to this notion of decentralization why it's kind of important. So what I'm about to describe to you is the basic machine learning pipeline that describes pretty much 99% of machine learning today. You have some company, let's call it AI Incorporated which wants to train a machine learning model to, wants to train a machine learning model to do some task, right? So it will create some model, instantiate it like a neural network or whatever and it will have users submit data to it. It will take their data. It will then train the neural network on it. Now these are green because it's trained and then it will give them back some service, you know, YouTube recommendations or something like that. And then the AI Incorporated will sell the data and that's how the business works, right? And this completely describes pretty much every technology company the way it works. They're all pretty much, you know, taking your data, training, you know, doing some data science in it so they can learn very interesting things and then either selling the data or selling the service as an API or something like that. And this describes, you know, the way Facebook's business model, Google's business model and, you know, pretty much every major tech company. So let's talk about like some problems with this, right? One problem is that there's a constant tension between the users and the company with respect to things like privacy, right? So you have this data. The data's very personal sometimes and, you know, so of course there's a tension to giving it up to a company. And then of course it's, you know, there's a convenience to, they want to create some convenience in your life, you know? So like giving you news that's most relevant to you, right? And of course the trade-off is that you have to give them data about yourself. And this is irreconcilable. This is always a tension, always a problem somehow. The second problem is that this notion of lost natural income. So when you are generating data for AI-incorporated, it's a labor that you're performing, right? You are generating data and you're doing it through the way that you're interacting with our service and you are not being paid for it or compensated in any way, right? And of course like the return is that you get some free service which is usually not a very good service, right? And many people, you know, and this is not just about machine learning, just in general, many people have been commenting on this for years that this is kind of like the economic model behind this is a little backwards. Like maybe it should make sense that when people generate data, there should be some incentive for them to do so rather than just getting some free service for it. The third problem is that the data is aggregated into just a few very large data silos, right? So Facebook, Google, I believe those two companies together are counted for something like 77, I think I read this today, 77% of all advertising income on online in 2017. And advertising is the product that's connected to having data. That's the business model of the internet. So these two companies soak up all of the data and a small number of people have ultimate control over that data where it goes, what it's used for, who has access to it and all sorts of other things, right? So this is kind of like creates this big power dynamic where several companies have the most precious resource of the 21st century, which is data about people, right? And people call data the new oil, right? And there's a lot of truth to that, I think. And this one is kind of related to the first one, the fourth, the privacy issue. Some products that companies would like to provide are very sensitive, right? So you can imagine that some companies out there would like to do, for example, to monitor your health and maybe warn you of some potential problem with your health, right? But there's a, of course, there's a disincentive to share that about yourself because you're afraid it might be used against you, which it often is, that's just a fact. And so again, you can see that this creates a very inefficient, possibly suboptimal model of interaction between customers and AI. So enter decentralization technology. So I'm gonna now shift gears and talk about decentralization stuff. And then we'll kind of look back to how these decentralization products can maybe either solve or at least improve some of the problems that we observe with machine learning. And I wanna first start by kind of giving you like a sort of taxonomy of decentralization technology because it's not just about Bitcoin and it's not just about blockchains, right? Decentralization can refer, at least like, I'm talking about mostly contemporary things. I imagine kind of there's like three categories of things. Decentralization is the outer bubble here where we're trying to decentralize all of the things that possibly that you do online. For example, storage of data, computation itself, network topology, things like that. And there's all of these different, either companies or open source initiative, many of which are decentralized themselves, not all of them are, which are attempting to actually try to create these services in a decentralized manner, which is to say that the service exists independent of some central authority. And this is just a few of them and I'll actually talk about a few of these in more detail later. But IPDB and IPFS, for example, I'll talk about later. These are all examples of decentralization technologies which aren't necessarily just blockchain stuff, right? A blockchain takes up most of the attention in the decentralization space but it's really only half the story. Then they go to the middle there and then we have blockchain stuff, right? So, and what most people associate with blockchain is Bitcoin, which was the first blockchain. But blockchains can be used for all sorts of other kinds of value exchanges, right? So, or keeping track of some evolving systems, right? So if you think about like business logic, want to decentralize business logic, you need to have some like decentralized computer that is actually computing the logical steps. And how do you do that in a decentralized way? Maybe on some computation that's distributed somehow. Again, we'll talk about these in more detail. And then at the lowest level here, we have cryptocurrencies, right? And that's kind of the most specific. And that's your Bitcoins and your Moneros and your Zcash and so on. And so just to dispel like a few maybe like cliches that I like to kind of start with. So if you ever like have Googled decentralization, you'll find that like half of every blog post or chapter that's written has this famous graphic that shows centralized, decentralized and distributed systems. And I think sometimes this is like, doesn't capture the whole story because decentralization exists on many axes. So any service can be decentralized in some ways and not decentralized in other ways. So for example, like if you consider Bitcoin, which is commonly called decentralized, well, that's only half the story because it's decentralized in the operation of the network. But of course, there is a developer community, there's a core developer community that's involved in developing it. There's a Bitcoin foundation, which is not official exactly because there's no notion of official in a decentralized entity, but it's de facto one organization that's developing it. So there is some centralization to Bitcoin. Of course, like there's also the mining pools that have emerged that have a huge amount of political power. And so decentralization is not like a binary thing. It's not either that you have a central authority or nothing, it's kind of this axis. And something can be decentralized in terms of operation, but maybe it's, let's say be decentralized in this operation, but decentralized or fully centralized in some political or authoritative. And decentralization is not just about Bitcoin. And of course, Bitcoin is skyrocketing in price right now. So of course, all of the news is about it. And Bitcoin has this kind of quasi-libertarian culture to it, but there's a lot more to decentralization than just Bitcoin. There's a lot more than just lock chains. And you'll talk about things like IPFS later, which don't have anything to do with lock chains, but are actually quite interesting in their own regard. And actually yesterday I talked about it much more linked at CIS. We talked about decentralized web technology. And these things are actually really important to be aware of. And what you find as you get into this ecosystem is that a lot of these new services are actually like very complimentary to each other. And that if you're building some sort of a, what you hope to be a decentralized service, you might be using multiple systems at once to handle various aspects of the product that you're creating. And finally, decentralization is not new. The, does anyone recognize this graphic on the left? This is the original schematic that described ARPANET, which was the internet, the very beginning of the internet in the late 1970s. At the time we had figured out how to have computers communicate with each other using telecommunications infrastructure. And along came a few people who said, wouldn't it be great if all of these different networks which had developed proprietary technologies to communicate amongst each other, wouldn't it be great if they could communicate with each other? Because all of them were using different protocols that they invented, like IBM have one thing and maybe the US maybe had another thing. And so along came a number of people who developed a whole bunch of the protocols that now govern all of internet telecommunications today. So you're, I'm sure everyone's here is familiar with HTTP and FTP and DNS and TCP and IP and all these protocols which define very precisely how to implement something that can interact with the internet. And then they start about convincing everybody to adopt it and then the internet was born. So when this happened, the idea, the original idea was the internet was going to create decentralized telecommunications among people. So suddenly all these computers can interact with each other. You have a computer at home or maybe you have a computer at your workplace and you can all talk to each other, right? And the original spirit of the internet was very much decentralized, right? This idea that we could communicate in a decentralized manner. But what happened? What happened is that all of these things happened. So some of the components, some of the nodes in this decentralized graph grew to be very, very large because they provided some superior service. And over time, a lot of the traffic began to concentrate in very, very small portions of this graph of people communicating with each other. So of course, like everyone recognizes all of these, you know, Twitter, Google and Facebook. And I also like to point out that, so the following companies, eBay, PayPal, Amazon, Airbnb and Uber, and I think a few others that you can mention, they were all like the whole idea of them was sort of like to be decentralized, right? So like eBay was like, oh, finally a buyer can sell, so someone can sell straight buyer, go straight from buyer to seller without some middleman in the middle. But what happened over time is that eBay grew and then they had to get involved in settling disputes and they just became another middleman, right? And this is the same as true of Amazon and then sharing economy apps or what they call themselves sharing economy apps. Like Airbnb and Uber, same thing, right? That was kind of like their selling point, but really they just threw to become, you know, a big sort of middleman. And so decentralization was effectively centralized. And this was very largely consolidated over the last 10 years, I'd say 10 to 15 years. Even in the 1990s, things were still like, you know, a little bit more disease. Things now have become very centralized. So this is manifested in some problems with the web today. And I'm gonna talk later about IPFS, which is kind of like the initiative that's trying to solve these problems in particular, but I'm gonna first introduce these problems, right? But what's wrong with the web today? It's slow and inefficient, right? And this is sort of contrary to people's perception. The internet has gotten faster over time, but it's beginning to plateau in this regard. And you can see that I don't have the address, but if you look up over the last 10 years or so, you'd seen that the price of storage has plummeted. You know, it's become extremely cheap to store things on computers, to back things up in the cloud and things like that. The price of that has plummeted, but the price of bandwidth has kind of stagnated. It's not getting cheaper as fast as storage, which is to say that we're accumulating more and more and more content, but not getting faster at traversing this content, which gives us the perception that the web is actually getting slower. And this is particularly true, like I think here, especially in India, where everyone is basically connected to the internet, but you can go in many places where it's extremely slow. And this is actually a big barrier to using a lot of services online. That's kind of one problem. Another problem is that the web is not really secure. It wasn't built to be secure. It was built sort of before the notion of cybersecurity was really like the kind of issue that it is today. So if you, you know, the way that security is added is well, first of all, why is it insecure, right? Because when you're interacting with services online, you're, you have to trust them, right? So if you're requesting some content from a service provider, you have to trust that they're not being malicious or giving you some computer virus or it's modern day form, which is advertising. But anyway, and you know, the way that we kind of made this more secure is through what's called the security through security on, through insecure channels. What is it? Security on the insecure channel, secure communication through insecure channels. So, so encrypting content while it goes through those pipes gets to you and doing all sorts of complicated cryptography, like to secure these insecure channels. And this, you know, works pretty well, but it can break down. And of course it breaks down because it's kind of a patch to the infrastructure of the internet. The infrastructure of the internet is insecure by default. You have to secure it with all of these appendages and all of these patches. The other thing is it's unreliable and impermanent. So how many people here remember GeoCities? So geocities.com slash Hollywood slash lot slash 6641, that was my first website. And you can no longer find it online because a couple of years ago, Yahoo bought GeoCities and then decided we're not gonna shut it down. And then this huge trove of early web content was completely lost forever. And of course this happens all the time. You know, how often do you get a 404 page when you try to click on the link for some, you know, resource that you're looking for. Often gone because, you know, some webmaster took it down or maybe the ISP that is hosting it decided to take it offline. You know, it's impermanent, which means that we're losing a lot of valuable, often very valuable, you know, content about ourselves. And you know, there's things like archive.org, right? Not the same one I mentioned before with an X but, you know, with a CH archive.org which tries to kind of keep, tries to just save a version of the web every day. But there's only so far this is another problem. And then this I already made reference to this notion of accumulated data. So this picture is of some random server form owned by Facebook. You can tell it's Facebook because of the blue glow. And it has, you know, thousands of servers. This is somewhere from random location, possibly a secret, actually, I don't even know where it is, like a new power or something. And it holds, you know, all of our data and a very small, very select group of people on the inside of Facebook have opened the power over that data. So when you interact with Facebook and you post something there, you are signing off your rights to that post. You know, you came up with it, your content but then Facebook has end game usage rights over it. So it can censor it if it doesn't like it or it's been asked to. Or maybe it decides that the service that you put it on is no longer relevant to their business model. So they just take it off. And then you can't use that same content for another page, right? So like, you know, there's some, you know, maybe like, okay, Facebook lets you cross post to Twitter and vice versa, but they're always quite reluctant to do this. And you'll see that there's a lot of friction for using your data for multiple services, right? So this is kind of like, because it's competitive advantage, right? Data is new oil, right, as we said. So it's in Facebook's best interest to keep that data for themselves and not let others use it or build products off of it. So these are all problems of the web, right? Now there's been like the emergence of peer-to-peer networks over the last 10, 15 years which have tried to create some alternatives to this. So a peer-to-peer network is, well, it's a decentralized computer network, right? So you have, this is the centralized model, the server-based model where you have some central server that deploys content to a bunch of clients. This is what we call a server client model, right? And all of the dominant protocols that the web works on, you know, from HTTP on down and IP and so on, they all start with assuming a server client model where you are requesting some content from a server and it is sending it to you, you are the client. In a peer-to-peer network, you get rid of the central server, and you just have communications among all the nodes, right? And we've seen, you know, way before blockchain, there's been a lot of peer-to-peer networks that are quite decentralized, right? So this is just a few of them. The top left here is a screenshot from Napster, who remembers Napster? This is of course like a way to discover lots of music back in the day. And then Napster was kind of succeeded to some degree in terms of peer-to-peer file sharing by BitTorrent, which is kind of a, again, it's a decentralized peer-to-peer network of people exchanging files. And then these are kind of like a few more recent ones. This one, well, this one here on the right is called Diaspora. How many people here have heard of Diaspora? Do you people? So Diaspora is essentially a decentralized Facebook, and it was this really, really, you know, well-intended initiative that was started, I'm not sure how long ago now, maybe someone knows better than me, I think. You know, maybe he's far back as 10 years ago in the aftermath of Facebook. But it never really scaled very well. So like today, I think Diaspora has something like a million users, which is, you know, like respectable, but you know, Facebook has a billion users, right? So it's quite a big difference in magnitude. And so it never really quite took off, right? And then you have, this is much more recent. This is called Mastodon. So Mastodon is a quasi-decentralized, probably more accurate to call it federated, a federated version of Twitter. So it's an open-source software which re-implements basically what Twitter has. And then anyone can start a Mastodon server, which participates in this network, this federated network of mutually interoperable Mastodon servers, which each have their own rules. They can all fork the software and adapt it and make their own set of custom rules and things like that. And so this is kind of an initiative that's more recent that's trying to basically make, again, like quasi-decentralized Twitter. It's like, I say quasi-decentralized, because I think like one Mastodon server accounts for like 50% of the accounts. So it's like not exactly, like as we said, decentralization is a continuum. So it's not really fully peer-to-peer, exactly, but it is like a step in that direction at least. And then this one at the bottom is specific to Germany, which is where I live right now, Freifunk, which is the free radio. So this is a, huge initiative that's quite popular in Germany to create a decentralized wireless internet network. So like a bunch of peer-to-peer nodes that share internet connection. And it's a pretty active initiative. So let's talk about cryptography. And this is gonna be, so the next section is, I wanna talk about like how we are going to decentralize some of these things. And then we'll, well once we have this sort of idea of like how things are decentralized, we'll turn back to AI and test and make things. I wanna keep everything so you know where I'm going with everything. But I wanna do like a very quick review of cryptography, which I'm not gonna talk about in any technical detail. On Sunday when we have our workshop, which actually raise your hand if you're gonna be at the workshop on Sunday. Just, okay, just a few of you, okay. So most won't be. That's fine. I'm just for those of you who are gonna be there. All of these like, I'm gonna talk about the more technical detail like how some of these protocols work. But for now I'm just gonna talk about applications to keep things like at a high level. So the basis for the most decentralization technology now, some of this newer stuff is cryptography. And cryptography is the science of obfuscating communication. How many people here are like familiar with cryptography to some degree? Not surprised. So you're all probably quite familiar with a lot. You know, we didn't have to spend too much time on them. These are just some of the applications of cryptography. So for example, top left here PGP mail. So PGP is a system of encrypted communications generally used for email. Stands for pretty good privacy, which is kind of an understatement. And it uses public and private key cryptography to secure communication. So in public and private key cryptography, any entity like a person generates one public and one private key. And you can use the public key to encrypt communications to that person. And then only the private key can be used decrypted. So if you're the person who is being communicated with, you have a private key and only you know what it is. And so therefore only you can open encrypted communications to you that was encrypted with a public key corresponding to the private key that you have. So, and this is of course used not just in email, but it's also used in apps like Signal and Telegram and even what's happening with it. So they use encryption to make it so that you have some amount of privacy through their service. You know, they don't, they don't, especially Signal and Telegram, they don't store your unencrypted communications anywhere. This gives you some level of privacy. Cryptography is also used to do the, as I called it, secure communications through insecure channels. So things like TLS and SSO, you know, if you've ever used HTTPS, this is secure internet browsing, right? And that's another application. And then these are related. So VPN is a way of browsing the internet securely through an intermediary, encrypting your traffic, encrypting your requests. And then finally Tor, right? Tor is the onion network. So this is kind of like a way of browsing the internet securely through a peer to peer network of volunteer nodes that encrypt your traffic in multiple layers of encryption in such a way that no node in network has all of the information necessary to know who is visiting what, right? So this is kind of a way of obfuscating web browsing. So these are all efforts to kind of make the web more secure, more private, you know? And they've been quite successful, of course, for a long time and you know, precede a lot of the assignment that we see with blockchains and blockchains kind of take this stuff. So getting into blockchains, the very first blockchain was Bitcoin, right? So of course everyone knows Bitcoin, you can't walk anywhere, turn a corner without hearing about it. Quite sick of this, I know. But Bitcoin, of course like, but Bitcoin is really impressive, right? It was really the first time we ever saw something like this work at a mass scale that involves money, right? Involves value transfer, you know, a peer to peer network that has money built into it. And really that never existed before Bitcoin or at least it didn't exist in a decentralized manner like Bitcoin. And Bitcoin has been around now for about seven years and what it is, of course, is it's a peer to peer currency, right? So you can exchange money directly between peer nodes in the peer, there's no central authority that issues the money. There's no central authority that governs it. Well, as we said, there's developers, there's miners and they all have their say, but the functioning of the network on the day-to-day level does not depend on any central authority to actually authenticate transfers of money, right? The way that we usually rely on banks to do so. And Bitcoin is secured using a very elaborate and very sort of like wasteful system of cryptography to basically make nodes expand in arbitrarily large amount of computation to authenticate transactions. And this is a way that, again, like I'll talk about this on Sunday and much more, much more, like we'll actually talk about the Bitcoin protocol, each step of it. But at a high level, you're using, you're making people solve very, very like arbitrary and meaningless computational puzzles in order to authenticate transactions. And this is a way of reaching consensus on the order of transactions and therefore how much money everyone has in a completely decentralized way. And that's kind of like what Bitcoin is. And this is the only investment advice I'll give you. Don't invest in Bitcoin. It wastes so much energy, like this, these are just some graphs I found the other day. Bitcoin network now uses as much electricity as Denmark populate in the whole country of Denmark. And it's projected that in sometime in 2018, it will overtake the United States as the biggest expenditure of electricity in the world. And the higher its price goes up, the more people make from mining and then it just drives up the energy usage even more. So it's kind of like a runaway out of control train. As I said, there's lots of amazing things about Bitcoin, but of course it's also the case that there's this huge ecosystem now of really interesting projects, including other cryptocurrencies many of which are actually trying to solve this, this energy waste problem and inefficiency problem. And maybe this little relining is that if Bitcoin keeps on rising at some point, the money will kind of like drop down into these other ventures. I don't know if that will happen, but that's kind of like the state of things today. So getting back to the decentralized web. So as I said, it's not all just blockchains, right? So this is a really cool initiative called IPFS, which stands for Interplanetary File System, which is a open source initiative to try to re-decentralize the worldwide web and basically reinvent the protocols that the World Wide Web works with from the ground up. So starting with HTTP and IP and all these protocols to basically just get rid of them and start to scratch and do so by creating a decentralized graph. And again, like this, and I'm gonna talk about in more detail on Sunday when we have the workshop, but at a high level what IPFS is, is it's trying to replace the server client model of the World Wide Web where you have some, you know, a few very big central servers that are providing these services, replace it with a fully decentralized peer-to-peer network of nodes that are trading content with each other. And it uses a lot of really quite recent innovations that have come out of academia in general for efficiently hosting content and kind of trading content. In some ways, I like to think of it like a few people have made this analogy, including Juan Dene who kind of started IPFS, that IPFS is kind of like it for the internet. How many people here are developers and know of GitHub? So imagine if the whole internet were like GitHub, all of the content where there was version control for all of the content. So as you go through versions of the web, all of the previous versions are saved. You have this kind of trail of commits and so nothing ever goes away. You can always recover something. Now of course that creates certain kinds of risks, like what if there's content on the web that you don't want to be there? Like some content doesn't need to go away sometimes. And the answer to that is that you also get rid of these data silos and you make the data layer something that's controlled by the user ultimately. So if you're a content creator and you put up some content, ultimately you can make the content such that no one can access it, right? Or you can revoke the service's ability to use it basically like putting the content creators of data in charge of their own data, making data a sort of public utility. And so this is kind of the IPFS venture. And then also websites and services don't have any central origin, they don't have a point of failure and security is built into the protocol in like really clever ways. We'll talk about like on Sunday, we'll talk about this notion of location or content addressing and so on. So I wanna, let's actually like, we'll just take a break right now. We'll kind of continue this after a short break and talk about some more. Okay, so getting back, so I just mentioned IPFS, right? IPFS kind of takes care of one element of this which is the creation of a data structure for the web that kind of is able to efficiently find the nodes to distribute content. But that doesn't solve everything for a decentralized web. There also has to be some way of storing the data and especially large data. And so this kind of ties into, there's now been a few initiatives that have started up that are interested in creating a decentralized file storage mechanism. So like right now, the way file storage generally works is you have servers, right? You have small servers that might host your content. And a lot of it is now shifting to really large systems like for example, Google Cloud and Azure and Amazon S3 and things like that. So there are a number of initiatives that say, wouldn't it be nice if maybe we can create a decentralized market for file storage? So you have some excess hard drive space in your computer. Maybe you can make it available for other people to host their content on it. And maybe you can get paid for it too, right? So people should be paid for this as a service. So you create a market, some people buy, some people sell. And in general, like because there's no middlemen to kind of take a cut, it should be much more, much cheaper, much more equitable and so on. And much more efficient because it makes use of all this sort of excess hard drive space that we have. And Filecoin is actually like, I think the sister project to IPFS, it's meant to be very complimentary, it's gorgeous, there's another one that's in this space. And of course there's like lots of, there's a lot of outstanding questions to how this would work effectively, like how do you make, how do you ensure that people are actually storing the data, how do you trust them? To keep your data, of course, encryption works its way into interference and you don't actually give people the raw data but the encrypted version of it. And there's some fault tolerance built into the system by hosting the content in multiple places at one time. So if one of the nodes fails and your content can still be recovered, all of those are like questions that are being worked on right now. And a few people have asked me during the break, how can I sort of get involved in some of these projects? And basically all the ones that I'm listing right now are all effectively open source projects. You can find them online, a lot of them have like Slack channel or Gitter channel or of course all of them are, basically all of them are open source projects on GitHub. They're all looking for contributors, they have documentation, they're sort of like new projects. They're all really hungry for people to contribute. And this is just like a sampling of a few other services that are really, really small sampling. Like this is not by any means like even my favorite ones. This is just like a few that are trying to provide some, like a decentralized version of some, something which is done by a tech company right now. So like for example, Schemit is this company that's trying to create something akin to like a Reddit which tries to incentivize people to post content by giving them tokens whenever they get up close in the content. So we haven't talked about tokens, but this is a whole other aspect to the sort of blockchain space, you can create tokens which are a store of value, which can be used in some closed sort of sort of user system. There are things like Open Bizarre, which is kind of like the truly decentralized version of Amazon. This is a really cool one that's actually open Berlin or at least open Berlin, which is kind of like a decentralized version of SoundCloud. So for musicians to host and screen content and kind of share music and things like that. Then you have IOTA, which is trying to tokenize internet of things devices and just lots of things in this ecosystem. And actually like I have a graphic here that just shows like just how many hundreds of these things there are. And this is just again, like a sampling. I think these are just some companies that were present at the last Ethereum conference. I'll mention Ethereum in a second. And then this one is actually quite interesting also. There's two companies, probably three or four, besides for these Gollum and Trubit, which are working on decentralized computation. So kind of equivalent to Filecoin and storage except for computation. So instead of access hard disk space, access CPU power or GPU power. You have some sort of a computation you have to do, maybe a scientific computation, like machine learning, protein folding, GCC and things like that. You can have it done on excess computation that all of us have on our devices. And again, you can create a marketplace for it and create a token which incentivizes people to participate in it. And that provides another element of decentralized computational services. Because if you're creating a service, there's multiple things you have to take care of at once. There's the computation, there's the storage, there's some sort of a, like if you're working with databases, there has to be some way of querying databases. All of these things either rely on a centralized service or they rely on media peer-to-peer graph. These are all trying to provide each of those components. And then this is really new and this is done by actually some friends of mine in Berlin, there's a new thing called Ocean Protocol which is trying to create a, basically a decentralized data marketplace. So I mentioned before that we have this big problem of data being siloed into companies, right? Or what if instead you had the way that we organized data was into one large kind of collective, we might think of as a global data commons, right? And services, your Facebooks and your Googles and your Reddit and so on, rather than holding your data themselves and accumulating it, they are simply a service layer on top of this resource that they share with each other. And anyone who puts data into this big decentralized data global data commons has usage rights over it. So they can license it, they can provide all of the rules of how it can be used, they can revoke access. Ultimately, the person who puts the data onto it has a final say of how it's used rather than the service provider or the company. And then of course there's like just lots and lots of different use cases for this. The people that are making this talk about how, for example, companies can share data in situations where they otherwise would not have. So for example, like you have self-driving car companies which have to train, they have to train AIs to drive cars. And for that you need a lot of data. And it's really expensive to acquire that data, especially for something like self-driving cars where if it makes a mistake, of course there's a lot of destruction. So it would be great if you're able to kind of get a head start by sharing data. But the thing is that these companies that have this, data is their bread and butter. It's the thing that gives their company value. So what is the incentive for them to share it? Well, if you create a sort of marketplace for them where they can exchange data and maybe pull it together and license it to each other, then you can maybe provide the right sort of balance of incentives to actually collaborate. And this makes it a lot, this will really, really improve. First of all, it'll improve the state of AI because for two reasons. One is that now there will be a way of sort of greasing these gears to allow data to be accumulated where it's needed for multiple sources. And also it's also useful because most of the data is not actually, you have to be inside of one of these companies in order to do a machine learning experiment. And this will maybe create a mechanism by which it's not just simply who's inside and who's outside these companies, but rather anyone can participate if they have the right objective in mind or an objective that's useful to other people. So this is kind of another really interesting object. I wanna talk really quickly about smart contracts and Ethereum, so smart contracts. So all of this stuff depends a little bit on being able to have rules and logic that is carried out and agreed upon. And in a, now this is easy to do when you have a decentralized service, you have one computer that's kind of handling all the logic. But what if you want to be able to do this thing in a decentralized way? So for that you need to have, you need to have some way of doing more than complication. For something like Bitcoin of course is really simple. It's just people exchanging money. But if you have an application which is somewhat more complicated, you'll have to write, you'd have to do something a little bit more complex. And smart contracts are a way of enabling a class of applications that are much more general than simply exchanging tokens in each other. And this is kind of a concept that's been around since the 1990s, Nick Dabow first wrote about it. And the idea of a smart contract is it's a contract which is digital program and it's effectively self enforcing. So if you can create a contract which is computed over without necessarily needing some like after the fact permission from the parties, then you can create a system by which logic can be executed in a way that everyone can agree is secure and will work and transparent and all these things. Nick Dabow made this analogy of a smart contract to a vending machine. So a vending machine is like the physical manifestation of a smart contract, which is an analogy I really, really like. So you can imagine a vending machine, it's just like a program that has a few rules in it. You'll put in $2 and get back a candy bar. Put in $2 and request a candy bar that I don't have and you get your $2 back. You get like, you have this sort of like very simple, very simple logic programmed into it and it would cost you more money to break into it than it's worth because of course like the vending machine is really hard to break into and you wouldn't break into it just to steal some chip. So this is kind of like the idea behind Ethereum, which kind of emerged in the wake of this phenomenon that we started to observe a few years ago that a lot of people began to consider like second generation applications on top of Bitcoin or more complex ways of interaction besides for just changing money. As people began to think about developing these, Ethereum kind of emerged as a, that promised to become a platform where you can deploy smart contracts that carry out much more complex logic. So for example, things like a vending machine except online, you might say. And again, like I don't have time to get into the Ethereum protocol. We'll talk about that on Sunday. But basically it's de facto leader for deploying smart contracts on a decentralized blockchain. And this is, yeah, as we talked about, smart contracts are sort of the basis for all of the stuff that Ethereum does. And with smart contracts, you can create what are called decentralized applications and I'll get into decentralized autonomous organizations. So as I get into these slides, I just want to mention that we're starting to get into, like keep an eye out on the things that we're talking about because we're starting to get into things of more complex, that are more complex in their construction and things that have some sort of complex behavior built in. So this is where the AI element begins to become relevant. And I'm gonna introduce the AI stuff after we mentioned these out. So decentralized applications can be defined very roughly as kind of like this, right? You have a sort of entity which has maybe one or more smart contracts in it which contains some logic for operating up an organization or some behavior. And maybe it can store money as well. So like a simple thing. Now decentralized applications have existed for a long time. As I mentioned, BitCore is a decentralized application. It doesn't store money, right? But it has logic programming built into it and it has people at the outside. So decentralized application is something that, you know, it operates autonomously, is independent of every individual person that is interacting with it. But it's effectively a way for people to interact through it. And in the blockchain world, this often carries some association with cryptocurrency, right? So cryptocurrency is used to transmit value. And this is really the difference between second generation decentralized applications and first generation ones, like peer-to-peer network, I mentioned like BitCore, none of them have a way of exchanging money. Like you don't exchange money in BitCore, which is good because before blockchain, there would be really no way of doing this securely, not any way of doing it without some sort of a central intermediary, which would defeat the whole purpose of a decentralized cloud sharing system you begin with, right? And now the big difference is that now, we can actually decentralize applications that carry value inside of them, like monetary value, economic value. And this dramatically increases the class of applications for which decentralization is relevant. Now, then there's this kind of murky notion of a decentralized organization or a decentralized autonomous organization. These two things kind of blend into each other a little bit. And I think like the best way of looking at it is a decentralized autonomous organization or a DAO, a DAO, as we'll call it, to save breath. The difference between that and a decentralized application is that a DAO can also have some built-in capital or like built-in assets that it manages. So for example, maybe Deans or usage rights to something, or maybe assets like cars or houses. So for example, suppose maybe you have a car sharing service and if you want it to be structured as a decentralized autonomous organization, you would have a smart contract which manages access rights to these cars. So a person might want to rent a car and so they deposit some money into the DAO. There's a contract in there that says that if you give me money, I will program this car to open up when you come next to it. And of course the difference between this and a previous slide just a decentralized application is that the car is now entirely managed by the decentralized autonomous organization. And this is kind of a really, really, this really begins to get us into much more advanced territory. And like you can have cars that are managed by decentralized applications. And when I say manage, when you hear the word manage, you think people, like you think people managing things. But as I said, like there's no people in the middle here. There's no people managing those resources. There's no bosses, there's no middle managers, there's no employees, there's just a contract which manages access to a car. That's a pretty interesting phenomenon and one that is still very, very new. And there's not really a whole lot of things that have actually done this just yet, I suppose. Because, well, again, maybe it hasn't been quite enough time yet, but I think all the technology is there to create a completely decentralized consciousness, I would say. Another way of decentralized autonomous organizations can be increasingly complicated. So maybe they're not even just, like maybe you have multiple DAOs which interact with each other. Because basically all they have at the edges of each of these systems is a way to transmit messages. And the messages will trigger code to execute. And the code will execute and perform some action by giving you access to a car or a house. And, but there's nothing to stop one DAO from transmitting messages to another DAO. So one smart contract can send the message to another smart contract. And you can create extremely complex and elaborate sequences or chains of communication between these different DAOs that all exist on the completely decentralized platform. I know we're getting into really abstract territory so that's kind of like, it's hard to do anything about because, again, we don't have a lot of concrete examples of this. We just have like the tools emerging to build these kinds of services. So it's kind of like that's sort of the best that you can do right now. But I do have a few examples that I'll give you. I'm gonna skip this slide actually just in the interest time. I wanna give you a few concrete examples of DAOs that we could possibly encode as a smart contract that lives on the blockchain, like the interior blockchain. So for example, how many people here are familiar with Kickstarter? So Kickstarter would be a really, really easy DAO. Here's how it works. You have a smart contract that says, someone can create a page and say I want to raise money for this activity and I need to raise this much money in this much time. And then you allow people to put money into the smart contract and then the rules of the smart contract say that if you raise a certain minimum amount of money before a certain date, those funds will then be deployed to the creator. And if it does not meet the criteria, then the money is returned to the people who donated it. It's a really, really simple application, right? But it doesn't require a company in middle that's there to manage it. Now, you might be asking like, is this useful or is it helpful? Like what's wrong with Kickstarter? And that's kind of like, not necessarily in the scope of this conversation, but you can imagine that if you have a system in which there is no central authority, then you have much less, you have fewer ways of controlling things or maybe censoring things. There's a lot of things out there that are disfavored from these kinds of platforms. Now, you might agree with a lot of times like with particular cases of something being dismissed from a platform, but you may also disagree sometimes. And that's kind of like, that becomes a really big and kind of existential question like, who decides these kinds of things? And that is the kind of question that people are beginning to have. And you can also create, for example, co-ops that collaborate on insurance, let's say, you know, insurance can be something that can be encoded in a smart contract. If you know how it works in my home, the United States, my home country, it's basically like an embarrassing scandal right now. You know, like the health insurance scheme is basically like profits off of making people sick. I don't think that's an exaggeration. It's basically how things are in America. And I don't know if there's any like really good solution to this except to maybe just remove the insurance company from the business of insurance. Maybe it's something that all of us can enter in agreement with a smart contract that says that if someone gets sick, you know, there's some funds that get deployed. Now, there's all sorts of questions about that. Like how do you prevent fraud? Like who's there to prevent fraud? Who decides how much money needs to be paid? All of those are outstanding questions. And there are ways of encoding human judgment into decentralized autonomous organizations. Doesn't mean no humans. It just means that humans are sort of at the edges. And so one concept that you see a lot in the blockchain space is this notion of Oracle, which are like entities that carry some information that are sort of, it might be people that are making decisions, but they're basically third parties. They are disinterested in the outcome of some transactions and they can be, let's say in an insurance situation, they can be empowered to be judges. They're, I mean, it's not by any means a solved problem, but it's something that is increasingly more and more realistic as we get, as we sort of evaluate these cases. And you know, I don't think any such thing exists yet, but it may be reasonable to expect this, but it may in the future. I'd like a few more experimental ones, and I kind of like, I'll be talking about, again, like I keep saying this in the broken record, we'll talk about these in more detail on Sunday, but I'll just mention that at a high level, things like prediction markets and putarchy, which are things that I myself know very little about. I just wanna like mention the one sentence I can tell you about them. And then, you know, if it intrigues you, it's something worth looking up, and it's something that I've kind of become familiar with recently. You, Charlie, is this idea that you can basically fundamentally change the way that we do government. Now, the way you think of government right now, or governance in general, is in a democracy, let's say, is that people vote on policies or politicians, right, representatives, or in case of like a direct democracy, you vote on proposals and laws that will do something that you think is necessary, right? And this can be like really, really, well, of course, like, this is the best democracy that we have so far, I guess, but it has a lot of flaws, right? So one thing is that oftentimes, people make mistakes in the sense that like, they believe a certain policy will have a certain income, and they turn out to be wrong, right? Because sometimes it's very complicated to evaluate policies, right? So in the future, the idea is, I wanna say I'm not endorsing this necessarily, because I don't know if it's good, if it works, there's been lots of debates online about this, but I think it's an interesting topic, worth sort of analyzing. In the future, instead of voting for policies, you vote for desired outcomes, so things that you want the society wants to achieve. So for example, like, and things that are measurable. So for example, it might be something as simple as, want this high of a GDP, right? Or maybe in a more specific case, like maybe the society decides, okay, we want to reduce, we wanna reduce infant mortality to below this percentage by this year, something that can be measured, and then you vote on the outcome. You vote that you want those things to happen rather than any specific policies that are designed to actually achieve those. So you vote for the policy, and then once the policies are voted, or sorry, you vote for the outcome, and once the outcomes are voted in, you create a prediction market, which is a market in which people can basically place bets on policies that will possibly achieve those outcomes. And the idea is that you are incentivizing people to put money, to stake money into policies that they think are going to actually achieve those things. And without getting into the details, try to fully understand, using like basically like actually like first principle, like first capitalism principle, that somehow the market will eventually converge onto solutions that are effective, like efficient, effective at achieving those things. I'm not sure if that's correct. Lots of good debates about that, but the idea is, the idea is very much in line with a lot of the core ethos of decentralization. If you create the correct incentive, you will get like at least the best behavior towards those desired outcomes that may be better available to you. And then yeah, lots of other ones, of course, there's lots of financial instruments like mutual funding and so on. We're gonna talk about Numeri in a couple of slides, which is another good example. So of course, like if corporations are decentralized, then the perceived benefits that people talk about are that it's more transparent and it's more inclusive because it's kind of like fluid. People decentralize karma's organization is effectively designed so that anyone can participate in it, or at least in principle, like it's able to interface with any person that bears an account in the system and possibly resistant to collusion because all of the actions are public, they're transparent and so on. And but that doesn't mean at all as well. There's dangers too, so what happens if it goes wrong? Like what if it does something wrong? Who's accountable for that? And these questions are mirrored in the AI space. The people are asking like, when if a self-driving car makes an accident happen, who is accountable for that accident, accident? Is it the person that developed the software? Is it the company that deployed the contract? Is it, you know, who is it? It's really unclear. These are all emerging questions that aren't any like very obvious answers to right now. And the questions are very much, very much important to decentralization space as well because when you create decentralized autonomous organizations, by definition, they don't have people that are necessarily accountable to them. So what happens when they do something wrong? Maybe for us to figure out, right? And how they deal with externalities, like how to decentralize autonomous organizations, deal with global warming, like climate change. Those are also, I think, very much unsettled questions. Okay, so now we're gonna get into some like really heavy, very AI down. So getting back to the AI concept. So what we've just described are, you know, decentralized autonomous organizations are these entities which can perform very complex sequences of logic without any central authority to stop them or to start them. They interact with human beings, you know, at the edges but ultimately they are independent entities that have their own resources. They didn't have their own cars, maybe they have their own houses. Now what kinds of behaviors can we make, can we enable if we now add AI to the mix? And what do I mean by AI? Like I don't necessarily mean artificial general intelligence, which is still like a ways off, but maybe even like just like, even relatively reasonably realistic AI. So just more complexity, more sophisticated intelligence behavior. So like for example, if you have a DAO which manages cars, right, can you also make the car sort of do complicated things like look for clients, you know, like maybe deploy advertisements, maybe manage its own money, you know, comply with the law and effectively own itself. Like maybe you can have a car sharing service which owns itself. So I'll make one example. So in Berlin where I live, there is a company called Car2Go which is a car sharing service, which there's no drivers. It's basically like, it's like a car rental service except there's no garages. It basically just cars parked on the street. You can log into your phone and you go, I need a car. You find the nearest one that's in park somewhere. You go and find it. It's programmed to open up only to your phone. So it has some NFC chip reader or something like that. It opens up, you drive it. Drive it wherever you need to go. You drop it off where it is and then you leave it there for the next person who happens to be nearby and needs a car. Now this is managed by a company right now. It's called Car2Go. But in principle, there's no reason why this couldn't be programmed as a smart contract. Now there's a lot of questions you might have like, where does it get the cars, for example, right? So that's not clear. Like maybe people donate it or something like that. But effectively the whole system can be potentially managed by a decentralized autonomous organization. And you have the capacity for increasingly elaborate rules to govern it. And maybe it may be like, for example, the cars have to do lots of things. Like they have to figure out when does it need maintenance. Maybe it can subcontract. It can hire human clients to perform maintenance, to wash it, things like that. To donate the cars, like maybe we'll buy the cars from people, like all this stuff is, effectively there's no reason why it can't be done. Well I shouldn't say there's no reason why it can't be done without management, but it's increasingly realistic that you could do this without human management, right? So that's kind of like a thought that I'd like to put into your head. The idea that maybe something like car to go can be managed entirely by an AI and have humans interacting with it on the surface, but essentially be autonomous independently. There's another company that they call Numeri, which is another probably I think the most advanced example of both decentralized technology and AI. Numeri is not exactly a decentralized company. It actually is a company, but it's on its way possibly to becoming a decentralized company. What it is is it's a hedge fund. It's a hedge fund which does what hedge funds do, it invests in companies, right? It tries to find, you know, buy stocks and things like that and so on. Except it has a really big difference from other hedge funds, which is that what it does is it makes all of its data public and it allows any data scientist in the world to run machine learning experiments on their data and submit predictions to them for what stocks to buy and so on. And what it does is it takes all of those predictions and it combines them and then it determines a policy of what stocks to buy and so on from all of these predictions being put into like a meta model that accumulates all these models and then binds them and then learns from them. And they encrypt all the data using what's called a homomorphic encryption, which is something I'm gonna talk about in the next slide, depending on how much time I have. And with homomorphic encryption, this is a type of technology. I'm gonna describe it in more detail in a second, but it basically allows them to share their data with all these data scientists, except encrypt it. Encrypt it in such a way that none of these data scientists can take their data, of course, which is very valuable, but what they can do is actually perform machine learning experiments over the data and then submit predictions back to them. So effectively like everyone is able to collaborate on one data set and essentially it turns what used to be a zero sum game of many hedge funds competing with each other or clients money and so on, into one application in which all these data scientists can effectively collaborate together on financial policy. And it's pretty interesting, right? And I'll talk again, technical details on Sunday, but for now I'll mention one neat thing about this, I was actually just reading this yesterday, is that they have an ambition to not only to do what they're doing now, but to fully decentralize the company so that it's no longer like one hedge fund, but it's rather like, well it is still one hedge fund, but it's a hedge fund where there's not one entity which controls the data set. And if that were to happen, then I think at the most optimistic level, you could almost see this evolving. I don't know if it ever will become something like this, but at least like if we're being optimistic, it could evolve into making something like a hedge fund or investment banking in general, into a public utility, right? Like a communal property. And of course finance is like really important now, if there's society, like it's the way that we determine how to allocate resources into things, into projects that will improve society, right? But the problem with finance right now is that it's completely centralized and so a lot of companies are simply just greedy, right? Like this is kind of what caused the 2008 crash is just a lot of like elaborate schemes to transfer wealth from one part of the population to the other, right? So ideally, of course, but it's impossible to get rid of finance, right? It's something that we have to do as a society. What if finance were a public good, right? Like is that possible? Is that possible with a smart contract? Is it possible with decentralized kind of organization? And if it is, and if it's running AI, right? If it's running machine learning to actually determine what the optimal strategies are, then you have something that's like quite unprecedented, right? You have this like completely autonomous, intelligent algorithm, which is determining how finance is done. I wish I could like speak more concretely about it, but it's one of these things that like this is really hard to pin down because there's still sort of like very abstract concept in there, you know, and then they're kind of percolating right now, but we'll see kind of like over the next few years how things like are developed. Okay, this is the last thing that I want to talk about which is open mind. And I guess like, I don't know, I've been going really long for like, is there one thing with like another tender, before we should take that tender? So machine learning which we introduced earlier today, we mentioned these, we started, we posed these problems, right? There's a privacy convenience trade off, there's always a tension between the client and the server. There's lost natural income, and there's data that's aggregated into startups, large power imbalances, and products are very sensitive, right? So mental health applications or physical health applications. You know, like, you can read lots and lots of things about how Facebook can infer extremely sensitive information about you just from who you associate with and all of that is possible for machine learning, right? So these are all like things that create conflicts of interest when it comes to products that use machine learning. So is there a way that we can maybe fix this, right? And I'm gonna describe to you one really, really new venture which is called Open Mind. This is a really, really cool project that, again, like actually people have been asking me like, what projects have participated? If you're interested in all the things I'm talking about, like a lot of this is leading to this because this is kind of a company that's attempting to combine all of these technologies to solve these problems. Whether they do so or not remains to be seen. It needs a lot of help, needs contributors and it's just basically it's not a company at all. It's just this open source project of the community of people that are dedicated to solving it, have a Slack channel and GitHub and so on. I'm gonna describe the proposal that has been made by the community of people to create what they call encrypted decentralized artificial intelligence. So the sort of next is all the things that we're talking about. So let's review. I mentioned this before. This is how centralized machine learning work. You have AI incorporated, you have a company. And by the way, I'm just summarizing like if you go to search for Open Mind on YouTube, you'll find a description that goes into much more details than I'm about to, which will describe exactly how this works. I'm just kind of summarizing it. So in normal centralized machine learning you have your company that has machine learning model users give their data to the company. Company takes the data, trains a machine learning model, gives services back to the user and sells access to the model or the data itself to third parties for profit. And this has all of the problems that we just mentioned. So again, like just to review what are the problems of this? Data senders, we mentioned, the data is aggregated into one company. There's the privacy tension, lost natural income. On the flip side, the model is secure. That's kind of a nice thing about it. The model never leaves AI incorporated, so they have. Now there's an emerging way of doing machine learning called federated learning. So if you've ever used, for example, like Google auto-complete, you actually have, in most cases I think, for the most part, or like Google translate also, you typically will have, or at least can have, the model downloaded to your phone where you actually have a copy of the model and it processes on your phone rather than being like an API that you submit a request to, you actually do the machine learning on your phone or on your computer or so on. And what happens is that in federated learning, AI incorporated will actually share the model. It'll send the model to the users. The users have the model locally and they never give their data directly to AI incorporated. Instead they interact with the service and they generate what are called gradients. Now we haven't talked about what gradients are. In machine learning, when you have a neural network, something like that, it is an algorithm which is characterized by a whole bunch of weights, parameters that define the behavior of the model. And the idea in training a neural network or a training machine learning model is that you have to find the correct weight that make it behave accurately or meet the objective that you require of it. And the way this is done in training is that you give it lots and lots of examples, you figure out what the correct answer should be and what you actually have and you then calculate the difference that you have to make to all of the parameters in order to change the diff, like the, and this is called the gradient. But the gradient is just a whole bunch of like change all of these parameters by this much and it'll make the model get slightly better. So in federated learning, the users have the model and they simply send the gradients back to AI incorporated. AI incorporated then takes those gradients and then combines them and then trains the model and then sends them a new version of the model. So, and then, you know, gives them a service and now no longer has the user's data. So actually the data base should still be here. The users no longer have the data, but now AI incorporated can still sell, you know, like the usage of that model, right? Because it's still like a valuable usage. It doesn't have to sell a data anymore. So this is kind of like a nice thing for many applications, but there's still some weaknesses, right? So the nice thing now is that the user never has to submit their data to the company. However, it turns out, for reasons, for technically reasons we won't get into today, that the privacy is actually not preserved or not directly anyway, because it turns out that in machine learning, the gradients can actually give hints about what's inside their data. And you can actually do a very good job reconstructing or inferring the original data entirely from the gradient, right, that they give hints about the user's data. There's a whole field called differential privacy, which is dedicated to this problem, right? And it turns out that you can actually learn a lot about users from the gradient. And that's why if Google says like, oh, we're not actually saving your data, it's like true but slightly disingenuous because it really turns out that they can still know everything. This is how we'll federated learning system work on Apple as well. They can still actually, if they want to, learn a lot about you. Now, the lost natural income is still a problem. Users are not getting paid for submitting this data. And now we have a new problem, which is that the model can be stolen. So, like AI Incorporated has just sent these neural networks to the user that they would do the service. So what's to prevent the user from stealing it and making their own AI Incorporated, right? This is a new problem. So the way that, so the next thing that we're gonna introduce to this formula, and by the way, I just saw there's a warning that's gonna get increasingly complicated. I'm just giving you the very high level view. As I said, you can find like a more detailed view of this at an interview, but it is pretty interesting. The next thing we're gonna add to the system is what's called homomorphic encryption. So homomorphic encryption is a very, very fascinating and very, very bleeding edge research topic in cryptography. It's a form of cryptography that allows you to encrypt data in such a way that you can perform mathematical functions on top of the encrypted data, right? The encrypted data is called Cypher text. Often it's called Cypher. And you can actually do mathematical operations like addition and subtraction on the Cypher text in such a way that when you decrypt the data, those mathematical operations are respected, they're preserved. So if you encrypt the number three and then take the Cypher, that result, and multiply it by two, and then take that and then decrypt it, you'll actually get six, right? This is not normally, you can't normally do this in most forms of cryptography, right? Or you could do addition and so on. And it turns out, first of all, like this kind of cryptography has only existed really for a few years. Like maybe 20 years ago, we came up with a basic way of doing it, but it worked a billion times too slow. Like it was literally orders of many orders of magnitude slower than conventional cryptography. So it was impractical to use for most things. Now it's only like 1,000 times as slow as I'm actually making these numbers up. I don't know exactly the exact numbers, but let's say like, it's improved a lot, but it's still very costly. So, but it's become realistic due to certain kinds of operations, right? So now, imagine you take this kind of encryption and now we're going to work it into the system. And by the way, this is what numer, numer I use homomorphic encryption to. So here's what's gonna happen. And corporators gonna generate a public and private key, which is used for the homomorphic encryption. We're going to encrypt the model with the homomorphic encryption. Now the model's weights are red, right? This means it's encrypted. So it's encrypted and it then sends those encrypted models to the user. And the encrypted models can still actually, because they're homomorphically encrypted, they can actually still perform all the operations of a neural network and generate an output that is useful to the user. So what happens is they generate the output. The users can now use these homomorphically encrypted neural network, except now they can't actually steal them, right? But otherwise everything works exactly the same. User gets the service, sends gradients back to AI Incorporated, AI Incorporated and then train the model and then go on about, you know, send the new, send the new network to update the model to it and business as usual. So selling access to the model. Now what's, okay, so what's the proposal come here? So now you might be thinking like, okay, now privacy is really preserved, right? Because now actually the gradients are also encrypted, right? Well, it turns out this is not fully true yet because it turns out also, and this is now getting to really obscure stuff, but in machine learning you can design a model in such a way that all it's actually doing is it's just copying the data and copying it into the model itself and sending it back to AI Incorporated. And now because the model is encrypted, there's no way for the users to find out that that's happening. So malicious AI Incorporated could still steal the data. Now, however, the model is now not be, not itself be stolen by the user. So at least that's a nice thing. And then the lost natural income fully remains, right? We haven't solved that yet. Okay, now we're getting more complicated. I told you this is gonna get crazy. So now add a smart contract in the mix. Here's how it works. AI Incorporated, as before, generates a public and private key to encrypt the data, to encrypt the network, encrypt the model, and puts it into a smart contract. Now the smart contract has the model and it's encrypted. And it also puts in some money from cryptocurrency, let's say, into the smart contract. And now the smart contract is separate from AI Incorporated. On the blockchain, entirely public, it's encrypted, the model's encrypted so it can't be stolen. But now, rather than AI Incorporated doing all the processing, the processing is done in a smart contract which is separated from AI Incorporated. And now the smart contract will send the encrypted model to the users. Users have the model. They receive a service, generate gradients, and then what happens is the smart contract will pay them for access to that data, right? So that's kind of a nice, that's a new thing. Users are now being paid for generating data, for generating the gradient after they've been verified somehow. And then the smart contract will take those gradients, add them to the network, send it back to AI Incorporated, AI Incorporated will encrypt it, and train it, and then, again, profit, right? That is the business model. So what's happening here now? Users, now we can safely say that the user's privacy is entirely guaranteed. And the reason why it's entirely guaranteed is because, as I mentioned before, that you can learn a lot about the users from the gradients that they submit, right? But the gradients no longer go to AI Incorporated. They only go into the smart contract where they are mixed together and added to the network. And then once they're mixed together, once all these gradients are added together, there's no more way of reconstructing information about any individual user. We don't know of any way of doing that. So now we know that we can actually guarantee users' privacy in a system like this. And that's really new. Like, that doesn't exist in any real machine learning system that I know of, where you can truly respect people's privacy. And of course, that will have far-reaching implications we'll get into in a second. But going back to the pros and cons, privacy's maintained. Another nice thing is now the users are compensated for the data. So now this is really the economic model that the internet should be based on, right? When you create data, you are performing a service, right? You are doing some labor, and this users can be compensated for it. Can you hold questions to the end of the video? Now, and the model is also secure from theft. And the model is secure from theft because it's incorporated, right? That's been there since before. Now there's still one more problem. We're gonna have to introduce like one more complication to make it work, which is who holds the keys, right? So like here I actually put the keys into AI incorporated, right? But this wouldn't work, right? Because then actually all of this falls apart because AI incorporated could then, smart contract is public, so they could at any point, like let's say after one user has submitted the gradients, they could download the model and encrypt it and then do the differential privacy stuff to determine what the user's data is, right? So AI incorporated cannot actually be trusted to hold the keys, but who should hold the keys? Well, you can't put it on the smart contract itself because then anybody can decrypt it, so users can steal the data and incorporated can spy on them and so on, or maybe someone, you know, totally outside the whole system can steal it. So who should store the keys? Okay, so now we're gonna add one more feature and this is called an oracle in open mind parlance called the capsule, which implements what we often call an oracle in the blockchain state. But an oracle is, it's kind of like an insurance company. It performs the service of protecting the keys that are used for this entire process, right? And the oracle has to put up a large, it's like an insurance company, they have to put up a large, they have to stake a large amount of money for the privilege of protecting these keys. And if they do a good job, they get paid a little bit, right? And if they do something malicious, like steal the keys, they lose a ton of money, right? So basically like an insurance company, like a guardian of the keys, right? And so here's how it works. Finally, this is the final model. AI incorporated sends a blueprint of the network, like an untrained network unencrypted to the oracle. Now it's important that this is done publicly because now this prevents the problem that we mentioned before that a malicious AI incorporated could potentially steal users' data by making the model such that it's simply just copies data, surreptitiously, right? So now, because everyone can see this process, they can be verified that they didn't do that. They put it in the oracle, they generate the keys, and they transfer the public key, and this is not a homomorphic key, this is just a normal, like, RSA key. They send it to the capsule, capsule has the public key that they incorporated, and then capsule generates its own keys that are secret from everybody. And these are the keys that are used to perform the homomorphic encryption. So now oracle will encrypt the model with its private keys, homomorphically encrypted, send the model to the smart contract. Everything else is the same, basically. And then the money goes from AI incorporated to the smart contract. Smart contract sends the model to the users. Users have the model, get the service, give gradients back to the contract. Contract adds those gradients together, pays the users, sends the model back to the capsule. The capsule will decrypt the model, and then it will decrypt it with a homomorphic key. And then it will re-encrypt it, sorry, it will train it. Actually, I think this is wrong, I think it could get trained inside here, after double check. But basically, the main idea is that it decrypts it, and then it re-encrypts it with the public key of AI incorporated, and then sends the encrypted model back to AI incorporated. It has the final model, which you can then decrypt with its private key, which no one ever saw. Now, this process is fully transparent, what's happened, except that no point anymore are any of the privacy boundaries violated. So none of the users ever saw the weights, or could never steal the model. AI incorporated never has any way of decrypting the model and stealing people's gradients or stealing people's data. And so what's happened is AI incorporated was able to train the model in a fully blind way. They don't know anybody's data, so they can perform a service to that user without knowing anything about them, which is really powerful. Like imagine, we were talking before about this notion of sensitive services, right? And they're sensitive for the simple reason that the company is interested to know information about that user, which is possibly very sensitive. Now it can actually perform a service without knowing anything about the user, which is really unprecedented, it's cool. And of course, business as usual, AI incorporated can sell the model, so on, so it's preserved, it's preserved and everything. Okay, so now everything, we have everything, right? Users are maintaining data, privacy is maintained, users are compensated, the model is secure from the deck, the keys are secure, and everything, right? So this is kind of an ongoing project, and for anyone who's interested in it and wants a really much better exploration of it, highly encourage you to visit their website, there's a lot of information that you can get involved. Okay, so this is the last couple of slides, and I just want to pose some questions here, right? So now I want to, this is kind of my ruminations about this, but both what we saw in particular with Numeria and with OpenMind, could you have a system in which all of the users are equivalent to AI incorporated? In other words, there's no difference between these users and AI incorporated. You have a pool of users which agree to train a model which is decentralized that performs a service for them, and they'll all be paid for it, right? And they're willing to actually stake money into it because the model performs a valuable service to them. So now you have maybe a means of having a machine learning model which is effectively a public good, but the public utility that's shared by everyone. And you can possibly use cryptography to enforce that, that no individual person can actually co-opt the model and use it for something else, possibly storing it as a shared secret or something like that, like a fancy cryptographic scheme. There's lots of research into how these kinds of things can be implemented using fancy cryptography, multi signatures, like oracles and stuff like that, but like the question that I have is, and I don't really know actually, I don't know enough about it to answer one way or the other, can the users be equivalent to AI incorporated? Could the model be a shared secret? If it is a shared secret and it's hosted autonomously, sorry, and if it's hosted as a shared secret, can it be made run autonomously? You know, if it's on a smart contract on a blockchain, it'd be performing a service entirely autonomously, and can those models then control their own resources? The model has a notion of money that only it has and it can pay users to give it data, and it can also sell its services to, you know, license it to third party, and so that's a way of getting money into the system. So in other words, can you have a machine learning model which is a completely decentralized service provider that manages its own resources, right? And then the last thing that this is like on my mind is like, could this be an artist also? Like there's now, there's lots of these algorithms that are developing machine learning art, you know, machine learning based art that's got a big interest of mine. So like, what if it were doing that and it were generating art and it were selling it, like in an open marketplace, like it's a, you know, art can be sold. And this actually was written about by Trent McConaughey, he's actually started the Ocean Protocol and here's the, here's one recipe for this. And it's a pretty simple one. You can imagine more, much more elaborate ones, but imagine you have a decentralized autonomous organization which is running, which is hosting a machine learning model which creates generative art, makes multiple editions of them, timestamps them like the Claire's copyright, sells them on some, you know, possibly on even centralized services or maybe like an open bizarre or whatever. And then transfers the rights to the content, to the buyer, it gets paid, has its own wallet, right? It uses the proceeds to pay for the computation to actually make the thing. And then maybe it, I don't know, maybe it redistributes, like, what does it do with the money? Like, well, maybe it just pays the creator, but it's kind of boring or maybe it just begins to accumulate money. So like the question is, can you have a machine, like an artist, decentralized autonomous organization which isn't generating art, accumulating its own wealth? Can, can it do that? Can it become a millionaire? Like, can you have this entity which is acquiring wealth on the blockchain that's totally separate from all, from everyone, right? Which is a really, really crazy thing, but possibly, you know, like actually realistic. And then this gets into the question of like, okay, well, you have an entity which has its own resources, like, who's involved? Like, can this, do we have the legal infrastructure to actually do this? And it turns out that in some places we do, right? So this is a picture from a city in Switzerland called Zoo, which has been nicknamed by the blockchain people, Crypto Valley. And it's called Crypto Valley because it has some laws that are particularly eccentric. It allows you to register corporations that effectively have no people in them. You can have a corporation that has zero people. And we know, and by corporation I mean truly, a corporation that has all the legal rights of corporations that are given to them. So if that's true, you can have a corporation with zero people, and that corporation, and by the way, like, this is where all the ICOs are happening in places like this, because of course, it has all the laws that are necessary to have these centralized autonomous organizations that function as companies. So if you can have this, and you can have a corporation with zero people of accumulating its own wealth, you know, doing all the things that we associate with humans, like, is that feasible? Are there any legal barriers to it? You know, can it be made to comply with the law? Can it be entirely autonomous of people, right? So this seems to be a case that we have all of it, like, all of the technology and legal infrastructure is in place for personless entities have rights that are respected by legal systems in the world. So this is like, kind of like a very, very new thing. And I suppose, like, you know, maybe it remains to be seen, like, what we do with this, or whether or not any of these things ever pan out. I just wanted to, like, plant that seed in your brains, you know, like, whether these, whether these kinds of things can be made to work. And if they can, like, what should they do, right? So, like, I use art as an example, because, you know, art is this kind of thing that everyone responds to, right? But they can do all sorts of things. They can do also things that we don't like, right? And that's kind of, like, important because, and that's why I want to be in this space, despite maybe reservations about ethical reservations about what these kinds of entities can be made to do. Because, like, what they do is really, like, up to us, right? They can be made to do the most progressive things in the world, and they can be made to do the most evil things in the world. And, you know, we haven't talked about the evil ones, right? But there's no reason why they can't do any of the things that people do. And so this is a very double-edged story. And so it's really sort of up to us to build it, right? In this, in a certain way. So that's all I have for you. I'll kind of stick around if you guys have, like, some questions. And for those of you who are coming on Sunday, the workshop, we'll talk about this in, like, much more detail. I hope you can join us. And that's all.