1 00:00:00,000 --> 00:00:09,044 *preroll music* 2 00:00:09,044 --> 00:00:14,049 Herald: Our next talk is going to be about AI and it's going to be about proper AI. 3 00:00:14,049 --> 00:00:17,730 It's not going to be about deep learning or buzz word bingo. 4 00:00:17,730 --> 00:00:22,590 It's going to be about actual psychology. It's going to be about computational metapsychology. 5 00:00:22,590 --> 00:00:25,750 And now please welcome Joscha! 6 00:00:25,750 --> 00:00:33,050 *applause* 7 00:00:33,050 --> 00:00:35,620 Joscha: Thank you. 8 00:00:35,620 --> 00:00:37,710 I'm interested in understanding how the mind works, 9 00:00:37,710 --> 00:00:42,640 and I believe that the most foolproof perspective at looking ... of looking at minds is to understand 10 00:00:42,640 --> 00:00:46,600 that they are systems that if you saw patterns at them you find meaning. 11 00:00:46,600 --> 00:00:51,700 And you find meaning in those in very particular ways and this is what makes us who we are. 12 00:00:51,700 --> 00:00:55,239 So they way to study and understand who we are in my understanding is 13 00:00:55,239 --> 00:01:01,149 to build models of information processing that constitutes our minds. 14 00:01:01,149 --> 00:01:05,640 Last year about the same time, I've answered the four big questions of philosophy: 15 00:01:05,640 --> 00:01:08,510 "Whats the nature of reality?", "What can be known?", "Who are we?", 16 00:01:08,510 --> 00:01:14,650 "What should we do?" So now, how can I top this? 17 00:01:14,650 --> 00:01:18,720 *applause* 18 00:01:18,720 --> 00:01:22,849 I'm going to give you the drama that divided a planet. 19 00:01:22,849 --> 00:01:26,470 Some of a very, very big events, that happened in the course of last year, 20 00:01:26,470 --> 00:01:30,080 so I couldn't tell you about it before. 21 00:01:30,080 --> 00:01:38,489 What color is the dress *laughs**applause* 22 00:01:38,489 --> 00:01:44,720 I mean ahmm... If you have.. do not have any mental defects you can clearly see it's white 23 00:01:44,720 --> 00:01:46,550 and gold. Right? 24 00:01:46,550 --> 00:01:48,720 [voices from audience] 25 00:01:48,720 --> 00:01:53,009 Turns out, ehmm.. most people seem to have mental defects and say it is blue and black. 26 00:01:53,009 --> 00:01:57,500 I have no idea why. Well Ok, I have an idea, why that is the case. 27 00:01:57,500 --> 00:02:01,170 Ehmm, I guess that you got too, it has to do with color renormalization 28 00:02:01,170 --> 00:02:04,720 and color renormalization happens differently apparently in different people. 29 00:02:04,720 --> 00:02:09,000 So we have different wireing to renormalize the white balance. 30 00:02:09,000 --> 00:02:12,650 And it seems to work in real world situations in pretty much the same way, 31 00:02:12,650 --> 00:02:18,000 but not necessarily for photographs. Which have only very small fringe around them, 32 00:02:18,000 --> 00:02:20,600 which gives you hint about the lighting situation. 33 00:02:20,600 --> 00:02:27,000 And that's why you get this huge divergencies, which is amazing! 34 00:02:27,000 --> 00:02:29,660 So what we see that our minds can not know 35 00:02:29,660 --> 00:02:33,250 objective truths in any way. Outside of mathematics. 36 00:02:33,250 --> 00:02:36,340 They can generate meaning though. 37 00:02:36,340 --> 00:02:38,760 How does this work? 38 00:02:38,760 --> 00:02:42,010 I did robotic soccer for a while, and there you have the situation, 39 00:02:42,010 --> 00:02:45,150 that you have a bunch of robots, that are situated on a playing field. 40 00:02:45,150 --> 00:02:48,480 And they have a model of what goes on in the playing field. 41 00:02:48,480 --> 00:02:52,050 Physics generates data for their sensors. They read the bits of the sensors. 42 00:02:52,050 --> 00:02:55,900 And then they use them to.. erghmm update the world model. 43 00:02:55,900 --> 00:02:59,020 And sometimes we didn't want to take the whole playing field along, 44 00:02:59,020 --> 00:03:03,380 and the physical robots, because they are expensive and heavy and so on. 45 00:03:03,380 --> 00:03:06,480 Instead if you just want to improve the learning and the game play of the robots 46 00:03:06,480 --> 00:03:07,800 you can use the simulations. 47 00:03:07,800 --> 00:03:11,200 So we've wrote a computer simulation of the playing field and the physics, and so on, 48 00:03:11,200 --> 00:03:15,210 that generates pretty some the same data, and put the robot mind into the simulator 49 00:03:15,210 --> 00:03:17,040 robot body, and it works just as well. 50 00:03:17,040 --> 00:03:20,590 That is, if you the robot, because you can not know the difference if you are the robot. 51 00:03:20,590 --> 00:03:24,460 You can not know what's out there. The only thing that you get to see is what is the structure 52 00:03:24,460 --> 00:03:27,530 of the data at you system bit interface. 53 00:03:27,530 --> 00:03:30,090 And then you can derive model from this. 54 00:03:30,090 --> 00:03:32,960 And this is pretty much the situation that we are in. 55 00:03:32,960 --> 00:03:38,180 That is, we are minds that are somehow computational, 56 00:03:38,180 --> 00:03:40,700 they are able to find regularity in patterns, 57 00:03:40,700 --> 00:03:44,530 and they are... we.. seem to have access to something that is full of regularity, 58 00:03:44,530 --> 00:03:46,630 so we can make sense out of it. 59 00:03:46,630 --> 00:03:48,930 [ghulp, ghulp] 60 00:03:48,930 --> 00:03:52,800 Now, if you discover that you are in the same situation as these robots, 61 00:03:52,800 --> 00:03:56,180 basically you discover that you are some kind of apparently biological robot, 62 00:03:56,180 --> 00:03:58,530 that doesn't have direct access to the world of concepts. 63 00:03:58,530 --> 00:04:02,140 That has never actually seen matter and energy and other people. 64 00:04:02,140 --> 00:04:04,890 All it got to see was little bits of information, 65 00:04:04,890 --> 00:04:06,270 that were transmitted through the nerves, 66 00:04:06,270 --> 00:04:07,870 and the brain had to make sense of them, 67 00:04:07,870 --> 00:04:10,470 by counting them in elaborate ways. 68 00:04:10,470 --> 00:04:12,720 What's the best model of the world that you can have with this? 69 00:04:12,720 --> 00:04:16,530 What will the state of affairs, what's the system that you are in? 70 00:04:16,530 --> 00:04:20,920 And what are the best algorithms that you should be using, to fix your world model. 71 00:04:20,920 --> 00:04:23,310 And this question is pretty old. 72 00:04:23,310 --> 00:04:27,750 And I think that has been answered for the first time by Ray Solomonoff in the 1960. 73 00:04:27,750 --> 00:04:30,840 He has discovered an algorithm, that you can apply when you discover 74 00:04:30,840 --> 00:04:33,540 that you are an robot, and all you have is data. 75 00:04:33,540 --> 00:04:34,870 What is the world like? 76 00:04:34,870 --> 00:04:40,990 And this algorithm is basically a combination of induction and Occam's razor. 77 00:04:40,990 --> 00:04:45,710 And we can mathematically prove that we can not do better than Solomonoff induction. 78 00:04:45,710 --> 00:04:51,380 Unfortunately, Solomonoff induction is not quite computable. 79 00:04:51,380 --> 00:04:54,450 But everything that we are going to do is some... is going to be some approximation 80 00:04:54,450 --> 00:04:55,820 of Salomonoff induction. 81 00:04:55,820 --> 00:04:59,400 So our concepts can not really refer to the facts in the world out there. 82 00:04:59,400 --> 00:05:02,380 We do not get the truth by referring to stuff out there, in the world. 83 00:05:02,380 --> 00:05:07,960 We get meaning by suitably encoding the patterns at our systemic interface. 84 00:05:07,960 --> 00:05:12,270 And AI has recently made a huge progress in encoding data at perceptual interfaces. 85 00:05:12,270 --> 00:05:15,900 Deep learning is about using a stacked hierarchy of feature detectors. 86 00:05:15,900 --> 00:05:21,280 That is, we use pattern detectors and we build them into a networks that are arranged in 87 00:05:21,280 --> 00:05:23,030 hundreds of layers. 88 00:05:23,030 --> 00:05:26,500 And then we adjust the links between these layers. 89 00:05:26,500 --> 00:05:29,380 Usually some kind of... using some kind of gradient descent. 90 00:05:29,380 --> 00:05:33,220 And we can use this to classify for instance images and parts of speech. 91 00:05:33,220 --> 00:05:37,950 So, we get to features that are more and more complex, they started as very, very simple patterns. 92 00:05:37,950 --> 00:05:41,290 And then get more and more complex, until we get to object categories. 93 00:05:41,290 --> 00:05:44,199 And now this systems are able in image recognition task, 94 00:05:44,199 --> 00:05:47,480 to approach performance that is very similar to human performance. 95 00:05:47,480 --> 00:05:52,040 Also what is nice is that it seems to be somewhat similar to what the brain seems to be doing 96 00:05:52,040 --> 00:05:53,740 in visual processing. 97 00:05:53,740 --> 00:05:57,570 And if you take the activation in different levels of these networks and you 98 00:05:57,570 --> 00:06:01,430 erghm... improve the... that... erghmm... enhance this activation a little bit, what 99 00:06:01,430 --> 00:06:03,500 you get is stuff that look very psychedelic. 100 00:06:03,500 --> 00:06:09,620 Which may be similar to what happens, if you put certain illegal substances into people, 101 00:06:09,620 --> 00:06:13,650 and enhance the activity on certain layers of their visual processing. 102 00:06:13,650 --> 00:06:21,540 [BROKEN AUDIO]If you want to classify the differences what we do if we want quantify 103 00:06:21,540 --> 00:06:33,030 this you filter out all the invariences in the data. 104 00:06:33,030 --> 00:06:36,360 The pose that she has, the lighting, the dress that she is on.. has on, 105 00:06:36,360 --> 00:06:38,020 her facial expression and so on. 106 00:06:38,020 --> 00:06:42,900 And then we go to only to this things that is left after we've removed all the nuance data. 107 00:06:42,900 --> 00:06:47,410 But what if we... erghmm want to get to something else, 108 00:06:47,410 --> 00:06:49,850 for instance if we want to understand poses. 109 00:06:49,850 --> 00:06:53,240 Could be for instance that we have several dancers and we want to understand what they 110 00:06:53,240 --> 00:06:54,400 have in common. 111 00:06:54,400 --> 00:06:58,330 So our best bet is not just to have a single classification based filtering, 112 00:06:58,330 --> 00:07:01,199 but instead what we want to have is to take the low level input 113 00:07:01,199 --> 00:07:05,180 and get a whole universe of features, that is interrelated. 114 00:07:05,180 --> 00:07:07,220 So we have different levels of interrelations. 115 00:07:07,220 --> 00:07:08,960 At the lowest levels we have percepts. 116 00:07:08,960 --> 00:07:11,580 On the slightly higher level we have simulations. 117 00:07:11,580 --> 00:07:16,920 And on even higher level we have concept landscape. 118 00:07:16,920 --> 00:07:19,300 How does this representation by simulation work? 119 00:07:19,300 --> 00:07:22,229 Now imagine you want to understand sound. 120 00:07:22,229 --> 00:07:23,669 [Ghulp] 121 00:07:23,669 --> 00:07:26,710 If you are a brain and you want to understand sound you need to model it. 122 00:07:26,710 --> 00:07:31,070 Unfortunatly we can not really model sound with neurons, because sound goes up to 20kHz, 123 00:07:31,070 --> 00:07:36,660 or if you are old like me maybe to 12 kHz. 20 kHz is what babies could do. 124 00:07:36,660 --> 00:07:41,240 And... neurons do not want to do 20 kHz. That's way too fast for them. 125 00:07:41,240 --> 00:07:43,250 They like something like 20 Hz. 126 00:07:43,250 --> 00:07:45,590 So what do you do? You need to make a Fourier transform. 127 00:07:45,590 --> 00:07:49,650 The Fourier transform measures the amount of energy at different frequencies. 128 00:07:49,650 --> 00:07:52,500 And because you can not do it with neurons, you need to do it in hardware. 129 00:07:52,500 --> 00:07:54,180 And turns out this is exactly what we are doing. 130 00:07:54,180 --> 00:07:59,860 We have this cochlea which is this snail like thing in our ears, 131 00:07:59,860 --> 00:08:06,669 and what it does, it transforms energy of sound in different frequency intervals into 132 00:08:06,669 --> 00:08:08,009 energy measurments. 133 00:08:08,009 --> 00:08:10,479 And then gives you something like what you see here. 134 00:08:10,479 --> 00:08:12,550 And this is something that the brain can model, 135 00:08:12,550 --> 00:08:16,210 so we can get a neurosimulator that tries to recreate this patterns. 136 00:08:16,210 --> 00:08:21,370 And we can predict the next input from the cochlea that then understand the sound. 137 00:08:21,370 --> 00:08:23,410 Of course if you want to understand music, 138 00:08:23,410 --> 00:08:25,160 we have to go beyond understanding sound. 139 00:08:25,160 --> 00:08:29,340 We have to understand the transformations that sound can have if you play it at different pitch. 140 00:08:29,340 --> 00:08:33,599 We have to arrange the sound in the sequence that give you rhythms and so on. 141 00:08:33,599 --> 00:08:35,889 And then we want to identify some kind of musical grammar 142 00:08:35,889 --> 00:08:38,799 that we can use to again control the sequencer. 143 00:08:38,799 --> 00:08:42,529 So we have stucked structures. That simulate the world. 144 00:08:42,529 --> 00:08:44,319 And once you've learned this model of music, 145 00:08:44,319 --> 00:08:47,309 once you've learned the musical grammar, the sequencer and the sounds. 146 00:08:47,309 --> 00:08:51,779 You can get to the structure of the individual piece of music. 147 00:08:51,779 --> 00:08:54,399 So, if you want to model the world of music. 148 00:08:54,399 --> 00:08:58,279 You need to have the lowest level of percepts then we have the higher level of mental simulations. 149 00:08:58,279 --> 00:09:01,910 And... which give the sequences of the music and the grammars of music. 150 00:09:01,910 --> 00:09:05,149 And beyond this you have the conceptual landscape that you can use 151 00:09:05,149 --> 00:09:08,249 to describe different styles of music. 152 00:09:08,249 --> 00:09:12,130 And if you go up in the hierarchy, you get to more and more abstract models. 153 00:09:12,130 --> 00:09:13,860 More and more conceptual models. 154 00:09:13,860 --> 00:09:16,449 And more and more analytic models. 155 00:09:16,449 --> 00:09:18,160 And this are causal models at some point. 156 00:09:18,160 --> 00:09:20,999 This causal models can be weakly deterministic, 157 00:09:20,999 --> 00:09:22,980 basically associative models, which tell you 158 00:09:22,980 --> 00:09:27,339 if this state happens, it's quite probable that this one comes afterwords. 159 00:09:27,339 --> 00:09:29,389 Or you can get to a strongly determined model. 160 00:09:29,389 --> 00:09:32,730 Strongly determined model is one which tells you, if you are in this state 161 00:09:32,730 --> 00:09:33,879 and this condition is met, 162 00:09:33,879 --> 00:09:35,589 You are are going to go exactly in this state. 163 00:09:35,589 --> 00:09:40,110 If this condition is not met, or a different condition is met, you are going to this state. 164 00:09:40,110 --> 00:09:41,449 And this is what we call an alghorithm. 165 00:09:41,449 --> 00:09:46,769 it's.. now we are on the domain of computation. 166 00:09:46,769 --> 00:09:48,730 Computation is slightly different from mathematics. 167 00:09:48,730 --> 00:09:51,179 It's important to understand this. 168 00:09:51,179 --> 00:09:54,699 For a long time people have thought that the universe is written in mathematics. 169 00:09:54,699 --> 00:09:58,399 Or that.. minds are mathematical, or anything is mathematical. 170 00:09:58,399 --> 00:10:00,439 In fact nothing is mathematical. 171 00:10:00,439 --> 00:10:04,529 Mathematics is just the domain of formal languages. It doesn't exist. 172 00:10:04,529 --> 00:10:07,300 Mathematics starts with a void. 173 00:10:07,300 --> 00:10:11,939 You throw in a few axioms, and if you've chosen a nice axioms, then you get infinite complexity. 174 00:10:11,939 --> 00:10:13,679 Most of which is not computable. 175 00:10:13,679 --> 00:10:16,270 In mathematics you can express arbitrary statements, 176 00:10:16,270 --> 00:10:18,269 because it's all about formal languages. 177 00:10:18,269 --> 00:10:20,369 Many of this statements will not make sense. 178 00:10:20,369 --> 00:10:22,469 Many of these statements will make sense in some way, 179 00:10:22,469 --> 00:10:24,429 but you can not test whether they make sense, 180 00:10:24,429 --> 00:10:26,740 because they're not computable. 181 00:10:26,740 --> 00:10:29,929 Computation is different. Computation can exist. 182 00:10:29,929 --> 00:10:32,459 It's starts with an initial state. 183 00:10:32,459 --> 00:10:34,739 And then you have a transition function. You do the work. 184 00:10:34,739 --> 00:10:38,449 You apply the transition function, and you get into the next state. 185 00:10:38,449 --> 00:10:41,249 Computation is always finite. 186 00:10:41,249 --> 00:10:43,689 Mathematics is the kingdom of specification. 187 00:10:43,689 --> 00:10:47,290 And computation is the kingdom of implementation. 188 00:10:47,290 --> 00:10:50,629 It's very important to understand this difference. 189 00:10:50,629 --> 00:10:55,329 All our access to mathematics of course is because we do computation. 190 00:10:55,329 --> 00:10:57,459 We can understand mathematics, 191 00:10:57,459 --> 00:10:59,939 because our brain can compute some parts of mathematics. 192 00:10:59,939 --> 00:11:04,439 Very, very little of it, and to very constrained complexity. 193 00:11:04,439 --> 00:11:06,860 But enough, so we can map some of the infinite complexity 194 00:11:06,860 --> 00:11:10,410 and noncomputability of mathematics into computational patterns, 195 00:11:10,410 --> 00:11:12,279 that we can explore. 196 00:11:12,279 --> 00:11:14,410 So computation is about doing the work, 197 00:11:14,410 --> 00:11:16,939 it's about executing the transition function. 198 00:11:19,730 --> 00:11:22,899 Now we've seen that mental representation is about concepts, 199 00:11:22,899 --> 00:11:25,670 mental simulations, conceptual representations 200 00:11:25,670 --> 00:11:29,110 and this conceptual representations give us concept spaces. 201 00:11:29,110 --> 00:11:30,970 And the nice thing about this concept spaces is 202 00:11:30,970 --> 00:11:33,399 that they give us an interface to our mental representations, 203 00:11:33,399 --> 00:11:36,290 We can use to address and manipulate them. 204 00:11:36,290 --> 00:11:39,119 And we can share them in cultures. 205 00:11:39,119 --> 00:11:40,899 And this concepts are compositional. 206 00:11:40,899 --> 00:11:43,639 We can put them together, to create new concepts. 207 00:11:43,639 --> 00:11:48,230 And they can be described using higher dimensional vector spaces. 208 00:11:48,230 --> 00:11:50,319 They don't do simulation and prediction and so on, 209 00:11:50,319 --> 00:11:53,119 but we can capture regularity in our concept wisdom. 210 00:11:53,119 --> 00:11:55,220 With this vector space you can do amazing things. 211 00:11:55,220 --> 00:11:57,589 For instance, if you take the vector from "King" to "Queen" 212 00:11:57,589 --> 00:12:01,009 is pretty much the same vector as to.. between "Man" and "Woman" 213 00:12:01,009 --> 00:12:04,110 And because of this properties, because it's really a high dimentional manifold 214 00:12:04,110 --> 00:12:07,569 this concepts faces, we can do interesting things, like machine translation 215 00:12:07,569 --> 00:12:09,470 without understanding what it means. 216 00:12:09,470 --> 00:12:13,929 That is without doing any proper mental representation, that predicts the world. 217 00:12:13,929 --> 00:12:16,989 So this is a type of meta representation, that is somewhat incomplete, 218 00:12:16,989 --> 00:12:21,199 but it captures the landscape that we share in a culture. 219 00:12:21,199 --> 00:12:25,089 And then there is another type of meta representation, that is linguistic protocols. 220 00:12:25,089 --> 00:12:27,699 Which is basically a formal grammar and vocabulary. 221 00:12:27,699 --> 00:12:29,619 And we need this linguistic protocols 222 00:12:29,619 --> 00:12:32,869 to transfer mental representations between people. 223 00:12:32,869 --> 00:12:36,019 And we do this by basically scanning our mental representation, 224 00:12:36,019 --> 00:12:38,660 disassembling them in some way or disambiguating them. 225 00:12:38,660 --> 00:12:43,040 And then we use it as discrete string of symbols to get it to somebody else, 226 00:12:43,040 --> 00:12:46,429 and he trains an assembler, that reverses this process, 227 00:12:46,429 --> 00:12:51,389 and build something that is pretty similar to what we intended to convey. 228 00:12:51,389 --> 00:12:53,569 And if you look at the progression of AI models, 229 00:12:53,569 --> 00:12:55,600 it pretty much went the opposite direction. 230 00:12:55,600 --> 00:13:00,279 So AI started with linguistic protocols, which were expressed in formal grammars. 231 00:13:00,279 --> 00:13:05,209 And then it got to concepts spaces, and now it's about to address percepts. 232 00:13:05,209 --> 00:13:09,689 And at some point in near future it's going to get better at mental simulations. 233 00:13:09,689 --> 00:13:11,730 And at some point after that we get to 234 00:13:11,730 --> 00:13:14,769 attention directed and motivationally connected systems, 235 00:13:14,769 --> 00:13:16,600 that make sense of the world. 236 00:13:16,600 --> 00:13:20,290 that are in some sense able to address meaning. 237 00:13:20,290 --> 00:13:23,489 This is the hardware that we have can do. 238 00:13:23,489 --> 00:13:25,629 What kind of hardware do we have? 239 00:13:25,629 --> 00:13:28,480 That's a very interesting question. 240 00:13:28,480 --> 00:13:32,230 It could start out with a question: How difficult is it to define a brain? 241 00:13:32,230 --> 00:13:35,439 We know that the brain must be somewhere hidden in the genome. 242 00:13:35,439 --> 00:13:38,290 The genome fits on a CD ROM. It's not that complicated. 243 00:13:38,290 --> 00:13:40,399 It's easier than Microsoft Windows. *laughter* 244 00:13:40,399 --> 00:13:45,549 And we also know, that about 2% of the genome is coding for proteins. 245 00:13:45,549 --> 00:13:48,429 And maybe about 10% of the genome has some kind of stuff 246 00:13:48,429 --> 00:13:51,239 that tells you when to switch protein. 247 00:13:51,239 --> 00:13:52,829 And the remainder is mostly garbage. 248 00:13:52,829 --> 00:13:57,170 It's old viruses that are left over and has never been properly deleted and so on. 249 00:13:57,170 --> 00:14:01,420 Because there are no real code revisions in the genome. 250 00:14:01,420 --> 00:14:08,119 So how much of this 10% that is 75 MB code for the brain. 251 00:14:08,119 --> 00:14:09,469 We don't really know. 252 00:14:09,469 --> 00:14:13,399 What we do know is we share almost all of this with mice. 253 00:14:13,399 --> 00:14:15,769 Genetically speaking human is a pretty big mouse. 254 00:14:15,769 --> 00:14:21,049 With a few bits changed, so.. to fix some of the genetic expressions 255 00:14:21,049 --> 00:14:25,879 And that is most of the stuff there is going to code for cells and metabolism 256 00:14:25,879 --> 00:14:27,999 and how your body looks like and so on. 257 00:14:27,999 --> 00:14:33,679 But if you look at erghmm... how much is expressed in the brain and only in the brain, 258 00:14:33,679 --> 00:14:35,170 in terms of proteins and so on. 259 00:14:35,170 --> 00:14:45,639 We find it's about... well of the 2% it's about 5%. That is only the 5% of the 2% that 260 00:14:45,639 --> 00:14:46,799 is only in the brain. 261 00:14:46,799 --> 00:14:50,199 And another 5% of the 2% is predominantly in the brain. 262 00:14:50,199 --> 00:14:52,069 That is more in the brain than anywhere else. 263 00:14:52,069 --> 00:14:54,249 Which gives you some kind of thing like a lower bound. 264 00:14:54,249 --> 00:14:59,379 Which means to encode a brain genetically base on the hardware that we are using. 265 00:14:59,379 --> 00:15:03,539 We need something like at least 500 kB of code. 266 00:15:03,539 --> 00:15:06,670 Actually ehmm.. this... we very conservative lower bound. 267 00:15:06,670 --> 00:15:08,720 It's going to be a little more I guess. 268 00:15:08,720 --> 00:15:11,449 But it sounds surprisingly little, right? 269 00:15:11,449 --> 00:15:13,709 But in terms of scientific theories this is a lot. 270 00:15:13,709 --> 00:15:16,519 I mean the universe, according to the core theory 271 00:15:16,519 --> 00:15:19,420 of the quantum mechanics and so on is like so much of code. 272 00:15:19,420 --> 00:15:20,569 It's like half a page of code. 273 00:15:20,569 --> 00:15:23,100 That's it. That's all you need to generate the universe. 274 00:15:23,100 --> 00:15:25,489 And if you want to understand evolution it's like a paragraph. 275 00:15:25,489 --> 00:15:29,609 It's couple lines you need to understand evolutionary process. 276 00:15:29,609 --> 00:15:32,199 And there is a lots, lots of details, that's you get afterwards. 277 00:15:32,199 --> 00:15:34,220 Because this process itself doesn't define 278 00:15:34,220 --> 00:15:37,259 how the animals are going to look like, and in similar way is.. 279 00:15:37,259 --> 00:15:41,269 the code of the universe doesn't tell you what this planet is going to look like. 280 00:15:41,269 --> 00:15:43,279 And what you guys are going to look like. 281 00:15:43,279 --> 00:15:45,949 It's just defining the rulebook. 282 00:15:45,949 --> 00:15:49,209 And in the same sense genome defines the rulebook, 283 00:15:49,209 --> 00:15:51,569 by which our brain is build. 284 00:15:51,569 --> 00:15:56,399 erghmmm,.. The brain boots itself into developer process, 285 00:15:56,399 --> 00:15:58,119 and this booting takes some time. 286 00:15:58,119 --> 00:16:01,069 So subliminal learning in which initial connections are forged 287 00:16:01,069 --> 00:16:04,910 And basic models are build of the world, so we can operate in it. 288 00:16:04,910 --> 00:16:06,999 And how long does this booting take? 289 00:16:06,999 --> 00:16:09,669 I thing it's about 80 mega seconds. 290 00:16:09,669 --> 00:16:14,319 That's the time that a child is awake until it's 2.5 years old. 291 00:16:14,319 --> 00:16:16,449 By this age you understand Star Wars. 292 00:16:16,449 --> 00:16:20,029 And I think that everything after understanding Star Wars is cosmetics. 293 00:16:20,029 --> 00:16:26,799 *laughter**applause* 294 00:16:26,799 --> 00:16:32,820 You are going to be online, if you get to arrive old age for about 1.5 giga seconds. 295 00:16:32,820 --> 00:16:37,929 And in this time I think you are going to get not to watch more than 5 milion concepts. 296 00:16:37,929 --> 00:16:41,600 Why? I don't know real... If you look at this child. 297 00:16:41,600 --> 00:16:45,480 If a child would be able to form a concept let say every 5 minutes, 298 00:16:45,480 --> 00:16:48,529 then by the time it's about 4 years old, it's going to have 299 00:16:48,529 --> 00:16:51,549 something like 250 thousands concepts. 300 00:16:51,549 --> 00:16:54,119 And... so... a quarter million. 301 00:16:54,119 --> 00:16:56,809 And if we extrapolate this into our lifetime, 302 00:16:56,809 --> 00:16:59,799 at some point it slows down, because we have enough concepts, 303 00:16:59,799 --> 00:17:01,230 to describe the world. 304 00:17:01,230 --> 00:17:04,410 Maybe it's something... It's I think it's less that 5 million. 305 00:17:04,410 --> 00:17:07,140 How much storage capacity does the brain has? 306 00:17:07,140 --> 00:17:12,319 I think that the... the estimates are pretty divergent, 307 00:17:12,319 --> 00:17:14,930 The lower bound is something like a 100 GB, 308 00:17:14,930 --> 00:17:18,569 And the upper bound is something like 2.5 PB. 309 00:17:18,569 --> 00:17:21,890 There is even... even some higher outliers this.. 310 00:17:21,890 --> 00:17:25,630 If you for instance think that we need all those synaptic vesicle to store information, 311 00:17:25,630 --> 00:17:27,530 maybe even more fits into this. 312 00:17:27,530 --> 00:17:31,740 But the 2.5 PB is usually based on what you need 313 00:17:31,740 --> 00:17:34,760 to code the information that is in all the neurons. 314 00:17:34,760 --> 00:17:36,770 But maybe the neurons do not really matter so much, 315 00:17:36,770 --> 00:17:39,930 because if the neuron dies it's not like the word is changing dramatically. 316 00:17:39,930 --> 00:17:44,270 The brain is very resilient against individual neurons failing. 317 00:17:44,270 --> 00:17:48,930 So the 100 GB capacity is much more what you actually store in the neurons. 318 00:17:48,930 --> 00:17:51,380 If you look at all the redundancy that you need. 319 00:17:51,380 --> 00:17:54,230 And I think this is much closer to the actual Ballpark figure. 320 00:17:54,230 --> 00:17:58,130 Also if you want to store 5 hundred... 5 million concepts, 321 00:17:58,130 --> 00:18:02,330 and maybe 10 times or 100 times the number of percepts, on top of this, 322 00:18:02,330 --> 00:18:05,490 this is roughly the Ballpark figure that you are going to need. 323 00:18:05,490 --> 00:18:07,110 So our brain 324 00:18:07,110 --> 00:18:08,320 is a prediction machine. 325 00:18:08,320 --> 00:18:11,490 It... What it does is it reduces the entropy of the environment, 326 00:18:11,490 --> 00:18:14,610 to solve whatever problems you are encountering, 327 00:18:14,610 --> 00:18:17,790 if you don't have a... feedback loop, to fix them. 328 00:18:17,790 --> 00:18:20,240 So normally if something happens, we have some kind of feedback loop, 329 00:18:20,240 --> 00:18:23,440 that regulates our temperature or that makes problems go away. 330 00:18:23,440 --> 00:18:26,050 And only when this is not working we employ recognition. 331 00:18:26,050 --> 00:18:29,250 And then we start this arbitrary computational processes, 332 00:18:29,250 --> 00:18:31,830 that is facilitated by the neural cortex. 333 00:18:31,830 --> 00:18:34,940 And this.. arhmm.. neural cortex has really do arbitrary programs. 334 00:18:34,940 --> 00:18:37,870 But it can do so with only with very limited complexity, 335 00:18:37,870 --> 00:18:42,070 because really you just saw, it's not that complex. 336 00:18:42,070 --> 00:18:43,900 The modeling of the world is very slow. 337 00:18:43,900 --> 00:18:46,570 And it's something that we see in our eye models. 338 00:18:46,570 --> 00:18:48,150 To learn the basic structure of the world 339 00:18:48,150 --> 00:18:49,330 takes a very long time. 340 00:18:49,330 --> 00:18:52,650 To learn basically that we are moving in 3D and objects are moving, 341 00:18:52,650 --> 00:18:54,030 and what they look like. 342 00:18:54,030 --> 00:18:55,130 Once we have this basic model, 343 00:18:55,130 --> 00:18:59,300 we can get to very, very quick understanding within this model. 344 00:18:59,300 --> 00:19:02,110 Basically encoding based on the structure of the world, 345 00:19:02,110 --> 00:19:03,610 that we've learned. 346 00:19:03,610 --> 00:19:07,100 And this is some kind of data compression, that we are doing. 347 00:19:07,100 --> 00:19:09,740 We use this model, this grammar of the world, 348 00:19:09,740 --> 00:19:12,150 this simulation structures that we've learned, 349 00:19:12,150 --> 00:19:15,190 to encode the world very, very efficently. 350 00:19:15,190 --> 00:19:17,740 How much data compression do we get? 351 00:19:17,740 --> 00:19:19,860 Well... if you look at the retina. 352 00:19:19,860 --> 00:19:24,610 The retina get's data in the order of about 10Gb/s. 353 00:19:24,610 --> 00:19:27,500 And the retina already compresses these data, 354 00:19:27,500 --> 00:19:31,120 and puts them into optic nerve at the rate of about 1Mb/s 355 00:19:31,120 --> 00:19:34,030 This is what you get fed into visual cortex. 356 00:19:34,030 --> 00:19:36,370 And the visual cortex does some additional compression, 357 00:19:36,370 --> 00:19:42,110 and by the time it gets to layer four of the first layer of vision, to V1. 358 00:19:42,110 --> 00:19:46,880 We are down to something like 1Kb/s. 359 00:19:46,880 --> 00:19:50,720 So if we extrapolate this, and you get live to the age of 80 years, 360 00:19:50,720 --> 00:19:54,140 and you are awake for 2/3 of your lifetime. 361 00:19:54,140 --> 00:19:56,930 That is you have your eyes open for 2/3 of your lifetime. 362 00:19:56,930 --> 00:19:59,040 The stuff that you get into your brain, 363 00:19:59,040 --> 00:20:03,700 via your visual perception is going to be only 2TB. 364 00:20:03,700 --> 00:20:05,370 Only 2TB of visual data. 365 00:20:05,370 --> 00:20:06,680 Throughout all your lifetime. 366 00:20:06,680 --> 00:20:09,430 That's all you are going to get ever to see. 367 00:20:09,430 --> 00:20:11,160 Isn't this depressing? 368 00:20:11,160 --> 00:20:12,790 *laughter* 369 00:20:12,790 --> 00:20:16,540 So I would really like to eghmm.. to tell you, 370 00:20:16,540 --> 00:20:22,750 choose wisely what you are going to look at. *laughter* 371 00:20:22,750 --> 00:20:26,940 Ok. Let's look at this problem of neural compositionality. 372 00:20:26,940 --> 00:20:29,250 Our brains has this amazing thing that they can put 373 00:20:29,250 --> 00:20:31,510 meta representation together very, very quickly. 374 00:20:31,510 --> 00:20:33,150 For instance you read a page of code, 375 00:20:33,150 --> 00:20:35,190 you compile it in you mind into some kind of program 376 00:20:35,190 --> 00:20:37,700 it tells you what this page is going to do. 377 00:20:37,700 --> 00:20:39,110 Isn't that amazing? 378 00:20:39,110 --> 00:20:40,810 And then you can forget about this, 379 00:20:40,810 --> 00:20:43,910 disassemble it all, and use the building blocks for something else. 380 00:20:43,910 --> 00:20:45,230 It's like legos. 381 00:20:45,230 --> 00:20:48,000 How you can do this with neurons? 382 00:20:48,000 --> 00:20:50,160 Legos can do this, because they have a well defined interface. 383 00:20:50,160 --> 00:20:52,180 They have all this slots, you know, that fit together 384 00:20:52,180 --> 00:20:53,600 in well defined ways. 385 00:20:53,600 --> 00:20:54,530 How can neurons do this? 386 00:20:54,530 --> 00:20:57,280 Well, neurons can maybe learn the interface of other neurons. 387 00:20:57,280 --> 00:20:59,780 But that's difficult, because every neuron looks slightly different, 388 00:20:59,780 --> 00:21:04,830 after all this... some kind of biologically grown natural stuff. 389 00:21:04,830 --> 00:21:06,610 *laughter* 390 00:21:06,610 --> 00:21:10,620 So what you want to do is, you want to encapsulate this erhmm... 391 00:21:10,620 --> 00:21:13,020 diversity of the neurons to make the predictable. 392 00:21:13,020 --> 00:21:14,820 To give them well defined interface. 393 00:21:14,820 --> 00:21:16,410 And I think that nature solution to this 394 00:21:16,410 --> 00:21:19,770 is cortical columns. 395 00:21:19,770 --> 00:21:24,250 Cortical column is a circuit of between 100 and 400 neurons. 396 00:21:24,250 --> 00:21:26,860 And this circuit has some kind of neural network, 397 00:21:26,860 --> 00:21:28,650 that can learn stuff. 398 00:21:28,650 --> 00:21:31,070 And after it has learned particular function, 399 00:21:31,070 --> 00:21:35,320 and in between, it's able to link up these other cortical columns. 400 00:21:35,320 --> 00:21:37,120 And we have about 100 million of those. 401 00:21:37,120 --> 00:21:39,770 Depending on how many neurons you assume is in there, 402 00:21:39,770 --> 00:21:41,490 it's... erghmm we guess it's something, 403 00:21:41,490 --> 00:21:46,500 at least 20 million and maybe something like a 100 million. 404 00:21:46,500 --> 00:21:48,330 And this cortical columns, what they can do, 405 00:21:48,330 --> 00:21:50,280 is they can link up like lego bricks, 406 00:21:50,280 --> 00:21:54,130 and then perform, by transmitting information between them, 407 00:21:54,130 --> 00:21:55,990 pretty much arbitrary computations. 408 00:21:55,990 --> 00:21:57,540 What kind of computation? 409 00:21:57,540 --> 00:22:00,130 Well... Solomonoff induction. 410 00:22:00,130 --> 00:22:03,820 And... they have some short range links, to their neighbors. 411 00:22:03,820 --> 00:22:05,690 Which comes almost for free, because erghmm.. 412 00:22:05,690 --> 00:22:08,490 well, they are connected to them, they are direct neighborhood. 413 00:22:08,490 --> 00:22:10,050 And they have some long range connectivity, 414 00:22:10,050 --> 00:22:13,000 so you can combine everything in your cortex with everything. 415 00:22:13,000 --> 00:22:14,900 So you need some kind of global switchboard. 416 00:22:14,900 --> 00:22:17,630 Some grid like architecture of long range connections. 417 00:22:17,630 --> 00:22:18,900 They are going to be more expensive, 418 00:22:18,900 --> 00:22:20,640 they are going to be slower, 419 00:22:20,640 --> 00:22:23,590 but they are going to be there. 420 00:22:23,590 --> 00:22:26,070 So how can we optimize what these guys are doing? 421 00:22:26,070 --> 00:22:28,270 In some sense it's like an economy. 422 00:22:28,270 --> 00:22:31,460 It's not enduring based system, as we often use in machine learning. 423 00:22:31,460 --> 00:22:32,780 It's really an economy. You have... 424 00:22:32,780 --> 00:22:35,560 The question is, you have a fixed number of elements, 425 00:22:35,560 --> 00:22:37,970 how can you do the most valuable stuff with them. 426 00:22:37,970 --> 00:22:41,030 Fixed resources, most valuable stuff, the problem is economy. 427 00:22:41,030 --> 00:22:43,320 So you have an economy of information brokers. 428 00:22:43,320 --> 00:22:45,830 Every one of these guys, this little cortical columns, 429 00:22:45,830 --> 00:22:48,150 is very simplistic information broker. 430 00:22:48,150 --> 00:22:50,950 And they trade rewards against neg entropy, 431 00:22:50,950 --> 00:22:54,140 Against reducing entropy in the... in the world. 432 00:22:54,140 --> 00:22:55,790 And to do this, as we just saw 433 00:22:55,790 --> 00:22:58,890 that they need some kind of standardized interface. 434 00:22:58,890 --> 00:23:02,090 And internally, to use this interface they are going to 435 00:23:02,090 --> 00:23:03,880 have some kind of state machine. 436 00:23:03,880 --> 00:23:05,660 And then they are going to pass messages 437 00:23:05,660 --> 00:23:07,400 between each other. 438 00:23:07,400 --> 00:23:08,630 And what are these messages? 439 00:23:08,630 --> 00:23:11,100 Well, it's going to be hard to discover these messages, 440 00:23:11,100 --> 00:23:12,800 by looking at brains. 441 00:23:12,800 --> 00:23:14,800 Because it's very difficult to see in brains, 442 00:23:14,800 --> 00:23:15,450 what the are actually doing. 443 00:23:15,450 --> 00:23:17,250 you just see all these neurons. 444 00:23:17,250 --> 00:23:18,790 And if you would be waiting for neuroscience, 445 00:23:18,790 --> 00:23:20,970 to discover anything, we wouldn't even have 446 00:23:20,970 --> 00:23:22,590 gradient descent or anything else. 447 00:23:22,590 --> 00:23:23,720 We wouldn't have neuron learning. 448 00:23:23,720 --> 00:23:25,420 We wouldn't have all this advances in AI. 449 00:23:25,420 --> 00:23:28,230 Jürgen Schmidhuber said that the biggest, 450 00:23:28,230 --> 00:23:30,010 the last contribution of neuroscience to 451 00:23:30,010 --> 00:23:32,220 artificial intelligence was about 50 years ago. 452 00:23:32,220 --> 00:23:34,280 That's depressing, and it might be 453 00:23:34,280 --> 00:23:37,870 overemphasizing the unimportance of neuroscience, 454 00:23:37,870 --> 00:23:39,490 because neuroscience is very important, 455 00:23:39,490 --> 00:23:41,090 once you know what are you looking for. 456 00:23:41,090 --> 00:23:42,510 You can actually often find this, 457 00:23:42,510 --> 00:23:44,320 and see whether you are on the right track. 458 00:23:44,320 --> 00:23:45,860 But it's very difficult to take neuroscience 459 00:23:45,860 --> 00:23:47,940 to understand how the brain is working. 460 00:23:47,940 --> 00:23:49,290 Because it's really like understanding 461 00:23:49,290 --> 00:23:53,230 flight by looking at birds through a microscope. 462 00:23:53,230 --> 00:23:55,150 So, what are these messages? 463 00:23:55,150 --> 00:23:57,850 You are going to need messages, that tell these cortical columns 464 00:23:57,850 --> 00:24:00,160 to join themselves into a structure. 465 00:24:00,160 --> 00:24:01,990 And to unlink again once they're done. 466 00:24:01,990 --> 00:24:03,690 You need ways that they can request each other 467 00:24:03,690 --> 00:24:06,040 to perform computations for them. 468 00:24:06,040 --> 00:24:07,510 You need ways they can inhibit each other 469 00:24:07,510 --> 00:24:08,320 when they are linked up. 470 00:24:08,320 --> 00:24:10,990 So they don't do conflicting computations. 471 00:24:10,990 --> 00:24:12,940 Then they need to tell you whether the computation, 472 00:24:12,940 --> 00:24:14,110 the result of the computation 473 00:24:14,110 --> 00:24:16,730 that the are asked to do is probably false. 474 00:24:16,730 --> 00:24:19,340 Or whether it's probably true, but you still need to wait for others, 475 00:24:19,340 --> 00:24:21,990 to tell you whether the details worked out. 476 00:24:21,990 --> 00:24:24,240 Or whether it's confirmed true that the concepts 477 00:24:24,240 --> 00:24:26,730 that they stand for is actually the case. 478 00:24:26,730 --> 00:24:28,150 And then you want to have learning, 479 00:24:28,150 --> 00:24:29,630 to tell you how well this worked. 480 00:24:29,630 --> 00:24:31,390 So you will have to announce a bounty, 481 00:24:31,390 --> 00:24:34,380 that tells them to link up and kind of reward signal 482 00:24:34,380 --> 00:24:36,740 that makes do computation in the first place. 483 00:24:36,740 --> 00:24:38,680 And then you want to have some kind of reward signal 484 00:24:38,680 --> 00:24:40,550 once you got the result as an organism. 485 00:24:40,550 --> 00:24:42,280 But you reach your goal if you made 486 00:24:42,280 --> 00:24:45,810 the disturbance go away or what ever you consume the cake. 487 00:24:45,810 --> 00:24:47,710 And then you will have some kind of reward signal 488 00:24:47,710 --> 00:24:49,250 that's you give everybody. 489 00:24:49,250 --> 00:24:50,650 That was involved in this. 490 00:24:50,650 --> 00:24:52,720 And this reward signal facilitates learning, 491 00:24:52,720 --> 00:24:55,230 so the.. difference between the announce reward 492 00:24:55,230 --> 00:24:57,530 and consumption reward is the learning signal 493 00:24:57,530 --> 00:24:58,740 for these guys. 494 00:24:58,740 --> 00:25:00,210 So they can learn how to play together, 495 00:25:00,210 --> 00:25:02,700 and how to do the Solomonoff induction. 496 00:25:02,700 --> 00:25:04,660 Now, I've told you that Solomonoff induction 497 00:25:04,660 --> 00:25:05,280 is not computable. 498 00:25:05,280 --> 00:25:07,630 And it's mostly because of two things, 499 00:25:07,630 --> 00:25:09,280 First of all it's needs infinite resources 500 00:25:09,280 --> 00:25:11,200 to compare all the possible models. 501 00:25:11,200 --> 00:25:13,530 And the other one is that we do not know 502 00:25:13,530 --> 00:25:15,440 the priori probability for our Bayesian model. 503 00:25:15,440 --> 00:25:19,280 If we do not know how likely unknown stuff is in the world. 504 00:25:19,280 --> 00:25:22,520 So what we do instead is, we set some kind of hyperparameter, 505 00:25:22,520 --> 00:25:25,050 Some kind of default priori probability for concepts, 506 00:25:25,050 --> 00:25:28,110 that are encoded by cortical columns. 507 00:25:28,110 --> 00:25:30,580 And if we set these parameters very low, 508 00:25:30,580 --> 00:25:32,140 then we are going to end up with inferences 509 00:25:32,140 --> 00:25:35,250 that are quite probable. 510 00:25:35,250 --> 00:25:36,480 For unknown things. 511 00:25:36,480 --> 00:25:37,690 And then we can test for those. 512 00:25:37,690 --> 00:25:41,350 If we set this parameter higher, we are going to be very, very creative. 513 00:25:41,350 --> 00:25:43,670 But we end up with many many theories, 514 00:25:43,670 --> 00:25:45,140 that are difficult to test. 515 00:25:45,140 --> 00:25:48,470 Because maybe there are too many theories to test. 516 00:25:48,470 --> 00:25:50,650 Basically every of these cortical columns will now tell you, 517 00:25:50,650 --> 00:25:52,240 when you ask them if they are true: 518 00:25:52,240 --> 00:25:54,960 "Yes I'm probably true, but i still need to ask others, 519 00:25:54,960 --> 00:25:56,980 to work on the details" 520 00:25:56,980 --> 00:25:58,670 So these others are going to be get active, 521 00:25:58,670 --> 00:26:00,640 and they are being asked by the asking element: 522 00:26:00,640 --> 00:26:01,730 "Are you going to be true?", 523 00:26:01,730 --> 00:26:04,380 and they say "Yeah, probably yes, I just have to work on the details" 524 00:26:04,380 --> 00:26:05,930 and they are going to ask even more. 525 00:26:05,930 --> 00:26:07,980 So your brain is going to light up like a christmas tree, 526 00:26:07,980 --> 00:26:10,240 and do all these amazing computations, 527 00:26:10,240 --> 00:26:12,450 and you see connections everywhere, most of them are wrong. 528 00:26:12,450 --> 00:26:16,310 You are basically in psychotic state if your hyperparameter is too high. 529 00:26:16,310 --> 00:26:20,790 You're brain invents more theories that it can disproof. 530 00:26:20,790 --> 00:26:24,550 Would it actually sometimes be good to be in this state? 531 00:26:24,550 --> 00:26:27,850 You bet. So i think every night our brain goes in this state. 532 00:26:27,850 --> 00:26:31,720 We turn up this hyperparameter. We dream. We get all kinds 533 00:26:31,720 --> 00:26:34,100 weird connections, and we get to see connections, 534 00:26:34,100 --> 00:26:36,140 that otherwise we couldn't be seeing. 535 00:26:36,140 --> 00:26:38,080 Even though... because they are highly improbable. 536 00:26:38,080 --> 00:26:42,750 But sometimes they hold, and we see... "Oh my God, DNA is organized in double helix". 537 00:26:42,750 --> 00:26:44,640 And this is what we remember in the morning. 538 00:26:44,640 --> 00:26:46,870 All the other stuff is deleted. 539 00:26:46,870 --> 00:26:48,440 So we usually don't form long term memories 540 00:26:48,440 --> 00:26:51,480 in dreams, if everything goes well. 541 00:26:51,480 --> 00:26:56,670 If you accidentally trip this up.. your modulators, 542 00:26:56,670 --> 00:26:59,100 for instance by consuming illegal substances, 543 00:26:59,100 --> 00:27:01,690 or because you just gone randomly psychotic 544 00:27:01,690 --> 00:27:04,600 you was basically entering a dreaming state I guess. 545 00:27:04,600 --> 00:27:06,990 You get to a state when the brain starts inventing more 546 00:27:06,990 --> 00:27:10,860 concepts that it can disproof. 547 00:27:10,860 --> 00:27:13,600 So you want to have a state where this is well balanced. 548 00:27:13,600 --> 00:27:16,180 And the difference between highly creative people, 549 00:27:16,180 --> 00:27:20,070 and very religious people is probably a different setting of this hyperparameter. 550 00:27:20,070 --> 00:27:21,890 So I suspect that people that people that are genius, 551 00:27:21,890 --> 00:27:23,880 like people like Einstein and so on, 552 00:27:23,880 --> 00:27:26,600 do not simply have better neurons than others. 553 00:27:26,600 --> 00:27:29,130 What they mostly have is a slightly hyperparameter, 554 00:27:29,130 --> 00:27:33,860 that is very finely tuned, so they can get better balance than other people 555 00:27:33,860 --> 00:27:43,850 in finding theories that might be true, but can still be disprooven. 556 00:27:43,850 --> 00:27:49,480 So inventiveness could be a hyperparameter in the brain. 557 00:27:49,480 --> 00:27:54,169 If you want to measure the quality of belief that we have 558 00:27:54,169 --> 00:27:56,370 we are going to have to have some kind of some cost function 559 00:27:56,370 --> 00:27:58,710 which is based on motivational system. 560 00:27:58,710 --> 00:28:02,400 And to identify if belief is good or not we can abstract criteria, 561 00:28:02,400 --> 00:28:06,440 for instance how well does it predict the wourld, or how about does it reduce uncertainty 562 00:28:06,440 --> 00:28:07,590 in the world, 563 00:28:07,590 --> 00:28:10,020 or is it consistency and sparse. 564 00:28:10,020 --> 00:28:14,080 And then of course utility, how about does it help me to satisfy my needs. 565 00:28:14,080 --> 00:28:18,920 And the motivational system is going to evaluate all this things by giving a signal. 566 00:28:18,920 --> 00:28:24,200 And the first signal.. kind of signal is the possible rewards if we are able to compute 567 00:28:24,200 --> 00:28:25,020 the task. 568 00:28:25,020 --> 00:28:27,430 And this is probably done by dopamine. 569 00:28:27,430 --> 00:28:30,350 So we have a very small area in the brain, substantia nigra, 570 00:28:30,350 --> 00:28:33,610 and the ventral tegmental area, and they produce dopamine. 571 00:28:33,610 --> 00:28:38,180 And this get fed into lateral frontal cortext and the frontal lobe, 572 00:28:38,180 --> 00:28:41,920 which control attention, and tell you what things to do. 573 00:28:41,920 --> 00:28:46,020 And if we have successfully done what you wanted to do, 574 00:28:46,020 --> 00:28:49,300 we consume the rewards. 575 00:28:49,300 --> 00:28:51,940 And we do this with another signal which is serotonine. 576 00:28:51,940 --> 00:28:53,480 It's also announce to motivational system, 577 00:28:53,480 --> 00:28:55,870 to this very small are the Raphe nuclei. 578 00:28:55,870 --> 00:28:58,690 And it feeds into all the areas of the brain where learning is necessary. 579 00:28:58,690 --> 00:29:02,160 A connection is strengthen once you get to result. 580 00:29:02,160 --> 00:29:07,559 These two substances are emitted by the motivational system. 581 00:29:07,559 --> 00:29:09,710 The motivational system is a bunch of needs, 582 00:29:09,710 --> 00:29:11,510 essentially you regulate it below the cortext. 583 00:29:11,510 --> 00:29:14,490 They are not part of your mental representations. 584 00:29:14,490 --> 00:29:16,930 They are part of something that is more primary than this. 585 00:29:16,930 --> 00:29:19,360 This is what makes us go, this is what makes us human. 586 00:29:19,360 --> 00:29:22,290 This is not our rationality, this is what we want. 587 00:29:22,290 --> 00:29:27,000 And the needs are physiological, they are social, they are cognitive. 588 00:29:27,000 --> 00:29:28,960 And you pretty much born with them. 589 00:29:28,960 --> 00:29:30,470 They can not be totally adaptive, 590 00:29:30,470 --> 00:29:33,340 because if we were adaptive, we wouldn't be doing anything. 591 00:29:33,340 --> 00:29:35,390 The needs are resistive. 592 00:29:35,390 --> 00:29:38,290 They are pushing us against the world. 593 00:29:38,290 --> 00:29:40,170 If you wouldn't have all this needs, 594 00:29:40,170 --> 00:29:41,740 If you wouldn't have this motivational system, 595 00:29:41,740 --> 00:29:43,630 you would just be doing what best for you. 596 00:29:43,630 --> 00:29:45,150 Which means collapse on the ground, 597 00:29:45,150 --> 00:29:49,010 be a vegetable, rod, give into gravity. 598 00:29:49,010 --> 00:29:50,270 Instead you do all this unpleasant things, 599 00:29:50,270 --> 00:29:52,690 to get up in the morning, you eat, you have sex, 600 00:29:52,690 --> 00:29:54,120 you do all this crazy things. 601 00:29:54,120 --> 00:29:58,809 And it's only because the motivational system forces you to. 602 00:29:58,809 --> 00:30:00,850 The motivational system takes this bunch of matter, 603 00:30:00,850 --> 00:30:02,890 and makes us to do all these strange things, 604 00:30:02,890 --> 00:30:05,940 just so genomes get replicated and so on. 605 00:30:05,940 --> 00:30:10,470 And... so to do this, we are going to build resistance against the world. 606 00:30:10,470 --> 00:30:13,360 And the motivational system is in a sense forcing us, 607 00:30:13,360 --> 00:30:15,470 to do all this things by giving us needs, 608 00:30:15,470 --> 00:30:18,330 and the need have some kind of target value and current value. 609 00:30:18,330 --> 00:30:21,850 If we have a differential between the target value and current value, 610 00:30:21,850 --> 00:30:24,590 we perceive some urgency to do something about the need. 611 00:30:24,590 --> 00:30:26,680 And when the target value approaches the current value 612 00:30:26,680 --> 00:30:28,660 we get the pleasure, which is a learning signal. 613 00:30:28,660 --> 00:30:30,540 If it gets away from it we get a displeasure signal, 614 00:30:30,540 --> 00:30:31,870 which is also a learning signal. 615 00:30:31,870 --> 00:30:35,370 And we can use this to structure our understanding of the world. 616 00:30:35,370 --> 00:30:36,870 To understand what goals are and so on. 617 00:30:36,870 --> 00:30:40,020 Goals are learned. Needs are not. 618 00:30:40,020 --> 00:30:42,780 To learn we need success and failure in the world. 619 00:30:42,780 --> 00:30:45,940 But to do things we need anticipated reward. 620 00:30:45,940 --> 00:30:48,120 So it's dopamine that's makes brain go round. 621 00:30:48,120 --> 00:30:50,560 Dopamine makes you do things. 622 00:30:50,560 --> 00:30:52,750 But in order to do this in the right way, 623 00:30:52,750 --> 00:30:54,610 you have to make sure, that the cells can not 624 00:30:54,610 --> 00:30:55,880 produce dopamine themselves. 625 00:30:55,880 --> 00:30:59,100 If they do this they can start to drive others to work for them. 626 00:30:59,100 --> 00:31:01,870 You are going to get something like bureaucracy in your neural cortext, 627 00:31:01,870 --> 00:31:05,650 where different bosses try to set up others to they own bidding 628 00:31:05,650 --> 00:31:07,910 and pitch against other groups in nerual cortext. 629 00:31:07,910 --> 00:31:09,730 It's going to be horrible. 630 00:31:09,730 --> 00:31:12,210 So you want to have some kind of central authority, 631 00:31:12,210 --> 00:31:16,290 that make sure that the cells do not produce dopamine themselves. 632 00:31:16,290 --> 00:31:19,679 It's only been produce in very small area and then given out, 633 00:31:19,679 --> 00:31:21,059 and pass through the system. 634 00:31:21,059 --> 00:31:23,350 And after you're done with it's going to be gone, 635 00:31:23,350 --> 00:31:26,070 so there is no hoarding of the dopamine. 636 00:31:26,070 --> 00:31:29,770 And in our society the role of dopamine is played by money. 637 00:31:29,770 --> 00:31:32,150 Money is not reward in itself. 638 00:31:32,150 --> 00:31:35,570 It's in some sense way that you can trade against the reward. 639 00:31:35,570 --> 00:31:36,850 You can not eat money. 640 00:31:36,850 --> 00:31:40,500 You can take it later and take a arbitrary reward for it. 641 00:31:40,500 --> 00:31:45,400 And in some sense money is the dopamine that makes organizations 642 00:31:45,400 --> 00:31:48,410 and society, companies and many individuals do things. 643 00:31:48,410 --> 00:31:50,500 They do stuff because of money. 644 00:31:50,500 --> 00:31:53,309 But money if you compare to dopamine is pretty broken, 645 00:31:53,309 --> 00:31:54,850 because you can hoard it. 646 00:31:54,850 --> 00:31:57,400 So you are going to have this cortical columns in the real world, 647 00:31:57,400 --> 00:31:59,670 which are individual people or individual corporations. 648 00:31:59,670 --> 00:32:03,250 They are hoarding the dopamine, they sit on this very big pile of dopamine. 649 00:32:03,250 --> 00:32:07,890 They are starving the rest of the society of the dopamine. 650 00:32:07,890 --> 00:32:10,630 They don't give it away, and they can make it do it's bidding. 651 00:32:10,630 --> 00:32:13,970 So for instance they can pitch substantial part of society 652 00:32:13,970 --> 00:32:16,130 against understanding of global warming. 653 00:32:16,130 --> 00:32:20,110 because they profit of global warming or of technology that leads to global warming, 654 00:32:20,110 --> 00:32:22,850 which is very bad for all of us. *applause* 655 00:32:22,850 --> 00:32:28,850 So our society is a nervous system that lies to itself. 656 00:32:28,850 --> 00:32:30,429 How can we overcome this? 657 00:32:30,429 --> 00:32:32,480 Actually, we don't know. 658 00:32:32,480 --> 00:32:34,639 To do this we would need to have some kind of centrialized, 659 00:32:34,639 --> 00:32:36,660 top-down reward motivational system. 660 00:32:36,660 --> 00:32:39,010 We have this for instance in the military, 661 00:32:39,010 --> 00:32:42,520 you have this system of military rewards that you get. 662 00:32:42,520 --> 00:32:44,950 And this are completely controlled from the top. 663 00:32:44,950 --> 00:32:47,260 Also within working organizations you have this. 664 00:32:47,260 --> 00:32:49,600 In corporations you have centralized rewards, 665 00:32:49,600 --> 00:32:51,850 it's not like rewards flow bottom-up, 666 00:32:51,850 --> 00:32:55,120 they always flown top-down. 667 00:32:55,120 --> 00:32:57,850 And there was an attempt to model society in such a way. 668 00:32:57,850 --> 00:33:03,380 That was in Chile in the early 1970, the Allende government had the idea 669 00:33:03,380 --> 00:33:07,320 to redesign society or economy in society using cybernetics. 670 00:33:07,320 --> 00:33:12,590 So Allende invited a bunch of cyberneticians to redesign the Chilean economy. 671 00:33:12,590 --> 00:33:14,550 And this was meant to be the control room, 672 00:33:14,550 --> 00:33:17,460 where Allende and his chief economists would be sitting, 673 00:33:17,460 --> 00:33:19,709 to look at what the economy is doing. 674 00:33:19,709 --> 00:33:23,880 We don't know how this would work out, because we know how it ended. 675 00:33:23,880 --> 00:33:27,260 In 1973 there was this big putsch in Chile, 676 00:33:27,260 --> 00:33:30,290 and this experiment ended among other things. 677 00:33:30,290 --> 00:33:34,170 Maybe it would have worked, who knows? Nobody tried it. 678 00:33:34,170 --> 00:33:38,370 So, there is something else what is going on in people, 679 00:33:38,370 --> 00:33:40,030 beyond the motivational system. 680 00:33:40,030 --> 00:33:43,610 That is: we have social criteria, for learning. 681 00:33:43,610 --> 00:33:47,670 We also check if our ideas are normativly acceptable. 682 00:33:47,670 --> 00:33:50,510 And this is actually a good thing, because individual may shortcut 683 00:33:50,510 --> 00:33:52,590 the learning through communication. 684 00:33:52,590 --> 00:33:55,260 Other people have learned stuff that we don't need to learn ourselves. 685 00:33:55,260 --> 00:33:59,800 We can build on this, so we can accelerate learning by many order of magnitutde, 686 00:33:59,800 --> 00:34:00,970 which makes culture possible. 687 00:34:00,970 --> 00:34:04,190 And which makes many anything possible, because if you were on your own 688 00:34:04,190 --> 00:34:06,860 you would not be going to find out very much in your lifetime. 689 00:34:08,520 --> 00:34:11,270 You know how they say? Everything that you do, 690 00:34:11,270 --> 00:34:14,250 you do by standing on the shoulders of giants. 691 00:34:14,250 --> 00:34:17,779 Or on a big pile of dwarfs it works either way. 692 00:34:17,779 --> 00:34:27,089 *laughter**applause* 693 00:34:27,089 --> 00:34:30,379 Social learning usually outperforms individual learning. You can test this. 694 00:34:30,379 --> 00:34:33,949 But in the case of conflict between different social truths, 695 00:34:33,949 --> 00:34:36,659 you need some way to decide who to believe. 696 00:34:36,659 --> 00:34:39,498 So you have some kind of reputation estimate for different authority, 697 00:34:39,498 --> 00:34:42,399 and you use this to check whom you believe. 698 00:34:42,399 --> 00:34:45,748 And the problem of course is this in existing society, in real society, 699 00:34:45,748 --> 00:34:48,389 this reputation system is going to reflect power structure, 700 00:34:48,389 --> 00:34:51,699 which may distort your belief systematically. 701 00:34:51,699 --> 00:34:54,759 Social learning therefore leads groups to synchronize their opinions. 702 00:34:54,759 --> 00:34:57,220 And the opinions become ...get another role. 703 00:34:57,220 --> 00:35:02,180 They become important part of signalling which group you belong to. 704 00:35:02,180 --> 00:35:06,630 So opinions start to signal group loyalty in societies. 705 00:35:06,630 --> 00:35:11,170 And people in this, and that's the actual world, they should optimize not for getting the best possible 706 00:35:11,170 --> 00:35:12,619 opinions in terms of truth. 707 00:35:12,619 --> 00:35:17,289 They should guess... they should optimize for doing... having the best possible opinion, 708 00:35:17,289 --> 00:35:19,799 with respect to agreement with their peers. 709 00:35:19,799 --> 00:35:22,029 If you have the same opinion as your peers, you can signal them 710 00:35:22,029 --> 00:35:24,299 that you are the part of their ingroup, they are going to like you. 711 00:35:24,299 --> 00:35:28,160 If you don't do this, chances are they are not going to like you. 712 00:35:28,160 --> 00:35:34,049 There is rarely any benefit in life to be in disagreement with your boss. Right? 713 00:35:34,049 --> 00:35:39,230 So, if you evolve an opinion forming system in these curcumstances, 714 00:35:39,230 --> 00:35:41,220 you should be ending up with an opinion forming system, 715 00:35:41,220 --> 00:35:42,980 that leaves you with the most usefull opinion, 716 00:35:42,980 --> 00:35:45,400 which is the opinion in your environment. 717 00:35:45,400 --> 00:35:48,400 And it turns out, most people are able to do this effortlessly. 718 00:35:48,400 --> 00:35:50,969 *laughter* 719 00:35:50,969 --> 00:35:55,529 They have an instinct, that makes them adapt the dominant opinion in their social environment. 720 00:35:55,529 --> 00:35:56,599 It's amazing, right? 721 00:35:56,599 --> 00:36:01,040 And if you are nerd like me, you don't get this. 722 00:36:01,040 --> 00:36:08,999 *lauging**applause* 723 00:36:08,999 --> 00:36:12,999 So in the world out there, explanations piggyback on you group allegiance. 724 00:36:12,999 --> 00:36:15,900 For instance you will find that there is a substantial group of people that believes 725 00:36:15,900 --> 00:36:18,380 the minimum wage is good for the economy and for you 726 00:36:18,380 --> 00:36:20,549 and another one believes that its bad. 727 00:36:20,549 --> 00:36:23,470 And its pretty much aligned with political parties. 728 00:36:23,470 --> 00:36:25,970 Its not aligned with different understandings of economy, 729 00:36:25,970 --> 00:36:30,740 because nobody understands how the economy works. 730 00:36:30,740 --> 00:36:36,330 And if you are a nerd you try to understand the world in terms of what is true and false. 731 00:36:36,330 --> 00:36:40,680 You try to prove everything by putting it in some kind of true and false level 732 00:36:40,680 --> 00:36:43,589 and if you are not a nerd you try to get to right and wrong 733 00:36:43,589 --> 00:36:45,609 you try to understand whether you are in alignment 734 00:36:45,609 --> 00:36:49,559 with what's objectively right in your society, right? 735 00:36:49,559 --> 00:36:55,680 So I guess that nerds are people that have a defect in there opinion forming system. 736 00:36:55,680 --> 00:36:57,069 *laughing* 737 00:36:57,069 --> 00:37:00,609 And usually that's maladaptive and under normal circumstances 738 00:37:00,609 --> 00:37:03,099 nerds would mostly be filtered from the world, 739 00:37:03,099 --> 00:37:06,529 because they don't reproduce so well, because people don't like them so much. 740 00:37:06,529 --> 00:37:07,960 *laughing* 741 00:37:07,960 --> 00:37:11,119 And then something very strange happened. The computer revolution came along and 742 00:37:11,119 --> 00:37:14,170 suddenly if you argue with the computer it doesn't help you if you have the 743 00:37:14,170 --> 00:37:17,849 normatively correct opinion you need to be able to understand things in terms of 744 00:37:17,849 --> 00:37:26,029 true and false, right? *applause* 745 00:37:26,029 --> 00:37:29,779 So now we have this strange situation that the weird people that have this offensive, 746 00:37:29,779 --> 00:37:33,410 strange opinions and that really don't mix well with the real normal people 747 00:37:33,410 --> 00:37:38,119 get all this high paying jobs and we don't understand how is that happening. 748 00:37:38,119 --> 00:37:42,599 And it's because suddenly our maladapting is a benefit. 749 00:37:42,599 --> 00:37:47,300 But out there there is this world of the social norms and it's made of paperwalls. 750 00:37:47,300 --> 00:37:50,349 There are all this things that are true and false in a society that make 751 00:37:50,349 --> 00:37:51,549 people behave. 752 00:37:51,549 --> 00:37:57,390 It's like this japanese wall, there. They made palaces out of paper basically. 753 00:37:57,390 --> 00:38:00,339 And these are walls by convention. 754 00:38:00,339 --> 00:38:04,009 They exist because people agree that this is a wall. 755 00:38:04,009 --> 00:38:06,630 And if you are a hypnotist like Donald Trump 756 00:38:06,630 --> 00:38:11,109 you can see that these are paper walls and you can shift them. 757 00:38:11,109 --> 00:38:14,079 And if you are a nerd like me you can not see these paperwalls. 758 00:38:14,079 --> 00:38:20,230 If you pay closely attention you see that people move and then suddenly middair 759 00:38:20,230 --> 00:38:22,869 they make a turn. Why would they do this? 760 00:38:22,869 --> 00:38:24,360 There must be something that they see there 761 00:38:24,360 --> 00:38:26,549 and this is basically a normative agreement. 762 00:38:26,549 --> 00:38:29,690 And you can infer what this is and then you can manipulate it and understand it. 763 00:38:29,690 --> 00:38:32,640 Of course you can't fix this, you can debug yourself in this regard, 764 00:38:32,640 --> 00:38:34,690 but it's something that is hard to see for nerds. 765 00:38:34,690 --> 00:38:38,109 So in some sense they have a superpower: they can think straight in the presence 766 00:38:38,109 --> 00:38:39,079 of others. 767 00:38:39,079 --> 00:38:42,590 But often they end up in their living room and people are upset. 768 00:38:42,590 --> 00:38:45,810 *laughter* 769 00:38:45,810 --> 00:38:49,789 Learning in a complex domain can not guarantee that you find the global maximum. 770 00:38:49,789 --> 00:38:53,970 We know that we can not find truth because we can not recognize whether we live 771 00:38:53,970 --> 00:38:57,059 on a plain field or on a simulated plain field. 772 00:38:57,059 --> 00:39:00,579 But what we can do is, we can try to approach a global maximum. 773 00:39:00,579 --> 00:39:02,339 But we don't know if that is the global maximum. 774 00:39:02,339 --> 00:39:05,509 We will always move along some kind of belief gradient. 775 00:39:05,509 --> 00:39:09,110 We will take certain elements of our belief and then give them up 776 00:39:09,110 --> 00:39:12,650 for new elements of a belief based on thinking, that this new element 777 00:39:12,650 --> 00:39:15,049 of belief is better than the one we give up. 778 00:39:15,049 --> 00:39:17,079 So we always move along some kind of gradient. 779 00:39:17,079 --> 00:39:19,789 and the truth does not matter, the gradient matters. 780 00:39:19,789 --> 00:39:23,650 If you think about teaching for a moment, when I started teaching I often thought: 781 00:39:23,650 --> 00:39:27,489 Okay, I understand the truth of the subject, the students don't, so I have to 782 00:39:27,489 --> 00:39:30,069 give this to them and at some point I realized: 783 00:39:30,069 --> 00:39:33,450 Oh, I changed my mind so many times in the past and I'm probably not going to 784 00:39:33,450 --> 00:39:35,769 stop changing it in the future. 785 00:39:35,769 --> 00:39:38,710 I'm always moving along a gradient and I keep moving along a gradient. 786 00:39:38,710 --> 00:39:43,099 So I'm not moving to truth, I'm moving forward. 787 00:39:43,099 --> 00:39:45,230 And when we teach our kids we should probably not think about 788 00:39:45,230 --> 00:39:46,390 how to give them truth. 789 00:39:46,390 --> 00:39:51,039 We should think about how to put them onto an interesting gradient, that makes them 790 00:39:51,039 --> 00:39:55,079 explore the world, world of possible beliefs. 791 00:39:55,079 --> 00:40:03,150 *applause* 792 00:40:03,150 --> 00:40:05,359 And this possible beliefs lead us into local minima. 793 00:40:05,359 --> 00:40:08,150 This is inevitable. This are like valleys and sometimes this valleys are 794 00:40:08,150 --> 00:40:11,210 neighbouring and we don't understand what the people in the neighbouring 795 00:40:11,210 --> 00:40:15,700 valley are doing unless we are willing to retrace the steps they have been taken. 796 00:40:15,700 --> 00:40:19,569 And if you want to get from one valley into the next, we will have to have some kind 797 00:40:19,569 --> 00:40:21,789 of energy that moves us over the hill. 798 00:40:21,789 --> 00:40:27,739 We have to have a trajectory were every step works by finding reason to give up 799 00:40:27,739 --> 00:40:30,380 bit of our current belief and adopt a new belief, because it's somehow 800 00:40:30,380 --> 00:40:34,739 more useful, more relevant, more consistent and so on. 801 00:40:34,739 --> 00:40:38,349 Now the problem is that this is not monotonous we can not guarantee that 802 00:40:38,349 --> 00:40:40,499 we're always climbing, because the problem is, that 803 00:40:40,499 --> 00:40:44,599 the beliefs themselfs can change our evaluation of the belief. 804 00:40:44,599 --> 00:40:50,390 It could be for instance that you start believing in a religion and this religion 805 00:40:50,390 --> 00:40:54,299 could tell you: If you give up the belief in the religion, you're going to face 806 00:40:54,299 --> 00:40:56,500 eternal damnation in hell. 807 00:40:56,500 --> 00:40:59,489 As long as you believe in the religion, it's going to be very expensive for you 808 00:40:59,489 --> 00:41:02,430 to give up the religion, right? If you truly belief in it. 809 00:41:02,430 --> 00:41:05,109 You're now caught in some kind of attractor. 810 00:41:05,109 --> 00:41:08,680 Before you believe the religion it is not very dangerous but once you've gotten 811 00:41:08,680 --> 00:41:13,019 into the attractor it's very, very hard to get out. 812 00:41:13,019 --> 00:41:16,309 So these belief attractors are actually quite dangerous. 813 00:41:16,309 --> 00:41:19,920 You can get not only to chaotic behaviour, where you can not guarantee that your 814 00:41:19,920 --> 00:41:23,470 current belief is better than the last one but you can also get into beliefs that are 815 00:41:23,470 --> 00:41:26,849 almost impossible to change. 816 00:41:26,849 --> 00:41:33,739 And that makes it possible to program people to work in societies. 817 00:41:33,739 --> 00:41:37,529 Social domains are structured by values. Basically a preference is what makes you 818 00:41:37,529 --> 00:41:40,769 do things, because you anticipate pleasure or displeasure, 819 00:41:40,769 --> 00:41:45,339 and values make you do things even if you don't anticipate any pleasure. 820 00:41:45,339 --> 00:41:49,809 These are virtual rewards. They make us do things, because we believe 821 00:41:49,809 --> 00:41:51,799 that is stuff that is more important then us. 822 00:41:51,799 --> 00:41:55,109 This is what values are about. 823 00:41:55,109 --> 00:42:00,690 And these values are the source of what we would call true meaning, deeper meaning. 824 00:42:00,690 --> 00:42:05,220 There is something that is more important than us, something that we can serve. 825 00:42:05,220 --> 00:42:08,769 This is what we usually perceive as meaningful life, it is one which 826 00:42:08,769 --> 00:42:12,759 is in the serves of values that are more important than I myself, 827 00:42:12,759 --> 00:42:15,749 because after all I'm not that important. I'm just this machine that runs around 828 00:42:15,749 --> 00:42:20,789 and tries to optimize its pleasure and pain, which is kinda boring. 829 00:42:20,789 --> 00:42:26,329 So my PI has puzzled me, my principle investigator in the Havard department, 830 00:42:26,329 --> 00:42:29,349 where I have my desk, Martin Nowak. 831 00:42:29,349 --> 00:42:33,970 He said, that meaning can not exist without god; you are either religious, 832 00:42:33,970 --> 00:42:36,950 or you are a nihilist. 833 00:42:36,950 --> 00:42:42,789 And this guy is the head of the department for evolutionary dynamics. 834 00:42:42,789 --> 00:42:45,769 Also he is a catholic.. *chuckling* 835 00:42:45,769 --> 00:42:49,729 So this really puzzled me and I tried to understand what he meant by this. 836 00:42:49,729 --> 00:42:53,200 Typically if you are a good atheist like me, 837 00:42:53,200 --> 00:42:57,920 you tend to attack gods that are structured like this, religious gods, 838 00:42:57,920 --> 00:43:02,940 that are institutional, they are personal, they are some kind of person. 839 00:43:02,940 --> 00:43:08,239 They do care about you, they prescribe norms, for instance don't mastrubate 840 00:43:08,239 --> 00:43:10,060 it's bad for you. 841 00:43:10,060 --> 00:43:14,759 Many of this norms are very much aligned with societal institutions, for instance 842 00:43:14,759 --> 00:43:20,799 don't questions the authorities, god wants them to be ruling above you 843 00:43:20,799 --> 00:43:23,839 and be monogamous and so on and so on. 844 00:43:23,839 --> 00:43:28,979 So they prescribe norms that do not make a lot of sense in terms of beings that 845 00:43:28,979 --> 00:43:31,200 creates world every now and then, 846 00:43:31,200 --> 00:43:34,619 but they make sense in terms of what you should be doing to be a 847 00:43:34,619 --> 00:43:36,730 functioning member of society. 848 00:43:36,730 --> 00:43:40,799 And this god also does things like it creates world, they like to manifest as 849 00:43:40,799 --> 00:43:43,660 burning shrubbery and so on. There are many books that describe stories that 850 00:43:43,660 --> 00:43:45,700 these gods have allegedly done. 851 00:43:45,700 --> 00:43:48,819 And it's very hard to test for all these features which makes this gods very 852 00:43:48,819 --> 00:43:54,280 improbable for us. And makes Atheist very dissatisfied with these gods. 853 00:43:54,280 --> 00:43:56,569 But then there is a different kind of god. 854 00:43:56,569 --> 00:43:58,599 This is what we call the spiritual god. 855 00:43:58,599 --> 00:44:02,410 This spiritual god is independent of institutions, it still does care about you. 856 00:44:02,410 --> 00:44:06,489 It's probably conscious. It might not be a person. There are not that many stories, 857 00:44:06,489 --> 00:44:10,579 that you can consistently tell about it, but you might be able to connect to it 858 00:44:10,579 --> 00:44:15,259 spiritually. 859 00:44:15,259 --> 00:44:19,470 Then there is a god that is even less expensive. That is god as a transcendental 860 00:44:19,470 --> 00:44:23,489 principle and this god is simply the reason why there is something rather then 861 00:44:23,489 --> 00:44:28,150 nothing. This god is the question the universe is the answer to, this is the 862 00:44:28,150 --> 00:44:29,600 thing that gives meaning. 863 00:44:29,600 --> 00:44:31,489 Everything else about it is unknowable. 864 00:44:31,489 --> 00:44:34,190 This is the god of Thomas of Aquinus. 865 00:44:34,190 --> 00:44:38,089 The God that Thomas of Aquinus discovered is not the god of Abraham this is not the 866 00:44:38,089 --> 00:44:39,180 religious god. 867 00:44:39,180 --> 00:44:43,559 It's a god that is basically a principle that us ... the universe into existence. 868 00:44:43,559 --> 00:44:47,140 It's the one that gives the universe it's purpose. 869 00:44:47,140 --> 00:44:50,200 And because every other property is unknowable about this, 870 00:44:50,200 --> 00:44:52,010 this god is not that expensive. 871 00:44:52,010 --> 00:44:55,960 Unfortunately it doesn't really work. I mean Thomas of Aquinus tried to prove 872 00:44:55,960 --> 00:45:00,049 god. He tried to prove an necessary god, a god that has to be existing and 873 00:45:00,049 --> 00:45:02,779 I think we can only prove a possible god. 874 00:45:02,779 --> 00:45:05,339 So if you try to prove a necessary god, this god can not exist. 875 00:45:05,339 --> 00:45:11,650 Which means your god prove is going to fail. You can only prove possible gods. 876 00:45:11,650 --> 00:45:13,259 And then there is an even more improper god. 877 00:45:13,259 --> 00:45:15,890 And that's the god of Aristotle and he said: 878 00:45:15,890 --> 00:45:20,069 "If there is change in the universe, something in going to have to change it." 879 00:45:20,069 --> 00:45:23,640 There must be something that moves it along from one state to the next. 880 00:45:23,640 --> 00:45:26,289 So I would say that is the primary computational transition function 881 00:45:26,289 --> 00:45:35,079 of the universe. *laughing* *applause* 882 00:45:35,079 --> 00:45:38,439 And Aristotle discovered it. It's amazing isn't it? 883 00:45:38,439 --> 00:45:41,509 We have to have this because we can not be conscious in a single state. 884 00:45:41,509 --> 00:45:43,279 We need to move between states to be conscious. 885 00:45:43,279 --> 00:45:45,979 We need to be processes. 886 00:45:45,979 --> 00:45:50,859 So we can take our gods and sort them by their metaphysical cost. 887 00:45:50,859 --> 00:45:53,290 The 1st degree god would be the first mover. 888 00:45:53,290 --> 00:45:56,069 The 2nd degree god is the god of purpose and meaning. 889 00:45:56,069 --> 00:45:59,089 3rd degree god is the spiritual god. And the 4th degree god is this bound to 890 00:45:59,089 --> 00:46:01,229 religious institutions, right? 891 00:46:01,229 --> 00:46:03,720 So if you take this statement from Martin Nowak, 892 00:46:03,720 --> 00:46:07,759 "You can not have meaning without god!" I would say: yes! You need at least 893 00:46:07,759 --> 00:46:14,990 a 2nd degree god to have meaning. So objective meaning can only exist 894 00:46:14,990 --> 00:46:19,119 with a 2nd degree god. *chuckling* 895 00:46:19,119 --> 00:46:22,269 And subjective meaning can exist as a function in a cognitive system of course. 896 00:46:22,269 --> 00:46:24,180 We don't need objective meaning. 897 00:46:24,180 --> 00:46:27,410 So we can subjectively feel that there is something more important to us 898 00:46:27,410 --> 00:46:30,509 and this makes us work in society and makes us perceive that we have values 899 00:46:30,509 --> 00:46:34,329 and so on, but we don't need to believe that there is something outside of the 900 00:46:34,329 --> 00:46:36,869 universe to have this. 901 00:46:36,869 --> 00:46:40,650 So the 4th degree god is the one that is bound to religious institutions, 902 00:46:40,650 --> 00:46:45,400 it requires a belief attractor and it enables complex norm prescriptions. 903 00:46:45,400 --> 00:46:48,430 It my theory is right then it should be much harder for nerds to believe in 904 00:46:48,430 --> 00:46:52,039 a 4th degree god then for normal people. 905 00:46:52,039 --> 00:46:56,489 And what this god does it allows you to have state building mind viruses. 906 00:46:56,489 --> 00:47:00,269 Basically religion is a mind virus. And the amazing thing about these mind viruses 907 00:47:00,269 --> 00:47:02,489 is that they structure behaviour in large groups. 908 00:47:02,489 --> 00:47:06,130 We have evolved to live in small groups of a few 100 individuals, maybe somthing 909 00:47:06,130 --> 00:47:07,249 like a 150. 910 00:47:07,249 --> 00:47:10,059 This is roughly the level to which reputation works. 911 00:47:10,059 --> 00:47:15,369 We can keep track of about 150 people and after this it gets much much worse. 912 00:47:15,369 --> 00:47:18,290 So in this system where you have reputation people feel responsible 913 00:47:18,290 --> 00:47:21,349 for each other and they can keep track of their doings 914 00:47:21,349 --> 00:47:23,049 and society kind of sort of works. 915 00:47:23,049 --> 00:47:27,789 If you want to go beyond this, you have to right a software that controls people. 916 00:47:27,789 --> 00:47:32,420 And religions were the first software, that did this on a very large scale. 917 00:47:32,420 --> 00:47:35,319 And in order to keep stable they had to be designed like operating systems 918 00:47:35,319 --> 00:47:36,039 in some sense. 919 00:47:36,039 --> 00:47:39,930 They give people different roles like insects in a hive. 920 00:47:39,930 --> 00:47:44,529 And they have even as part of this roles is to update this religion but it has to be 921 00:47:44,529 --> 00:47:48,380 done very carefully and centrally because otherwise the religion will split apart 922 00:47:48,380 --> 00:47:51,719 and fall together into new religions or be overcome by new ones. 923 00:47:51,719 --> 00:47:54,259 So there is some kind of evolutionary dynamics that goes on 924 00:47:54,259 --> 00:47:55,930 with respect to religion. 925 00:47:55,930 --> 00:47:58,519 And if you look the religions, there is actually a veritable evolution 926 00:47:58,519 --> 00:47:59,739 of religions. 927 00:47:59,739 --> 00:48:04,789 So we have this Israelic tradition and the Mesoputanic mythology that gave rise 928 00:48:04,789 --> 00:48:13,019 to Judaism. *applause* 929 00:48:13,019 --> 00:48:16,299 It's kind of cool, right? *laughing* 930 00:48:16,299 --> 00:48:36,289 Also history totally repeats itself. *roaring laughter* *applause* 931 00:48:36,289 --> 00:48:41,889 Yeah, it totally blew my mind when I discovered this. *laughter* 932 00:48:41,889 --> 00:48:45,039 Of course the real tree of programming languages is slightly more complicated, 933 00:48:45,039 --> 00:48:48,599 And the real tree of religion is slightly more complicated. 934 00:48:48,599 --> 00:48:51,229 But still its neat. 935 00:48:51,229 --> 00:48:54,289 So if you want to immunize yourself against mind viruses, 936 00:48:54,289 --> 00:48:58,570 first of all you want to check yourself whether you are infected. 937 00:48:58,570 --> 00:49:02,809 You should check: Can I let go of my current beliefs without feeling that 938 00:49:02,809 --> 00:49:07,670 meaning departures me and I feel very terrible, when I let go of my beliefs. 939 00:49:07,670 --> 00:49:11,279 Also you should check: All the other people around there that don't 940 00:49:11,279 --> 00:49:17,019 share my belief, are they either stupid, or crazy, or evil? 941 00:49:17,019 --> 00:49:19,890 If you think this chances are you are infected by some kind of mind virus, 942 00:49:19,890 --> 00:49:23,710 because they are just part of the out group. 943 00:49:23,710 --> 00:49:28,059 And does your god have properties that you know but you did not observe. 944 00:49:28,059 --> 00:49:32,490 So basically you have a god of 2nd or 3rd degree or higher. 945 00:49:32,490 --> 00:49:34,589 In this case you also probably got a mind virus. 946 00:49:34,589 --> 00:49:37,259 There is nothing wrong with having a mind virus, 947 00:49:37,259 --> 00:49:39,920 but if you want to immunize yourself against this people have invented 948 00:49:39,920 --> 00:49:44,059 rationalism and enlightenment, basically to act as immunization against 949 00:49:44,059 --> 00:49:50,660 mind viruses. *loud applause* 950 00:49:50,660 --> 00:49:53,869 And in some sense its what the mind does by itself because, if you want to 951 00:49:53,869 --> 00:49:56,949 understand how you go wrong, you need to have a mechanism 952 00:49:56,949 --> 00:49:58,839 that discovers who you are. 953 00:49:58,839 --> 00:50:03,109 Some kind of auto debugging mechanism, that makes the mind aware of itself. 954 00:50:03,109 --> 00:50:04,779 And this is actually the self. 955 00:50:04,779 --> 00:50:08,339 So according to Robert Kegan: "The development of ourself is a process, 956 00:50:08,339 --> 00:50:13,400 in which we learn who we are by making thing explicit", by making processes that 957 00:50:13,400 --> 00:50:17,249 are automatic visible to us and by conceptualize them so we no longer 958 00:50:17,249 --> 00:50:18,859 identify with them. 959 00:50:18,859 --> 00:50:22,019 And it starts out with understanding that there is only pleasure and pain. 960 00:50:22,019 --> 00:50:25,180 If you are a baby, you have only pleasure and pain you identify with this. 961 00:50:25,180 --> 00:50:27,869 And then you turn into a toddler and the toddler understands that they are not 962 00:50:27,869 --> 00:50:31,059 their pleasure and pain but they are their impulses. 963 00:50:31,059 --> 00:50:34,259 And in the next level if you grow beyond the toddler age you actually know that 964 00:50:34,259 --> 00:50:38,880 you have goals and that your needs and impulses are there to serve goals, but its 965 00:50:38,880 --> 00:50:40,210 very difficult to let go of the goals, 966 00:50:40,210 --> 00:50:42,789 if you are a very young child. 967 00:50:42,789 --> 00:50:46,329 And at some point you realize: Oh, the goals don't really matter, because 968 00:50:46,329 --> 00:50:49,509 sometimes you can not reach them, but we have preferences, we have thing that we 969 00:50:49,509 --> 00:50:52,950 want to happen and thing that we do not want to happen. And then at some point 970 00:50:52,950 --> 00:50:55,869 we realize that other people have preferences, too. 971 00:50:55,869 --> 00:50:58,979 And then we start to model the world as a system where different people have 972 00:50:58,979 --> 00:51:01,940 different preferences and we have to navigate this landscape. 973 00:51:01,940 --> 00:51:06,420 And then we realize that this preferences also relate to values and we start 974 00:51:06,420 --> 00:51:09,700 to identify with this values as members of society. 975 00:51:09,700 --> 00:51:13,469 And this is basically the stage if you are an adult being, that you get into. 976 00:51:13,469 --> 00:51:16,910 And you can get to a stage beyond that, especially if you have people this, which 977 00:51:16,910 --> 00:51:20,059 have already done this. And this means that you understand that people have 978 00:51:20,059 --> 00:51:23,660 different values and what they do naturally flows out of them. 979 00:51:23,660 --> 00:51:26,849 And this values are not necessarily worse than yours they are just different. 980 00:51:26,849 --> 00:51:29,450 And you learn that you can hold different sets of values in your mind at 981 00:51:29,450 --> 00:51:33,019 the same time, isn't that amazing? and understand other people, even if 982 00:51:33,019 --> 00:51:36,660 they are not part of your group. If you get that, this is really good. 983 00:51:36,660 --> 00:51:39,269 But I don't think it stops there. 984 00:51:39,269 --> 00:51:43,019 You can also learn that the stuff that you perceive is kind of incidental, 985 00:51:43,019 --> 00:51:45,339 that you can turn it of and you can manipulate it. 986 00:51:45,339 --> 00:51:49,940 And at some point you also can realize that yourself is only incidental that you 987 00:51:49,940 --> 00:51:52,559 can manipulate it or turn it of. And that your basically some kind of 988 00:51:52,559 --> 00:51:57,420 consciousness that happens to run a brain of some kind of person, that navigates 989 00:51:57,420 --> 00:52:04,279 the world in terms to get rewards or avoid displeasure and serve values and so on, 990 00:52:04,279 --> 00:52:05,130 but it doesn't really matter. 991 00:52:05,130 --> 00:52:08,119 There is just this consciousness which understands the world. 992 00:52:08,119 --> 00:52:11,009 And this is the stage that we typically call enlightenment. 993 00:52:11,009 --> 00:52:14,549 In this stage you realize that you are not your brain, but you are a story that 994 00:52:14,549 --> 00:52:25,640 your brain tells itself. *applause* 995 00:52:25,640 --> 00:52:29,630 So becoming self aware is a process of reverse engineering your mind. 996 00:52:29,630 --> 00:52:32,890 Its a different set of stages in which to realize what goes on. 997 00:52:32,890 --> 00:52:33,799 So isn't that amazing. 998 00:52:33,799 --> 00:52:38,930 AI is a way to get to more self awareness? 999 00:52:38,930 --> 00:52:41,319 I think that is a good point to stop here. 1000 00:52:41,319 --> 00:52:44,499 The first talk that I gave in this series was 2 years ago. It was about 1001 00:52:44,499 --> 00:52:45,979 how to build a mind. 1002 00:52:45,979 --> 00:52:49,670 Last year I talked about how to get from basic computation to consciousness. 1003 00:52:49,670 --> 00:52:53,709 And this year we have talked about finding meaning using AI. 1004 00:52:53,709 --> 00:52:57,470 I wonder where it goes next. *laughter* 1005 00:52:57,470 --> 00:53:22,769 *applause* 1006 00:53:22,769 --> 00:53:26,489 Herald: Thank you for this amazing talk! We now have some minutes for Q&A. 1007 00:53:26,489 --> 00:53:31,190 So please line up at the microphones as always. If you are unable to stand up 1008 00:53:31,190 --> 00:53:36,430 for some reason please very very visibly rise your hand, we should be able to dispatch 1009 00:53:36,430 --> 00:53:40,099 an audio angle to your location so you can have a question too. 1010 00:53:40,099 --> 00:53:44,030 And also if you are locationally disabled, you are not actually in the room 1011 00:53:44,030 --> 00:53:49,069 if you are on the stream, you can use IRC or twitter to also ask questions. 1012 00:53:49,069 --> 00:53:50,989 We also have a person for that. 1013 00:53:50,989 --> 00:53:53,779 We will start at microphone number 2. 1014 00:53:53,779 --> 00:53:59,940 Q: Wow that's me. Just a guess! What would you guess, when can you discuss 1015 00:53:59,940 --> 00:54:04,559 your talk with a machine, in how many years? 1016 00:54:04,559 --> 00:54:07,400 Joscha: I don't know! As a software engineer I know if I don't have the 1017 00:54:07,400 --> 00:54:12,619 specification all bets are off, until I have the implementation. *laughter* 1018 00:54:12,619 --> 00:54:14,509 So it can be of any order of magnitude. 1019 00:54:14,509 --> 00:54:18,249 I have a gut feeling but I also know as a software engineer that my gut feeling is 1020 00:54:18,249 --> 00:54:23,450 usually wrong, *laughter* until I have the specification. 1021 00:54:23,450 --> 00:54:28,200 So the question is if there are silver bullets? Right now there are some things 1022 00:54:28,200 --> 00:54:30,569 that are not solved yet and it could be that they are easier to solve 1023 00:54:30,569 --> 00:54:33,469 than we think, but it could be that they're harder to solve than we think. 1024 00:54:33,469 --> 00:54:36,710 Before I stumbled on this cortical self organization thing, 1025 00:54:36,710 --> 00:54:40,719 I thought it's going to be something like maybe 60, 80 years and now I think it's 1026 00:54:40,719 --> 00:54:47,289 way less, but again this is a very subjective perspective. I don't know. 1027 00:54:47,289 --> 00:54:49,240 Herald: Number 1, please! 1028 00:54:49,240 --> 00:54:55,589 Q: Yes, I wanted to ask a little bit about metacognition. It seems that you kind of 1029 00:54:55,589 --> 00:55:01,329 end your story saying that it's still reflecting on input that you get and 1030 00:55:01,329 --> 00:55:04,900 kind of working with your social norms and this and that, but Colberg 1031 00:55:04,900 --> 00:55:11,839 for instance talks about what he calls a postconventional universal morality 1032 00:55:11,839 --> 00:55:17,420 for instance, which is thinking about moral laws without context, basically 1033 00:55:17,420 --> 00:55:23,069 stating that there is something beyond the relative norm that we have to each other, 1034 00:55:23,069 --> 00:55:29,579 which would only be possible if you can do kind of, you know, meta cognition, 1035 00:55:29,579 --> 00:55:32,599 thinking about your own thinking and then modifying that thinking. 1036 00:55:32,599 --> 00:55:37,229 So kind of feeding back your own ideas into your own mind and coming up with 1037 00:55:37,229 --> 00:55:43,779 stuff that actually can't get ... well processing external inputs. 1038 00:55:43,779 --> 00:55:48,469 Joscha: Mhm! I think it's very tricky. This project of defining morality without 1039 00:55:48,469 --> 00:55:53,119 societies exists longer than Kant of course. And Kant tried to give this 1040 00:55:53,119 --> 00:55:56,869 internal rules and others tried to. I find this very difficult. 1041 00:55:56,869 --> 00:56:01,069 From my perspective we are just moving bits of rocks. And this bits of rocks they 1042 00:56:01,069 --> 00:56:07,589 are on some kind of dust mode in a galaxy out of trillions of galaxies and how can 1043 00:56:07,589 --> 00:56:08,609 there be meaning? 1044 00:56:08,609 --> 00:56:11,180 It's very hard for me to say: 1045 00:56:11,180 --> 00:56:13,969 One chimpanzee species is better than another chimpanzee species or 1046 00:56:13,969 --> 00:56:16,559 a particular monkey is better than another monkey. 1047 00:56:16,559 --> 00:56:18,539 This only happens within a certain framework 1048 00:56:18,539 --> 00:56:20,160 and we have to set this framework. 1049 00:56:20,160 --> 00:56:23,700 And I don't think that we can define this framework outside of a context of 1050 00:56:23,700 --> 00:56:26,420 social norms, that we have to agree on. 1051 00:56:26,420 --> 00:56:29,650 So objectively I'm not sure if we can get to ethics. 1052 00:56:29,650 --> 00:56:33,769 I only think that is possible based on some kind of framework that people 1053 00:56:33,769 --> 00:56:38,339 have to agree on implicitly or explicitly. 1054 00:56:38,339 --> 00:56:40,630 Herald: Microphone number 4, please. 1055 00:56:40,630 --> 00:56:46,559 Q: Hi, thank you, it was a fascinating talk. I have 2 thought that went through my mind. 1056 00:56:46,559 --> 00:56:51,589 And the first one is that it's so convincing the models that you present, 1057 00:56:51,589 --> 00:56:56,709 but it's kind of like you present another metaphor of understanding the 1058 00:56:56,709 --> 00:57:01,670 brain which is still something that we try to grasp on different levels of science 1059 00:57:01,670 --> 00:57:07,469 basically. And the 2nd one is that your definition of the nerd who walks 1060 00:57:07,469 --> 00:57:10,950 and doesn't see the walls is kind of definition... or reminds me 1061 00:57:10,950 --> 00:57:15,229 Richard Rortys definition of the ironist which is a person who knows that their 1062 00:57:15,229 --> 00:57:20,799 vocabulary is finite and that other people have also a finite vocabulary and 1063 00:57:20,799 --> 00:57:24,599 then that obviously opens up the whole question of meaning making which has been 1064 00:57:24,599 --> 00:57:28,979 discussed in so many other disciplines and fields. 1065 00:57:28,979 --> 00:57:32,930 And I thought about Darridas deconstruction of ideas and thoughts and 1066 00:57:32,930 --> 00:57:36,300 Butler and then down the rabbit hole to Nietzsche and I was just wondering, 1067 00:57:36,300 --> 00:57:39,009 if you could maybe map out other connections 1068 00:57:39,009 --> 00:57:44,430 where basically not AI helping us to understand the mind, but where 1069 00:57:44,430 --> 00:57:49,819 already existing huge, huge fields of science, like cognitive process 1070 00:57:49,819 --> 00:57:53,359 coming from the other end could help us to understand AI. 1071 00:57:53,359 --> 00:57:59,680 Joscha: Thank you, the tradition that you mentioned Rorty and Butler and so on 1072 00:57:59,680 --> 00:58:02,989 are part of a completely different belief attractor in my current perspective. 1073 00:58:02,989 --> 00:58:06,209 That is they are mostly social constructionists. 1074 00:58:06,209 --> 00:58:10,880 They believe that reality at least in the domains of the mind and sociality 1075 00:58:10,880 --> 00:58:15,359 are social constructs they are part of social agreement. 1076 00:58:15,359 --> 00:58:17,190 Personally I don't think that this is the case. 1077 00:58:17,190 --> 00:58:19,630 I think that patterns that we refer to 1078 00:58:19,630 --> 00:58:23,890 are mostly independent of your mind. The norms are part of social constructs, 1079 00:58:23,890 --> 00:58:28,099 but for instance our motivational preferences that make us adapt or 1080 00:58:28,099 --> 00:58:32,719 reject norms, are something that builds up resistance to the environment. 1081 00:58:32,719 --> 00:58:35,660 So they are probably not part of social agreement. 1082 00:58:35,660 --> 00:58:41,569 And the only thing I can invite you to is try to retrace both of the different 1083 00:58:41,569 --> 00:58:45,640 belief attractors, try to retrace the different paths on the landscape. 1084 00:58:45,640 --> 00:58:48,529 All this thing that I tell you, all of this is of course very speculative. 1085 00:58:48,529 --> 00:58:52,390 These are that seem to be logical to me at this point in my life. 1086 00:58:52,390 --> 00:58:55,400 And I try to give you the arguments why I think that is plausible, but don't 1087 00:58:55,400 --> 00:58:59,109 believe in them, question them, challenge them, see if they work for you! 1088 00:58:59,109 --> 00:59:00,559 I'm not giving you any truth. 1089 00:59:00,559 --> 00:59:05,720 I'm just going to give you suitable encodings according to my current perspective. 1090 00:59:05,720 --> 00:59:11,739 Q:Thank you! *applause* 1091 00:59:11,739 --> 00:59:15,099 Herald: The internet, please! 1092 00:59:19,179 --> 00:59:26,029 Signal angel: So, someone is asking if in this belief space you're talking about 1093 00:59:26,029 --> 00:59:30,109 how is it possible to get out of local minima? 1094 00:59:30,109 --> 00:59:33,959 And very related question as well: 1095 00:59:33,959 --> 00:59:38,530 Should we teach some momentum method to our children, 1096 00:59:38,530 --> 00:59:41,599 so we don't get stuck in a local minima. 1097 00:59:41,599 --> 00:59:44,829 Joscha: I believe at some level it's not possible to get out of a local minima. 1098 00:59:44,829 --> 00:59:50,329 In an absolute sense, because you only get to get into some kind of meta minimum, 1099 00:59:50,329 --> 00:59:56,769 but what you can do is to retrace the path that you took whenever you discover 1100 00:59:56,769 --> 00:59:59,989 that somebody else has a fundamentally different set of beliefs. 1101 00:59:59,989 --> 01:00:02,769 And if you realize that this person is basically a smart person that is not 1102 01:00:02,769 --> 01:00:07,359 completely insane but has reasons to believe in their beliefs and they seem to 1103 01:00:07,359 --> 01:00:10,579 be internally consistent it's usually worth to retrace what they 1104 01:00:10,579 --> 01:00:12,180 have been thinking and why. 1105 01:00:12,180 --> 01:00:15,930 And this means you have to understand where their starting point was and 1106 01:00:15,930 --> 01:00:18,279 how they moved from their current point to their starting point. 1107 01:00:18,279 --> 01:00:22,219 You might not be able to do this accurately and the important thing is 1108 01:00:22,219 --> 01:00:25,369 also afterwards you discover a second valley, you haven't discovered 1109 01:00:25,369 --> 01:00:27,059 the landscape inbetween. 1110 01:00:27,059 --> 01:00:30,839 But the only way that we can get an idea of the lay of the land is that we try to 1111 01:00:30,839 --> 01:00:33,200 retrace as many paths as possible. 1112 01:00:33,200 --> 01:00:36,339 And if we try to teach our children, what I think what we should be doing is: 1113 01:00:36,339 --> 01:00:38,650 To tell them how to explore this world on there own. 1114 01:00:38,650 --> 01:00:43,900 It's not that we tell them this is the valley, basically it's given, it's 1115 01:00:43,900 --> 01:00:47,599 the truth, but instead we have to tell them: This is the path that we took. 1116 01:00:47,599 --> 01:00:51,239 And these are the things that we saw inbetween and it is important to be not 1117 01:00:51,239 --> 01:00:54,390 completely naive when we go into this landscape, but we also have to understand 1118 01:00:54,390 --> 01:00:58,170 that it's always an exploration that never stops and that might change 1119 01:00:58,170 --> 01:01:01,140 everything that you believe now at a later point. 1120 01:01:01,140 --> 01:01:05,700 So for me it's about teaching my own children how to be explorers, 1121 01:01:05,700 --> 01:01:10,950 how to understand that knowledge is always changing and it's always a moving frontier. 1122 01:01:10,950 --> 01:01:17,230 *applause* 1123 01:01:17,230 --> 01:01:22,259 Herald: We are unfortunately out of time. So, please once again thank Joscha! 1124 01:01:22,259 --> 01:01:24,069 *applause* Joscha: Thank you! 1125 01:01:24,069 --> 01:01:28,239 *applause* 1126 01:01:28,239 --> 01:01:38,749 *postroll music*