 I studied computer science. I took courses in AI and machine learning. I understand what machine learning is. I don't understand what AI is. So, I don't know, maybe you want to example it? Sure. I think AI will say, I mean, one of the things that I like the most about AI is getting a machine to do things that only people can do. So, I'm a shifting definition. How do you... Choose it. How do you see that? Machine learning is teaching a machine how to perform certain tasks or activities, the way a human would do it, or not as a human, but teaching a machine how to do something, right? And you generally can't teach... You can only... You can write an algorithm, you can write an algorithm that would do one thing, but if you wanted the machine to learn by itself, which is what machine learning implies, you start feeding it data, and then based on the base algorithm, it tries to understand how to use that data to make better decisions. So, what is AI? Because I hear a lot... It's a problem, sir. It's a... Machine learning is only one branch inside the whole area of AI. So, what else? You can say AI is automation. To some extent, it's not wrong. It's just to automate things. That people have been doing since the beginning, but now we're able to make it automated. So, that's in a very broad sense, but AI is... Internet also doing it, right? So, the problem is learning could be that... Wikipedia says it. A machine mimics cognitive functions such as learning and problem solving. So, I guess you're right. It's a very broad area. It covers a lot of things. It could be that more like brute force. Things like AlphaGo or Deep Blue just throw a lot of machines, right? It's more like brute force. It's not really like super intelligent, right? It's just like trying to use the fact that you can write many cycles of machines faster. That's another approach. I think we know the example of what Microsoft did with the bots. The Twitter bot went really badly online. I don't know if you guys read that. They still don't say it's artificial intelligence. It's not well-executed, but it's still... There was some learning. I guess you don't trust the internet to give you input here on learning and problem solving. I've had some experience writing a Twitter bot as well, just as a hobby. The thing about Twitter is that they try to catch bots that pretend not to be bots. So there's a team that finds fake accounts and they delete them. So one of the hobbies I have is to... I don't eat pads, but I keep them to the bots. So I write to governors that try to trick people to think that it's an actual person tweeting. So you can see there's something called intelligence as well. Mimicking intelligence. It goes back to the Turing test. Tend to be human in a conversation. So some technologists don't perform as easily to pretend to mimic a human to pretend because it's just text. Facebook is a bit more difficult because there's so many more mediums of sharing. There's more conversations going on in Facebook as well. So it's going to level me first. Are there any mediums of sharing? I'm just going to say that... One example. The way we use the phrase AI is like there's like all the shop. There's no like governors or the term. So it's difficult to put a box around it because firstly people use it. Salesmen might use it to sell something. And a researcher might recognize over what's in it and what's not in it. And depending on which researcher you go to, they will cut the domain into new ways. I was wondering whether just to put up as like a starting point like creating like concentric circles. And now I'm right in the middle. And context of the set you have AI. And context of the set you have AI. That you might talk about what's in one and not in the other. So I'm not sure about that. Does anyone know how the actual sharing test competition works? No, I have no idea. Well, I don't think it's a competition. It's still going on. It's still every day, every year. There is still a competition. I know the mechanics is... They get a test with people. Correct. With machines. Then you guess, and you guess people as well. You guess which are people. Yeah, so that's the important thing is that you have people who are testing. And on the other side there will be machines and there will be humans. But the interface is the same, right? Yeah. But from year to year they're like, how long do you spend? What type of questions you can ask? This can change. The funny thing about this competition is there are actually two of works. The one team wins. The one, the computer system which wins them. Which fools the most humans. Right? That's the best part. And the human who gets vote at least as a machine. Because there's another possibility, right? You are interacting with a human. But you may be marked as a... You may have to make a decision. Are you talking to a computer or are you talking to a human? So humans get marked as machines. So there are two words at best. The most human computer and the most human human. Wow. So when you're talking about thoughts, right? At TraderBot. So how do you... Alright, just getting a little more of the next level of that. How do you train the bot to act like they do? What are you doing? Right. So what I did initially was I searched for tweets online. And I threw them to database and I changed around the words to make the tweet look legitimate and not a copy. But later on I found certain language processes online where you can create insult things a lot. So I changed the way the engine works when I created something. So it's just English structure based on certain trends that's happening in Twitter, things like that. So you end up with a bit like a soft bot looking like a teen with a lot of anger issues. I don't know about that. That seems to be... But I don't take input from the internet. So there are different ways to do that I guess. But learning is the... What Microsoft tried to do is probably the impacts of everything you learn. But I guess it didn't execute it properly because it just took input from anyone. I think there was even a 4th century that just attacked the bot about all kinds of strange things. So it depends who you open way to get your data from. So is your bot learning on the go? Currently... Currently no, it's not learning on the go. I'm not letting it just take input from anyone. I'm collecting inputs and putting it in. So where would that, on a spectrum of AI versus machine learning, call like, is it... Because it's not learning, does it imply that it's just like an artificial intelligence? Because it's not learning, so then is that machine learning already being a bit separate? It's fancy I guess. That machine is a bit... Between machine learning and... I wouldn't call it machine learning yet. Maybe just control learning or something. I don't know whether that's the thing. Because essentially you've set its parameters, right? You're saying that... I mean it's very unlike a human where a human is taking input from the outside world, processing it and then trying to change. Whereas this one is very parameterized where you're saying you only can do this and this and there is no other way. It can't act on its... Well it does like on trends for example. So trending terms, one of the people is saying and trying to throw that into a language processor. So that is the part where it gets a bit of learning? A little bit. Does your language processors do anything with measuring emotions? Emotions? No. I do have certain guidelines. I try not to... I prevent the bots from tweeting. Because you'd be surprised how many strange words there are which you don't want the bots to be saying. Do you apply sentiment analysis before you post? Do I apply what? Sentiment analysis or something? I'm looking into trying to do that. Something to... Emotional, trying some emotions. James Penny Baker. He's done some research to pronounce and how it reveals certain emotions. Who's that again? James Penny Baker. James Penny Baker. Does anyone watch this series called Personal Thinkers? Yeah. It's a very interesting series that is run by Jonathan Nolan. Christopher Nolan's brother writes most of the other movies as well. It's essentially an AI system that I don't know, is very classified right now. But it's an artificial intelligence that takes data from all government feeds. And then, based on that, it can move its own location if it feels it's under threat. By that, it means that it could issue commands to people as if it was issued by a woman to move its location. To some extent, that's what the person is saying. It's a learning machine, so it essentially runs a lot of simulations. And based on that, it could tell you what is the more probabilistic positive outcome or negative outcome. I use this big data. It collects all the inputs from all the security cameras. So in that series, they don't explain any of the machines working, but in the initial stages of the machine, there is a part where the machine essentially learns emotions based on its administrators' past events. Has anyone got any experience in how does that kind of stuff work, where you kind of get a machine to understand emotion or empathize, right? So the machine essentially finds out that on this particular day, that man's father had died. So it kind of sympathizes with them and says, oh, I'm sorry for that, right? And that logic is kind of not present, because... I think humans are really complex creatures, so understanding emotions. Our emotions change even hour to hour sometimes. Yes, it's very... based on the conditions of the events, like the day you can suddenly decide if you hear something badly, so... You might as well be yours, and you should see your friends. Yes, while driving, there's an accident in the next car, and you see the police or the next car, and suddenly you become a man's officer. We don't need any form of biofeedback from a human, so it's really difficult, I guess, from just... maybe facial recognition could come up, I guess. Since sometimes fiction movies are a little bit of a present space, and we try to get an emotional or emotional profile of the person. You could also analyze the text strings, for example, tweets even from someone writing a tweet, and then you could measure some emotion and get some traits. Have you ever... Going back to your... to your algorithm that you wrote, have you ever taken into account the response of your tweets and tried to... Adjust your... Not much time I spent on this weekend, so I'm actually looking at... looking at the emotional aspect of it. Right now, I'm just trying not to get bad to put it down. I have like two plots, about a thousand to two thousand followers each. Are you sure all of them are... Are you sure all of them are plots? People are like experimenting with plots. So there won't be... But they've been online for a few years already. Yeah, so plots following other plots, and then you become just like... I mean, the plots are actually in some of the year, and then you are experimenting with plots. Like, as we say, you know, how much you are. And like a thousand, like... Like a thousand, like... Like a thousand, like... I'm just going to pick up on the point that you mentioned on looking at the words yourself. I think one area is in AI, like, probably in machine learning, you'll use a lot of supplies and super-buyers and in the AI limbo that falls under things like reasoning, things like learning and things like problem solving. These are the words that guys use in AI to kind of equivalent the machine learning to you. But beyond that, there are all these other things which are a bit funny. One is actually knowledge representation. So I think I want to actually mark this as a motion, like, very much. I was quite interested in how the way you represent like a chessboard to a program. What is it? Is it like, you know, 64 columns or like, category variable? Is there something more? So knowledge representation is an interesting one that is kind of within ML maybe not within ML. The one that's within ML and maybe not within ML is perception. And again, when we talk about things like imaging around that, it's powered by machine learning but for a program to know that it's hard. And the same program to also know that I'm near to a wall and the same program to also know that I'm going to do this thing because I'm near to a wall. That perception thing is something that is peace beyond the machine learning project but it's been all in AI. So I guess what I want to say is the knowledge representation of the world is very, very important. And now I'm going to a space that I've not done before around my own excess. When you use natural language processing and all these space represent words you want to see whether this is a happy, is this to the guy happy? Traditionally what you would do is throw in a bunch of dictionaries, these are the words, ecstatically it could be plus five. I'm only happy with plus two and okay, like terrible between things and I'm okay. This is a happy because it's more positive but another way to represent the same data is a word to back sense of that word to back that way where basically every word is represented by a black box of concepts. So if you look at what does the word when in word to back representation you'll find a matrix of like hundreds of numbers which we don't know what it is. And the reason why this is the case is because a man has a gender flavor to it. A man has a person flavor to it and some things all these things are encapsulated is we are representing the word man. So my colleague actually did a demo where we use king minus man equals queen king minus man equals queen. So I think representing data is at the heart of the data we need to make more intelligent things that are possible to make out of this data. And so I think to make a board recognize a motion of recognize meaning, the concept of meaning will really be around the way we use. I think the current record for the most complicated word is word to break. Like if I'm going to take a break, I broke this thing, break credit, give thanks, whatever. I think there's like 70 different meanings to the word to break and it's because of many meanings and characters they want word therapy to represent words in a way that's more complicated than the word. Man was one of the words that I came across on Twitter. He told me that very often man is hot. Old man. Like what? It's very interesting. So each word there's a whole matrix of data and then king minus man. So I think those of you who are interested as a whole one is every word is represented that's a word to break. And the other one is the more difficult one where every sentence is represented where it's like oh man if man thinks oh it's probably not real and you know that kind of thing. So this bit I'm not sure if I'm able to link to it. It's also a menu. From the take away from it is English is probably not the best one to do that in a setting. Is there any kind of what kind of line edge is best for a machine learning? Of course, yeah. Not really. The chess example that you're discussing and the probabilistic values that it gives does the machine remember or is it does it do it? Does it remember all its calculations that it does so that it can lose that in a future scenario? That's one. Second, is it efficient to do that? Definitely yes. One of the things that you do is you make moves within one game. This is what I'll be talking about. We talk about playing games and you make moves. You might move your movement. So after your move it's going to be my move again. I remember what I explored and the possibility space that I explored in my previous move I could build more on top of that rather than clean state every time. Is the probabilistic number changed? The number is 0.6, 0.7, 0.9 So it changes that and it keeps record of that over time. So that's how a single kind of body grows in the sense in the sense of exploring more possibilities as you play longer and longer. It's learning. It's the way I think the humans would learn. Although us as humans for the body it's easier to do it. As humans if you make a move you remember that move whether it worked or not but you would not have a set that you had thought about all these 5 other moves which the next time you play a game and you find a similar scenario you would simulate that one move. The machine says it's actually very simple statistics. It's just counting how many of the future games I might want. That's it. I mean in a very simple way if you just do that it would work pretty well. Just count all the futures I tried out 5% I have won. It's probably not a good place to be. 5% I have lost. It's just simple statistics. So it's like humans produce map of that. But in your case there was like trillions and trillions of possibilities. How do you integrate all of them? I think I can talk about that. So the idea is to get back to sampling. So we can't do all the trillions of possibilities that we over time will sample the space. Remember how many times we did this. Wouldn't you limit I am just actually really wrong. Wouldn't you limit say if there were 10 minutes to 46 I think right? Yes. If you limit to a certain number and get different machines to run different numbers and then go against each other where all the possibilities are mapped so if they all say if a certain state had 4 possible futures in Ireland you run against each of them and since it's a machine the other side could run the other possibilities in Ireland and then keep a record of all that which could then be collated into a single machine. Is that how it works? They do exactly that. What's why you have a cluster of machines? Because it's randomly random futures Each of the machines could look at a different random future and you aggregate the statistics and you do exactly that. I didn't I learned something. That's what I wanted to learn. That's how much things work. What happens in scenarios where there are defendants how do they resolve this? What do you mean by that? Like say for example there's a scenario where so the variables are not independent right? The variables are dependent where if one variable did something it would affect the other variable or the output of the other variable then something else would change. How do do scientists or researchers start to solve these problems? It's too hard of a mathematical problem or computational problem that it's worth making just guesses of certain things and running with them. There are many ways to paralyze or distribute the computations that we can utilize many machines so I think one of the ways you can do it is just do it at the very top where the very first move we're going to make and that's clearly going to be independent because after the risk is off we try to do it less than inside I guess it's more complicated because you say if I take another move with mine Yes so for example a certain example where say there's the idea of independent variables and variables and there's some curve around it like the idea was that a grandmother has finds a lady has 500 and she gives them every year a chocolate or on New York now she needs to decide out of 10 chocolates she knows from some past data that out of these 10 chocolates every child likes a certain chocolate but then there is a dependency so if the first child takes a red chocolate the second child because his tendency is to take whatever the first child takes he takes a red chocolate so trying to guess what is the best allocation of chocolate that needs to be there there are too many possibilities so what happens in those scenarios like chess is still a confined problem in this case because the dependent variables are changing and it seems very much dependent how do people look at such problems I think that the traditional technique there is always just to look at the part that are independent and try to break it along those lines I mean even look at probability, distribution of the multiple variables such a very complex form because it's going to be, you need a lot of numbers actually represent the distribution of the 10 variables for example each of the combination of 10 variables they have one probability so actually you need a very large even to represent this distribution it's very complicated so what people do is probably these three are dependent and these other three these two groups are not dependent so you can factor two separate distributions which are half the size in some sense and then if you get the joint distribution just multiply because it doesn't depend if it's a multiplication so the problem comes in that you just break it into smaller subgroups that are usually as independent as possible you figure out where are the groups that are independent and leave the dependent ones alone because they are dependent but you can't break it further I think this question is really an idea of a habit so quite long into it but I think I see like two approaches to approach the problem and it depends on who you are so within I guess like some of the size of people there's like almost like two religions one of the people who have gone up through the mathematical modeling and statistical modeling and when they see dependent variables they will use their intuition they will do interviews and try to explicitly account for all these things so you have like 10 variables plus a whole bunch of log variables and you will hand code these things like I will account for all these things that I expected there are schools of thought which I work via the the comm science and now kind of just like you meet the testers with black box I don't care how it works but it works better than it do so hard and so they will just like start that I won't use anything linear I'll use extremely non-linear thing like how I was drawing new network algorithms that would code up its own features and I don't even know what features they're coding up but you can account for dependencies, multiple entities all these things I don't know which ones they're counting for or how but they by choosing the right algorithm and putting in a really how kind of so depends on who you are I've seen both of these things go in parallel and the stats points are actually pretty good but they are understandable like you know yeah I've seen both of these marks and I got semi-accessible you know what I tell you what's going on that guy is like thank you but I don't care and every day you get me through guys like but it's a very classic example from statistics for classification it looks like spam filtering maybe it's an email spam or not and then the early days the base so that example I was trying to like this where I was trying to understand NaE base there actually is to just make a simplified assumption pretend that everything is independent even though they are dependent in reality but they just pretend they're independent and it looks really well actually even though it's not the assumption is wrong but I think we still don't fully understand why that's the case but don't just ignore the problem pretend it doesn't exist then move along and see but it still works out in me sorry we have to go okay we have a lightning dog scenario thanks a lot