 This speaker is Ziraj. He's a renowned data scientist, YouTube content creator, teacher, co-instructor. I think a lot of you have tried the e-learning course already. He has a mission to inspire and educate developers to build AI. He's also a DAB developer and entrepreneur, founder of a crowdfunding platform for developers called HABI. He developed several iOS apps, including E-Up, and then has worked on a host of open source work. Besides that, he likes to travel, he's a musician, post-modernist, and scuba diver. And then I was tempted to say hello to the world. I think I'll leave that to you. Let's put our hands together. Welcome. Thank you. Hi everybody. Testing 4876. That's the next code for Bitcoin. So buy it at that price. I'm just kidding. Yeah. The end of that description was this really old description that O'Reilly wrote when I wrote the book Decentralized Applications like four years ago, which at the time, HABI was, yes, a crowdsourcing platform that I helped create, but now it's deprecated and et cetera. You know how stuff moves in software. It's just like hearing gone in like 30 seconds. But yeah. So this talk is called Decentralized. Let me move that away. Let me just get a feel for how this is going to work. Decentralized artificial intelligence. That's the name of this talk. It's really combining two technologies that are very hot today. Deep learning and blockchain. We know about these two technologies, right? Even if you're in data science, you study blockchain at some point. There's articles. It's hard not to look at this space with some level of interest. So we're going to talk about how to combine these two technologies together to make something beautiful. So I want to start with a quote by a very well-known public figure, shall we say. The quote is, He alone who owns the youth gains the future. And this is by Adolf Hitler. I just want you to think about this quote for a second. He alone who owns the youth gains the future. What did he mean by that? Clearly, he knew what he was talking about. He was able to influence the youth to kill people by the hundreds of thousands. So I want to ask you guys this question. Who owns the youth? Single word answers. This is an open question. You can raise your hand. Who owns the youth? Facebook. Facebook, okay. Really smart guy. Any other, any other? Snapchat. Okay, you guys know what's up. Yeah, yeah, yeah. Exactly. Yeah, Facebook, Snapchat. Those are some possible answers. But let's talk about what the larger answer to this question is. Who owns the youth? Who owns the future? So in the 80s, the web was created, et cetera, et cetera, right? The worldwide web. We know that the web was created, but I remember going to CERN. CERN was in Geneva. I had a talk at CERN in Geneva, and it was amazing. Like the data scientists there, like, we were getting so technical. They knew exactly what was going on. Their questions were so beautifully orchestrated. They were challenging me. It was amazing. And those people, I mean, the amount of data that CERN has is amazing. Like, anyway, that's a tangent. When I was at CERN, I remember seeing Tim Berners-Lee's office, the creator of HTTP, the hypertext transfer protocol for the worldwide web. Arguably, without that protocol, everything that we're doing, carousel, everything about the web wouldn't exist. And I remember seeing his office and just standing there for a second and being like, wow, this is where it all began. And then I backed up for a second and I was just like, oh, my God. And all these kids started coming in. I was like, oh, I better back up. And these kids started coming in, in mass. And they were like, I don't know, eight or ten years old. It was like a school field trip. And they came in and they saw, and then the teacher was like, or the guide was like, and this is where the worldwide web was created. And the kids were like, oh, my God. And they took a Snapchat and immediately left. Some of the kids just, one of them, when the guide said that, they just kept walking. So that's who owns the future, Snapchat. Right, because the kids use Snapchat. So that was web 1.0. This man is web 2.0. Who knows who this guy is? Just shout it out. Jeff Bezos. Who is the richest guy in the world right now? Jeff Bezos. Yeah. In 1998, this was Jeff Bezos. He was selling books. That's all he wanted to do with Amazon.com. He was selling books. He had a very, very small office. His office was just a desk, right? And now Amazon is more powerful than Walmart. Its market cap is higher than Walmart. Its market cap is higher than any bookseller in the world, Barnes and Nobles, et cetera. All of these global brands, Amazon eats them up, literally, quite literally, Amazon acquires them. So Jeff Bezos owns Amazon. He's the largest shareholder in Amazon. And Amazon is just one example. We've got Snapchat. We've got Google. We've got Facebook. And these entities that we've created over time have been a result of entrepreneurs saying, how do we take web 1.0? And how do we capitalize off of it? How do we create something where we can make money off of it? Which is, it makes sense, right? This is a beautiful technology. And the initial aims for the web were beautiful. They were very pure. Let's take this decentralized internet that Tim Berners-Lee helped create. And let's create services on top of it. So let me take the server. And instead of having all of these computers talk to each other, which is very hard to do at the time, let me just make one central server. Let me use this HTTP protocol to create this get and set transfer protocol and store all of that data on this central server. Like that, right? And it made it easy to do things like have an Amazon, have an eBay, have a marketplace, have a social network, have a ride-sharing service. All this stuff that we use today is because we were able to create these centralized services. It worked really well, and there were no problems. In fact, it was a faster way to do things, right? It was a faster way. It was a more efficient way. It just worked. But what's happened now is really bad. Objectively speaking, this is not my opinion. It is objectively a bad thing. And this is not just coming from me. This is coming from the creators of the web, Tim Berners-Lee, Jerome Lanier, the pioneers of the internet, don't like what has happened today, what the internet has become. How do we fix this? What does web 3.0 look like? Not web 1.0 or 2.0, not having centralized services, but decentralized services. The original vision for the web was to have decentralized services. In a good economy, we monetize more and more, but in a bad economy, we monetize less and less. What is the most valuable resource we have today? Can anyone answer that question? What is the most valuable resource that any person has today? Knowledge. Yes, another word for knowledge is data. Data, exactly. But what are we doing with our data right now? We are giving it away for free in return for a free service, right? That's part of the terms of use for all of these services we use, Facebook's terminal service, Snapchat's terminal service. You let us use your data. We learn from your data. We can profit off of your data. We can sell your data, and in return, we'll let you use a service. That's fine, right? It works. But what's happened is this is shrinking the economy because all of that data is being concentrated in these central servers, these connected cores that everybody has to connect through. These are gateways for the web. We need that data because as automation technology gets better and better, as we learn to automate all of these things, the only agency, the only power you're going to have in this new world that we're very fast moving to is your data. But right now you're giving it away for free. So what value do you bring to the world when you don't own your data? Nothing. Zero. You have no value. If you give away your data, you have no value. So you need to own your data and profit from your data, right? This is akin to something like a basic income. When robotics and automation technology and AI and deep learning and machine learning, although that gets good enough in the next 5 to 10 years probably, everything will be automated. Everything. Everything that can be automated will be automated. I just made that up on the spot. You call it my law. There's accountability as well. So accountability is we don't know what they're doing with our data, right? First of all, we want to own our data. We want to profit off of our data. Right now we're giving it away for free and they're doing something with it and they could give it away to governments. They could use it to spy on people. I was having a conversation with a friend a few days ago. We were talking about something, a single topic. Later on, on her phone, in the ad, it said the topic that we were talking about. They are listening on our phones. They are listening in real time. Facebook is listening. Our phones, they are listening to us. They are using that data because there are no checks and balances. That's just what's happened. The Internet of Things is really a set of spy devices for these central authorities. And lastly, we are trying to solve AI. I'm going to talk about that at the very, very end, but we're trying to solve AI and we need a way to do that. And this is how we do that. All of these problems, a shrinking economy, data ownership, accountability and towards general intelligence, all of that can be solved if we move to Web 3.0. So you guys, with me, we need to move to Web 3.0, right? This is the plan. How do we do this? So I'm going to tell you how to do this. I'm going to tell you how to do this. I'm going to show you right now how to do this. You ready? Okay. Here we go. It all starts with something very, very simple. The linked list. Who here has seen the linked list before? Awesome. The linked list is a very, very simple data structure. I've got the code for it right here. Don't worry if you've never coded before. Don't worry about it. It's only 12 lines of code. 12 single lines in Python. And what this does is it creates a data structure called a linked list. And a linked list is just a very simple object that's linked to the next object. That's linked to the next object. That's linked to the next object. And it just goes on as much as you want. And we can make one in Python. We'll just say class node. This is a class. This thing, this square. And let's say, let's give it some functions. Okay, that's going to give us what's inside of here. Which is, if we could store it, the word hello. Get next. Get the next node in the list. Stack. Set data. What do I want to put in this thing? Hello, a number, an image, a video, whatever data. And set next. What's going to be the data in the next linked? The next node in the linked list. That's it. It's a simple data structure. This is going to help us move towards Web 3.0. You might be thinking, wait a second. No, this has been around since the 80s. How is this going to help with anything? So let's take this linked list. Let's take this linked list. And let's store a copy of it on, let's say, a million computers. Let's just say, hey guys, I've got this linked list. I want you to store a copy of it. And inside of this linked list, we're not going to store the word hello. That does nothing for us. Let's instead store a transaction. And not just one transaction, several transactions. In fact, millions of transactions between people. Payments. Let's store that. And let's try to make it so that no one can modify this linked list. No one can change the transactions in this linked list. How do we do that? Okay, what if we had a separate system of people and all they did was they solved random mathematical problems with their computer. And this required a lot of computing power. So much so, that in order for you to be able to change any of the data in any of these nodes, you would have to have more than 50% of the computing power of all of those nodes that are solving these random mathematical problems. Now what if the amount of computing power in those nodes was greater than the 500 fastest supercomputers in the world combined? Well then you would have to have more computing power than the 500 fastest supercomputers in the world combined to change the data in this linked list. What that means is no one could change the data in this linked list. So if we did that, we had a linked list that everybody stored a copy of that no one could change the data in. We would call it a blockchain. And that's what the Bitcoin blockchain is. It's a glorified linked list. That's it. It's a glorified linked list. And instead of calling them links or nodes, we call them blocks. And the reason it's called a chain is because it's not just pointing to a random node with a word hello in it. It's pointing to a random node with transactions in it. And the reason that the transactions are stored as this tree decision tree to be specific, we can call it a Merkle-Dag as well. We don't have to talk about that. It's because that's just a more efficient way of storing it and retrieving from it. But you can just think of it at an abstract level as just a list of transactions. And that's a blockchain. So this is a very new idea. This was 8, 9, 10 years ago when Satoshi invented this concept. It was the first time where we could say I'm going to store my transactions in this data structure. That's what the blockchain is. It's a data structure. I'm going to store my data in this thing. And no one's going to be able to modify those transactions. That's why we use banks. There are only a few real banks in the world that we trust enough to store our money in. Or something very valuable. Because we trust that they're not going to modify essentially the list of transactions in our account. They're not going to modify that. They're going to say this is the amount you had before and this is the amount you have now. Trust us. And that's it. We trust them. But what if we could store our money in a way that we didn't have to trust a bank. So there's not a third party. This is the first time that we've ever been able to do that in the advent of this data structure. That's cool. Okay. We don't need a bank. Great. So what? What else could we store in this thing besides transactions? We don't need to trust a third party to store our transactions. We don't need to trust a third party to store our data. We don't need to trust a third party to store our code. What about our web apps? What about our apps like Snapchat and Facebook and Google? What if we could store code in this thing? Not just transactions. And that is what Ethereum's blockchain is. They took Bitcoin's blockchain. They added a bunch of other data like logs, difficulty, time stamps, mixes, but really what matters is this. The code. You don't just store transactions in the Ethereum blockchain. You store code. What that means is every computer, every node, remember it's a glorified linked list, every node in the network is storing code. Instead of pushing code to a single server like Heroku, that's how we usually build web apps, we store it on a single node that everybody has to agree upon. So in order to modify that code you would have to have more computing power than the 500 fastest supercomputers in the world can buy. So this is a trustless way to build web apps and that's how we're going to get to web 3.0. So what does code on Ethereum look like? This this 17 line of code snippet is what you would call an initial coin offering. 17 lines of code is responsible for raising more money and the past two years that have been raised by every IPO in the past 10. We're talking about more than 100 billion US dollars worth of of investments from the public, the public. Every ICO that you're seeing these days is a result of this. Someone takes this as a variable that says how many shares or coins or tokens do I want and how much do I want to sell those tokens for? They take this code and they push it to the Ethereum blockchain via a push command in terminal or they could do it via a web app and then they have this public address, 25 characters or more and they say here's your address, ICO time invest and then people just you know, 30 million in a single day, this kind of stuff happens. This is very powerful and no one controls this, right? This is the advent of a new structure similar to a corporation, a new type of institution that no one controls and we can do this with Ethereum's blockchain because we don't have to trust a third party not just a bank but even a government, legal documents, things like that, we can build these with smart contracts things every time we've had to require to trust a third party but we can remove that and push that instead in the form of contracts to the Ethereum blockchain. There's a problem though. So if you think about, okay well I'm going to make a web app and it's going to pull from the weather.com API and it's just going to tell you the weather, that's it, okay? So it's called myweatherappdecentralized.com and all it displays is a single text the weather today in Singapore and it pulls that data from its API with the parameter Singapore now, it returns the result and it displays that whatever degrees Celsius. If we were to do that on the Ethereum blockchain it wouldn't work. Why? Because every single node in the network has to pull that data at the same time because they're all executing that code at the same time but the weather changes over time so they would all have to agree upon the time how do you do that? You can't, right? So the solution to this is to have what's called an oracle and the oracle is a trusted third party source that speaks to the web that the Ethereum blockchain can pull from. So you might be thinking wait a second, okay I thought this was trustless what does it deal with this oracle thing? There are networks of oracles out there delegations, four or five of them and they can all agree on they can have consensus on some data but this is a solution to pulling data from the web 2.0 in the form of APIs until we have a better solution but that's how it works right now. So you might be thinking, okay we can have our application level constructs on the Ethereum blockchain we can pull data in the form of oracles from the web but then how do we store huge databases onto the blockchain could we do that? I mean you could technically use downloading a copy of this blockchain would then have to store all of that data on their personal computer. Facebook has a lot of data if we were to decentralize Facebook and say this is the Ethereum version of Facebook everybody would have to store a copy of Facebook, that's just infeasible so what we need is not just an application level construct that says here are usernames here are tweets et cetera, we need it to point to some kind of decentralized data store a storage for data that is not owned by anyone. Now the best solution for this that I found is called the interplanetary file system so IPFS is big, there's a lot to go into about IPFS it's a beautiful system the way it works but essentially it's a group of computers that all agree on some data so if I want to push some data to Google Cloud or AWS right now, the services own that data so I would say this is a huge video let me push it to Google Cloud or AWS and they would store it how do we decentralize that with IPFS it's a network of computers who are saying I'm an IPFS node when you push that video to IPFS not everybody stores a copy of that video, one person will store a copy that video will be split up into several parts and replicated across the network so then you might be thinking well what if someone deletes their node well there's already replications everywhere how are these nodes incentivized well there's this idea of Filecoin where these nodes are paid to store this data so it's how to incentivize these nodes so that's kind of our complete list of solutions for this Web 3.0 we can have a data store called IPFS we can have Ethereum to store code and application level constructs and oracles to pull data from the web this is how Web 3.0 works this is the new stack so this is how it works Ethereum references content big data that's stored on IPFS it gets that static content the distributed application gets that content from IPFS and it gets the global state or Ethereum it's a global state machine so if we combine IPFS and Ethereum we can build distributed decentralized applications but what's a talk what is really a talk without talking about deep learning deep learning is where it's at we've got this thing called a neuron in our brain there's cells billions of these cells in our brain and they make us intelligent they give us consciousness and love and emotion all these things we feel it because of this thing it's a collection of them I don't know how they work neither do you, neither does nobody neither does the most renowned neuroscientist in the world no one knows how this thing works but in the 50s some guy was like I mean I get the basic idea there's some data that's being fed into this thing something's happening here and then it sends it out and then very roughly very very roughly he created a mathematical model for this thing he says he said we have these inputs let me just take these inputs these are a set of numbers let me apply some kind of summation to this and then some kind of non-linearity as in let me take this data and make sure that it can learn a function that can learn any type of function both a linear and a non-linear function that means any type of function and then output a result so it takes some data it does something to it and then it outputs that data and what happens here has been the topic of debate amongst many researchers like what do we do to it but the idea is the same where we take some input data apply some kind of function to it and then we have an output we have a lot of them and what happened is the deep learning revolution all of these things that humans are good at image recognition driving cars etc all these things for some reason this model that resembles the human brain works well sometimes even better than we do so this is kind of this is mind boggling on silicon these chips we roughly estimated the rules that are happening here is doing something that even humans can't even do could there be some universal law of intelligence out there that we're just slowly approximating that we're just slowly coming close to probably and this is our first step there deep learning incredible technology ok so how do we combine all three of these things how do we combine ipfs, ethereum and deep learning well there's one really great example that I know of called open mind that a friend of mine has built with his community if I'm a data scientist and you have some data if I wanna train my deep learning model on your data how do I do that in a way that involves no third party where you don't have to trust me that I'm gonna look at your data say it's very sensitive like patient data like for hospital records how do I train my model to learn from your data so I can make predictions about say cancer or some other disease without looking at your personal information so this is a perfect example of how you can put these three technologies together you can have ipfs store that data you can have ethereum build an application level construct for smart contracts so that data doesn't have to be stored on a central server and you could use deep learning to learn from that data so if you wanna know how all three of these can work together look up openminds, the website is openmind.org but it's a great example of how to combine all these three things together so where does this lead us eventually this is a graph of emergent higher level complexity simple ai agents that are ants but complex swarm behavior emerges right now we have ai and they live on servers they live in our phones, they live on laptops but if we could free these ai's from the confines of servers and let them live in a decentralized space they could speak to each other they could learn from each other and perhaps the real solution to higher level intelligence is to let that complexity emerge rather than trying to code it in why don't we just let it emerge right, we don't know how consciousness works but we have this basic rough idea of how it works, what if we just need to give it enough complexity and let that real intelligence emerge we just can't even understand it that is very likely the case and if we use blockchains and we use ethereum and we use ipfs and all these decentralized technologies we can create real artificial intelligences that are as smart if not 100,000 times smarter than all of us lastly you might be thinking why would we want to do that why do I care about that right water crises, the spread of infectious diseases weapons of mass destruction North Korea Donald Trump, interstate conflict failure of climate change adaption, energy price shock fiscal crises unemployment, biodiversity loss and ecosystem collapse it's too late for us to solve climate change it's too late, no matter what we do we're not going to solve it we are headed for destruction we are headed for oblivion there is nothing we can do we're screwed, I don't think so because those people don't know about the power of AI if we solve AI let's just say we solve it, we create an intelligence that's as smart if not smarter than all of us we can give it an objective a goal, a function that could be solve climate change it would look at data in a way no human ever could in a way that 10,000 plus scientists put together in a room for 100 years couldn't solve it could solve all of these things immediately it could move us towards this utopia that we didn't think even possible before if we do it the right way and if all of us together are thinking about ways to do this we all have to be thinking about how these technologies work together if you even barely understand how this technology works it is your responsibility to do something about it because there are millions of people across the world who don't even have food and water in Maslow's hierarchy of needs they don't have any of these things how can they come to these kinds of insights they can't everybody in this room it's your responsibility to work on this to help make this happen, how do you do that you learn from the internet you use this new university that we have this collective university to educate yourself on how these technologies work and how you can best influence the problems in our world using these technologies so I want to leave you with a single remark solve AI or die trying that's my motto, that's what I live by and if you want to learn more there's my YouTube channel, subscribe thank you very much thank you Suresh we have 5 minutes Q&A are there any questions I think that's only one I can check out the question is the example that you saw that we can with IBM we can have a secure data access on that so the algorithm can see why not the user that is accessing that data can see so how the security is going to work security is definitely an issue we have to solve when it comes to decentralized applications part of the reason that we centralized stuff was because one person could just secure everything it's much more secure to have data on a central server security is one really really great area to focus on if you're looking on something to work on that's one of those open fields there's so much a startup could do there distributed security, decentralized security in terms of a solution right now for security in general for decentralized applications one thing could be better ways of storing data that involve replication so replication is one example a lot of times data is just lost in IPFS or something like that it's just lost because there's not enough replication so looking at different ways different algorithms for replicating data that are efficient that would be a great way to solve security that would solve the update security nobody can modify it but if you want to provide the security on read access like I'm not going to modify it that's secure enough because it hasn't saved on multiple places but if I want to secure the read access that whoever is authorized and maybe authenticated already that can only access the is it like medium already provided or we have to build a security layer on top of that we definitely have to build an extra security layer on top of it like there are ideas within these communities for IPFS and Ethereum for security layers but we would need a different protocol entirely like one of these issues one of the ways that this brings up is identity we don't know who you are we don't know how to authenticate you identity protocols is a layer on top of these data and application level protocols that would be something that's very necessary people have tried this Microsoft even tried this with one ID it didn't work but the problem was they weren't thinking about it the right way how do I make an identity protocol that can speak to these other protocols seamlessly and there's this idea of cross-chain transfers atomic swaps from the Bitcoin community I think that would be a good place to look for inspiration Hi so I was wondering what is the algorithmic scaling for the full homomorphic encryption because if I'm not mistaken it's pretty large and on top of that most of the deep learning frameworks the deep learning models tend to be computational halt yeah so two questions are one how do we scale homomorphic encryption and two how do we solve the problem of deep learning being very computationally expensive yeah together because just increase the scale if you're going to do homomorphic encryption on a deep learning system it just goes crazy yeah definitely yeah you know what it is so there's this idea of one-shot learning from one or two examples like if you look at how humans learn you give me an image of something that I've never seen before and in two or three examples I'll be able to classify it whereas with deep learning it requires hundreds of thousands if not millions of images this is not even close to how we learn I think the best example of a very data efficient deep learning algorithm is it actually came out of Facebook and it was called the memory augmented neural network so right now we intertwine the processor with the memory in the neural network that's the weight values right those matrix weight values but what if we could separate that memory into a separate module and it could say here's the memory module it's just a glorified matrix and then we have the weight values well what happened when they trained this thing was it was able to recognize handwritten digits with only 15 examples this was like two years ago and then that project didn't get that much attention it should have because that was amazing I would say look at that project as an example Facebook AI research omni-glot dataset it's going to be the first link on Google but more efficient deep learning models is another open field of research I have two questions so the first question is I was wondering how can we as someone who is learning data science you know adapt to the changes in terms of if you are looking at a different perspective so currently for example we are using some of the typical activation functions which mathematically speaking you know like sigmoids and all these are mathematical models so if there's any mathematical evolution such that you know in the future perhaps it could be more than just that human brain we think that there will be a shift in terms of how we will get neural network or even the whole entire energy and space and the second question is I'm just curious as to you know AI these days they have robots that can jump literally like a terminator so in terms of cyber security I'm not sure how well it is but what if one day it happens to be at the hands of maybe someone who is extremist or terrorist say for example will we all be doomed because the weapons everything can be controlled and these robots might just go it might be to hear some of the robots talking about humans in a very negative way we've seen some the one I can't remember the name the lady who has been so there was something negative about the way she thinks about humans so I mean what if one day robots behave like this and they learn how to manufacture the mass weapon of destruction or whichever way they can behave like a terminator for example I don't know how would it become so these are my good questions yeah ok good questions so for the first question about different ways of looking at neural network in the future beyond the model that we have right now this is one of the most exciting areas of research in all of computer science right now how do we evolve the neural network architecture there's this idea of well let me just stack more layers let me change the weights let me add a new type of layer and these are ok this works but what if we change the entire hardware what if we said instead of running it on silicon let's run it on a quantum processor what happens then there's this idea of neuroscience of quantum nanotubules that's the seat of consciousness that actually lives in multiple dimensions who knows but there's a lot of great research in hardware as well so if you want to look at really really really different radical breakthroughs in neural networks look at the hardware not the algorithmic level on to your second question about Terminator and AI some people say that the bitcoin blockchain is some people say that bitcoin is one of those systems where you can sell drugs easily and you can but it's one of those systems that is very untraceable right this is not true bitcoin is the most traceable system in the world key example silk road a single transaction is stored in a public way that everyone can can view right they can view it they can validate it they can see exactly the trail of transactions now what happens if we take these transactions and instead we abstract them to to actions in a network this person liked this tweet this person did this with this money this person launched this nuke if we put everything on to this public ledger that everyone can view that's a way to democratize read access to actions by everyone so if there was some kind of government right for example probably the Singaporean government would do something like this because they're awesome where they said let me just make everything public and viewable by people right then every politician would have to record all their transactions on a blockchain right there would be no more there would be much less corruption there would be much less right because this thing this would be validated by everybody when it comes to nuclear weapons weapons of mass destruction if we are able to create these systems in a way that everyone can verify there's more transparency there would be less likely of a risk like that right because these people would have to be held accountable and that's really where this whole blockchain stuff plays in it's all about accountability do you think that with all the work we've done on proof of stake and blockchain 3.0 there might be a better technology than Ethereum coming off as soon yeah I do I do there are some promising blockchains out there I think one is Cardano right Charles Hoskinson was one of the former founders of Ethereum and he's been working on that with a really great team they're working on KYC compliance as a way to merge that with the existing financial infrastructure it's great I think the idea is that they have our greats and yes there could be something better proof of work for sure I mean proof of work is one of those really computationally expensive processes that wastes a lot of electricity across the world that we could put that could be put to better use why have these computers solved random mathematical problems why not instead have them solve protein folding or some real problem that matters so yes there's a lot of potential right now however in terms of building something that affects the population that affects consciousness that affects culture Ethereum and there are ways of doing that very fast and then we can update those systems later on one question do you think that reinforcement learning will work as in right now what I've seen so far is only on games chairs do you think you have more practical purpose like for example the other classifier that we can do something in the real world other than just playing games yeah right yes I do the idea behind reinforcement learning has been around since the 60s it's not new all of those concepts that AlphaGo used last year AlphaGo was it beat the best Go player in the world a little refresher and which was a huge deal because AI experts predicted that it would take 10 20 years for that to happen we did it in a decade less all of the components of AlphaGo Monte Carlo tree search right value networks policy networks neural networks these all had all been around for decades what was the difference then well they just synthesized some ideas that hadn't been put together and fed it more computing power than we'd ever been able to feed it before and then it gave us this result that no one thought possible so the real breakthrough there was just giving it computing power there was no difference in terms of the algorithm just a synthesis of old ideas into one bigger new idea which works to be real do I think that we could do better for sure I don't see why we aren't thinking of new ideas for reinforcement learning there's so much we could do in terms of search strategies for reinforcement learning there's so much we could do when we combine deep learning and reinforcement learning together games have been a test bed for AI because if we can generalize to different games maybe we could generalize to tasks that matter that's why these big institutions that have been focused on games is applied to real world applications some examples would be cooling data centers that's what Google did with the same reinforcement learning algorithm they use on a game they use to reduce the or increase the efficiency of cooling their data centers but I think it was 40 or 50% which is huge that's one example and that's just because Google did it guess what Google open sourced that algorithm if we look on github and we see these reinforcement learning algorithms we ourselves can take them and apply them in the news case while these researchers at the top institutions are working on games because they're trying to improve the generalization ability we can use them to solve problems today so we can take those algorithms we can use them to solve real problems some avenues like maybe three that I can think of for reinforcement learning specifically would be for example traffic control that's a reinforcement learning problem reward using an agent to take in reward cycling that back into the system another one would be increasing the efficiency of bandwidth between applications like peer-to-peer applications that's another one for reinforcement learning and a third one I'm just making these up as I go a third one would probably be you know how Google did this with the data center we could do that with just lights in general energy efficiency that would be another one for reinforcement learning all these ideas were we were talking about some event some timeline something that's happening over time that's a reinforcement learning problem because it's interacting with an environment it's not an input-output static data set it's a dynamic environment so think about an environment, think about time think about sequences and apply reinforcement learning there thank you we mentioned earlier about how we can have individual simple AI blockchain we can all communicate which are some sort of emergent behavior and that's the next step what if we don't like the emergent behavior because you can't control swan behavior right? like answer the colony individually you know what you do but you don't know how to behave in this system so what then? that's a great question so I met someone who worked on the ethics team and she was telling me that she was kind of going through this game with me she was saying well why should we release this source code why should we do this but the conclusion that she made me come up with which is what the conclusion that she had already come to the thought process was that DeepMind wanted to be the first to come up with the real intelligence they wanted to be the first to solve AI because they knew that they themselves would be benevolent with it they didn't trust that other people would so that's why they didn't open source it at the time however that's not good for us because if someone at DeepMind is not benevolent for whatever reason they could use that AI for evil by whatever your means of evil is how do we prevent that the best way to prevent any kind of malpractice any kind of mal-intent is to democratize this technology to give it to as many people as possible there's this famous quote by this English playwright Lord Acton also by Elon Musk but the quote is the best way and I'm paraphrasing the best way to prevent the rise of a despot someone who is an all-powerful dictator in the easiest ways to centralize that power so if we decentralize this power to AI, to blockchain, to all this technology to data specifically then we can better increase the probability or likelihood that we reach a good state in the future is there a possibility that someone who is a bad actor will use this technology? for sure I mean life is full of existential risks but doing this democratizing these technologies further increases the probability that it will happen because we'll come up with some sort of security system or fail-safe or check system that would prevent that from happening maybe there's a decentralized AI that goes rogue well there might be a decentralized security system that prevents that that someone created beforehand we just all have to be thinking about these things it's a good thought that's all I'm always grateful for all those speakers that shared their time for free this time around Shirok was coming here so let's put our hands together thank you Lucas do you need to say anything? Lucas do you need to say anything? okay I mean okay whatever we'll be here for a while for networking and stuff like that so thank you very much for coming