1 00:00:00,268 --> 00:00:01,871 Inside the vector space of AI 2 00:00:01,948 --> 00:00:03,730 What lies near the words relativity 3 00:00:04,147 --> 00:00:06,030 If you could open the mind of AI 4 00:00:06,067 --> 00:00:08,191 And locate the word relativity 5 00:00:08,227 --> 00:00:09,250 Then look around 6 00:00:09,747 --> 00:00:11,051 What do you think you would see 7 00:00:11,468 --> 00:00:14,811 You might guess light speed gravitational waves and curved spacetime 8 00:00:15,307 --> 00:00:15,691 That is right 9 00:00:15,987 --> 00:00:17,510 They really are nearby 10 00:00:17,628 --> 00:00:18,690 But you would never guess 11 00:00:18,690 --> 00:00:20,146 That among the neighbors of relativity 12 00:00:20,147 --> 00:00:22,751 Are things that do not belong to physics at all 13 00:00:23,556 --> 00:00:23,860 Hello 14 00:00:23,956 --> 00:00:24,459 Hi everyone 15 00:00:24,797 --> 00:00:25,799 I am Leo Wang 16 00:00:26,157 --> 00:00:30,035 We once made an episode called A Trillion Parameters Are Not Knowledge 17 00:00:30,037 --> 00:00:31,459 They Are the Geometry of Cognition 18 00:00:32,037 --> 00:00:34,600 In that episode we explored the vector space of AI 19 00:00:34,716 --> 00:00:37,476 What it is and how every word and concept 20 00:00:37,476 --> 00:00:39,955 Is not a symbol inside the mind of AI 21 00:00:39,957 --> 00:00:41,240 But a coordinate 22 00:00:41,516 --> 00:00:44,799 Scattered like stars across thousands of dimensions 23 00:00:45,037 --> 00:00:45,979 Concepts that are close 24 00:00:46,277 --> 00:00:47,019 Are related in meaning 25 00:00:47,557 --> 00:00:48,540 Concepts that are far apart 26 00:00:48,796 --> 00:00:49,879 Are distant in meaning 27 00:00:50,156 --> 00:00:53,600 King minus man plus woman equals queen 28 00:00:53,716 --> 00:00:56,680 We covered that kind of semantic arithmetic last time 29 00:00:57,201 --> 00:00:58,924 Today we are leaving the theory aside 30 00:00:59,000 --> 00:01:01,064 Today I want to do something more interesting 31 00:01:01,481 --> 00:01:04,564 I want to take you through the neighborhood of relativity 32 00:01:04,640 --> 00:01:06,435 And meet its neighbors 33 00:01:06,632 --> 00:01:07,855 Why choose relativity 34 00:01:08,393 --> 00:01:10,476 Because the general theory of relativity by Einstein 35 00:01:10,512 --> 00:01:12,471 May be one of the most 36 00:01:12,473 --> 00:01:15,235 Precise and beautiful theories in human history 37 00:01:15,473 --> 00:01:20,276 It describes the nature of spacetime as mass tells space how to curve 38 00:01:20,393 --> 00:01:22,795 And space tells matter how to move 39 00:01:22,953 --> 00:01:25,196 It is the pinnacle of fundamental physics 40 00:01:25,432 --> 00:01:29,355 If any concept belongs securely in the drawer of physics 41 00:01:29,473 --> 00:01:30,875 It is relativity 42 00:01:31,333 --> 00:01:32,815 But AI does not use drawers 43 00:01:33,012 --> 00:01:34,896 AI has no system of academic disciplines 44 00:01:35,133 --> 00:01:37,396 It does not see this is physics 45 00:01:37,732 --> 00:01:38,611 This is philosophy 46 00:01:38,613 --> 00:01:39,235 This is religion 47 00:01:39,852 --> 00:01:42,896 It sees the distances between every concept 48 00:01:43,292 --> 00:01:43,595 So 49 00:01:43,932 --> 00:01:46,015 When you enter the vector space of AI 50 00:01:46,053 --> 00:01:48,295 And locate the coordinate for general relativity 51 00:01:48,372 --> 00:01:49,315 Then draw a circle around it 52 00:01:49,973 --> 00:01:51,075 Who falls inside that circle 53 00:01:51,847 --> 00:01:54,389 The first ring holds no surprises curved spacetime 54 00:01:54,726 --> 00:01:58,570 Field equations Riemannian geometry equivalence and gravitational waves 55 00:01:58,886 --> 00:01:59,669 They are all family 56 00:01:59,927 --> 00:02:01,490 Close relatives within physics 57 00:02:02,006 --> 00:02:03,650 The second ring gets interesting 58 00:02:03,966 --> 00:02:06,529 You find the concept of relation sitting close by 59 00:02:07,006 --> 00:02:08,365 Not social relationships 60 00:02:08,366 --> 00:02:12,086 But the deeper idea of relational ontology where the nature of things 61 00:02:12,087 --> 00:02:13,885 Does not lie in what they are alone 62 00:02:13,886 --> 00:02:15,650 But in their relations with one another 63 00:02:16,087 --> 00:02:20,230 This is the central insight of general relativity there is no absolute space 64 00:02:20,606 --> 00:02:21,969 And no absolute time 65 00:02:22,206 --> 00:02:26,843 Space and time arise from relations between matter and energy 66 00:02:27,080 --> 00:02:28,202 Keep moving outward 67 00:02:28,360 --> 00:02:32,223 And you meet a neighbor that makes physicists frown dependent origination 68 00:02:32,839 --> 00:02:33,183 Yes 69 00:02:33,519 --> 00:02:35,262 The Buddhist teaching of dependent origination and emptiness 70 00:02:35,920 --> 00:02:36,862 When this exists that arises 71 00:02:37,200 --> 00:02:38,523 When this ceases that ceases 72 00:02:38,799 --> 00:02:41,642 Everything arises through causes and conditions 73 00:02:41,839 --> 00:02:44,843 Nothing has an independent and inherent nature 74 00:02:45,119 --> 00:02:45,963 Think about it 75 00:02:45,999 --> 00:02:49,303 Are these two ways of saying the same thing 76 00:02:49,839 --> 00:02:53,283 Einstein says space is not an independent stage 77 00:02:53,439 --> 00:02:55,563 It is a function of matter and energy 78 00:02:55,880 --> 00:02:59,918 Nagarjuna says no being is an independently existing entity 79 00:02:59,920 --> 00:03:01,403 It is a function of relationships 80 00:03:01,937 --> 00:03:03,855 AI does not know who Einstein is 81 00:03:03,856 --> 00:03:05,460 And it does not know who Nagarjuna is 82 00:03:05,696 --> 00:03:08,420 It does not know one was a twentieth century physicist 83 00:03:08,456 --> 00:03:10,980 And the other a second century Buddhist philosopher 84 00:03:11,257 --> 00:03:13,700 It simply read everything humanity has written 85 00:03:13,777 --> 00:03:17,499 And found two sides of the same idea in their words 86 00:03:17,777 --> 00:03:20,140 So it placed them close together 87 00:03:20,456 --> 00:03:21,480 This is not a coincidence 88 00:03:21,777 --> 00:03:23,620 And it is not AI talking nonsense 89 00:03:24,401 --> 00:03:25,563 In 2024 90 00:03:25,640 --> 00:03:28,924 A team of MIT researchers published an important paper 91 00:03:28,960 --> 00:03:30,824 Called the Platonic Representation Hypothesis 92 00:03:31,360 --> 00:03:34,903 They found something surprising different AI models 93 00:03:35,200 --> 00:03:36,883 Trained on different data 94 00:03:36,960 --> 00:03:39,720 And working across different modalities some with text 95 00:03:39,721 --> 00:03:40,919 Some with images 96 00:03:40,920 --> 00:03:44,439 And some with sound ultimately learn high dimensional representations 97 00:03:44,441 --> 00:03:45,644 That are converging 98 00:03:45,920 --> 00:03:46,704 Did you catch that 99 00:03:47,121 --> 00:03:47,903 Different AI systems 100 00:03:48,161 --> 00:03:49,244 Take different paths 101 00:03:49,281 --> 00:03:52,119 Yet their maps of the world look increasingly alike 102 00:03:52,476 --> 00:03:53,100 More specifically 103 00:03:53,316 --> 00:03:54,200 The researchers found 104 00:03:54,276 --> 00:03:58,040 That as a language model becomes larger and more capable 105 00:03:58,237 --> 00:04:00,196 Its internal structure of conceptual distances 106 00:04:00,196 --> 00:04:02,675 Moves closer to the distance structure 107 00:04:02,677 --> 00:04:03,799 Inside a vision model 108 00:04:04,196 --> 00:04:07,600 An AI trained only on text places cats and dogs close together 109 00:04:07,756 --> 00:04:11,320 And an AI trained only on images does the same 110 00:04:11,556 --> 00:04:13,639 They never exchanged any information 111 00:04:13,716 --> 00:04:16,880 Yet their ways of measuring the world are converging 112 00:04:17,357 --> 00:04:21,540 Researchers call this convergent destination a Platonic representation 113 00:04:22,670 --> 00:04:23,733 Why call it Platonic 114 00:04:24,230 --> 00:04:26,430 Because twenty four centuries ago Plato 115 00:04:26,430 --> 00:04:29,514 Proposed the famous allegory of the cave 116 00:04:29,750 --> 00:04:32,954 He said humans are like prisoners bound inside a cave 117 00:04:33,190 --> 00:04:36,353 Seeing only shadows cast by fire onto a wall 118 00:04:36,550 --> 00:04:40,153 The true Forms or essence of things exist outside the cave 119 00:04:40,350 --> 00:04:41,973 We can only see projections 120 00:04:42,271 --> 00:04:43,795 Never reality itself 121 00:04:44,111 --> 00:04:48,110 The MIT paper says different AI models resemble 122 00:04:48,111 --> 00:04:50,995 Different prisoners standing at different places in the cave 123 00:04:51,231 --> 00:04:52,695 One sees one side of the shadow 124 00:04:53,031 --> 00:04:54,634 Another sees the other side 125 00:04:54,872 --> 00:04:56,711 But as they become more capable 126 00:04:56,711 --> 00:05:00,274 The shapes they infer outside the cave become more alike 127 00:05:00,752 --> 00:05:03,435 The researchers also provided a mathematical explanation 128 00:05:03,671 --> 00:05:04,495 They showed 129 00:05:04,791 --> 00:05:06,714 That under idealized conditions 130 00:05:06,791 --> 00:05:09,875 Different learners eventually converge on one representation 131 00:05:10,072 --> 00:05:13,855 Its core is pointwise mutual information between underlying events 132 00:05:14,952 --> 00:05:15,853 In plain language 133 00:05:15,863 --> 00:05:19,875 AI ultimately learns no single point of view 134 00:05:19,911 --> 00:05:22,070 It learns the relationships 135 00:05:22,072 --> 00:05:23,894 Among the underlying events 136 00:05:24,192 --> 00:05:26,034 That generate those perceptions 137 00:05:26,432 --> 00:05:27,295 It learns causes 138 00:05:27,791 --> 00:05:28,774 Rather than effects 139 00:05:30,472 --> 00:05:31,335 What does that mean 140 00:05:31,792 --> 00:05:34,835 It means the vector space of AI is not arbitrary 141 00:05:34,992 --> 00:05:37,196 It is approaching some real structure 142 00:05:37,472 --> 00:05:39,835 Different AI systems use different channels of perception 143 00:05:39,992 --> 00:05:42,636 Yet the world skeletons they draw are converging 144 00:05:42,833 --> 00:05:44,792 Like astronomers in different countries 145 00:05:44,792 --> 00:05:47,191 Using different telescopes and frequency bands 146 00:05:47,193 --> 00:05:49,636 To produce the same map of the universe 147 00:05:50,564 --> 00:05:53,006 Let us take this idea a step further 148 00:05:53,244 --> 00:05:56,846 When AI places relativity and dependent origination side by side 149 00:05:57,004 --> 00:05:58,406 It is not making an analogy 150 00:05:58,444 --> 00:05:59,927 Or playing a word game 151 00:06:00,124 --> 00:06:03,966 It is saying that across all experience encoded in human language 152 00:06:04,004 --> 00:06:07,446 These two concepts invoke the same deep semantic structure 153 00:06:07,603 --> 00:06:11,863 They share an ontology where relations come before entities 154 00:06:12,260 --> 00:06:12,883 You may say 155 00:06:13,101 --> 00:06:14,123 Why is that a big deal 156 00:06:14,700 --> 00:06:18,263 People noticed similarities between Buddhism and quantum physics long ago 157 00:06:18,402 --> 00:06:19,202 That is true 158 00:06:19,460 --> 00:06:21,483 But those comparisons were made by people 159 00:06:21,781 --> 00:06:23,740 A thinker would read physics 160 00:06:23,741 --> 00:06:24,604 And then read Buddhism 161 00:06:24,900 --> 00:06:26,539 And subjectively think well 162 00:06:26,541 --> 00:06:27,683 These two seem a little alike 163 00:06:28,661 --> 00:06:30,783 That comparison always faces one challenge 164 00:06:30,793 --> 00:06:33,284 Are you forcing a similarity 165 00:06:33,781 --> 00:06:36,443 Are you pretending different things are the same 166 00:06:37,110 --> 00:06:38,953 But AI is different 167 00:06:39,110 --> 00:06:42,314 It has no motive to connect physics with Buddhism 168 00:06:42,470 --> 00:06:44,393 It has no need for interdisciplinary dialogue 169 00:06:44,430 --> 00:06:45,793 No urge to publish a book 170 00:06:45,831 --> 00:06:49,314 And no vanity about uniting East and West 171 00:06:49,711 --> 00:06:53,393 It is doing one painfully dull task predicting the next word 172 00:06:53,790 --> 00:06:54,134 Then 173 00:06:54,391 --> 00:06:55,869 As a byproduct of prediction 174 00:06:55,870 --> 00:06:57,614 It spontaneously arranges concepts in parameter space 175 00:06:57,790 --> 00:07:00,230 Unconsciously and purely by statistical patterns 176 00:07:00,230 --> 00:07:02,634 Placing these concepts in the same neighborhood 177 00:07:03,071 --> 00:07:05,233 It is like finding something you never searched for 178 00:07:05,391 --> 00:07:07,429 You simply stand high enough 179 00:07:07,430 --> 00:07:08,874 Look down and realize 180 00:07:08,910 --> 00:07:11,533 These two rivers flow into the same sea 181 00:07:11,533 --> 00:07:14,196 Let me give you a few more examples 182 00:07:14,393 --> 00:07:17,996 Find the term observer effect in the vector space of AI 183 00:07:18,314 --> 00:07:21,992 It is a core idea in quantum mechanics where observation itself 184 00:07:21,994 --> 00:07:23,996 Changes the system being observed 185 00:07:24,354 --> 00:07:25,516 Look at its neighbors 186 00:07:25,674 --> 00:07:27,857 Beyond wave function collapse and Heisenberg 187 00:07:28,153 --> 00:07:30,797 And physical terms such as the measurement problem 188 00:07:30,833 --> 00:07:33,357 You also find mindfulness nearby 189 00:07:33,633 --> 00:07:34,696 That sounds crazy right 190 00:07:35,193 --> 00:07:35,576 Mindfulness 191 00:07:35,874 --> 00:07:37,232 A concept from meditation 192 00:07:37,234 --> 00:07:39,477 And the observer effect in quantum mechanics 193 00:07:39,513 --> 00:07:41,876 Are neighbors in the semantic space of AI 194 00:07:42,273 --> 00:07:43,477 But think about it carefully 195 00:07:43,513 --> 00:07:46,007 Their deep structures really are alike 196 00:07:46,243 --> 00:07:48,962 The observer effect says whatever you observe 197 00:07:48,964 --> 00:07:50,406 You also change 198 00:07:50,683 --> 00:07:51,923 In quantum mechanics 199 00:07:51,923 --> 00:07:55,166 An electron is in superposition before it is measured 200 00:07:55,324 --> 00:07:57,406 It is here and there at once 201 00:07:57,683 --> 00:07:58,426 The moment you observe it 202 00:07:58,683 --> 00:08:00,887 It collapses into a definite state 203 00:08:01,204 --> 00:08:04,406 Observation itself participates in creating reality 204 00:08:05,642 --> 00:08:06,586 What does mindfulness say 205 00:08:07,163 --> 00:08:08,185 Whatever you become aware of 206 00:08:08,403 --> 00:08:10,045 You begin to free yourself from 207 00:08:10,322 --> 00:08:12,325 When you are unaware of your anxiety 208 00:08:12,442 --> 00:08:13,521 The anxiety is you 209 00:08:13,522 --> 00:08:14,505 And you are the anxiety 210 00:08:14,843 --> 00:08:16,245 The two are fused together 211 00:08:16,522 --> 00:08:18,706 But once you notice I am anxious 212 00:08:19,122 --> 00:08:21,726 A space appears between you and the anxiety 213 00:08:22,002 --> 00:08:23,041 The anxiety remains 214 00:08:23,043 --> 00:08:25,646 But its control over you suddenly weakens 215 00:08:25,882 --> 00:08:28,445 Awareness changes what is being observed 216 00:08:28,762 --> 00:08:32,565 Both share one premise awareness is not passive recording 217 00:08:32,723 --> 00:08:35,086 Awareness is active participation 218 00:08:35,122 --> 00:08:37,325 It changes the object of awareness 219 00:08:37,682 --> 00:08:41,006 Quantum mechanics describes this with equations 220 00:08:41,162 --> 00:08:44,406 Buddhism presents it through introspective experience 221 00:08:45,243 --> 00:08:46,266 The forms are completely different 222 00:08:46,563 --> 00:08:48,126 But the skeleton is the same 223 00:08:48,483 --> 00:08:51,045 And that skeleton is what AI sees 224 00:08:51,408 --> 00:08:51,910 Here is another example 225 00:08:51,910 --> 00:08:53,690 The concept of emergence 226 00:08:53,967 --> 00:08:58,530 It comes from complexity science where simple rules create complex behavior 227 00:08:58,688 --> 00:09:01,851 Ants have no architect yet build intricate nests 228 00:09:02,087 --> 00:09:04,170 Look at its neighbors in vector space 229 00:09:04,247 --> 00:09:06,567 Alongside self organization and nonlinearity 230 00:09:06,568 --> 00:09:08,111 You find the cluster the Dao gives birth to one 231 00:09:08,408 --> 00:09:09,246 One gives birth to two 232 00:09:09,247 --> 00:09:09,871 Two gives birth to three 233 00:09:10,208 --> 00:09:13,010 And three gives birth to all things 234 00:09:13,408 --> 00:09:15,687 What Laozi described twenty five centuries ago 235 00:09:15,688 --> 00:09:17,927 And what complexity science studies today 236 00:09:17,928 --> 00:09:20,780 Are close neighbors in the eyes of AI 237 00:09:21,017 --> 00:09:21,896 But here 238 00:09:21,897 --> 00:09:24,859 I need to add a very honest and important warning 239 00:09:25,056 --> 00:09:27,300 We previously discussed Wittgenstein 240 00:09:27,336 --> 00:09:29,855 And whether AI can align any two concepts 241 00:09:29,857 --> 00:09:31,136 Because it is approaching truth 242 00:09:31,137 --> 00:09:32,959 Or creating the most refined illusion of language 243 00:09:33,456 --> 00:09:37,095 The main point was that AI can make any two concepts 244 00:09:37,097 --> 00:09:38,420 Appear connected 245 00:09:38,497 --> 00:09:40,735 It only needs to search vector space 246 00:09:40,737 --> 00:09:43,379 For associations near both concepts 247 00:09:43,497 --> 00:09:46,940 Then follow them to generate fluent prose 248 00:09:47,097 --> 00:09:50,440 Ask how recipes resemble quantum mechanics 249 00:09:50,816 --> 00:09:52,496 And it can write three thousand words 250 00:09:52,497 --> 00:09:54,379 That even sound convincing 251 00:09:55,017 --> 00:09:57,979 So I want to draw a very clear line here 252 00:09:58,815 --> 00:10:01,858 When AI places relativity near dependent origination 253 00:10:01,975 --> 00:10:04,445 It tells us a fact about language 254 00:10:04,455 --> 00:10:07,177 Across thousands of years of human writing 255 00:10:07,254 --> 00:10:09,317 The contexts in which these concepts appear 256 00:10:09,575 --> 00:10:12,653 The ways they are discussed and the ideas linked to them 257 00:10:12,654 --> 00:10:14,258 Overlap to a remarkable degree 258 00:10:14,455 --> 00:10:16,417 That is an objective statistical finding 259 00:10:16,654 --> 00:10:18,138 But does that finding 260 00:10:18,254 --> 00:10:20,413 Mean they describe 261 00:10:20,415 --> 00:10:21,797 The same ultimate reality 262 00:10:22,575 --> 00:10:23,317 That inference 263 00:10:23,695 --> 00:10:25,018 Is beyond what AI can do for you 264 00:10:25,134 --> 00:10:26,898 You must decide for yourself 265 00:10:27,215 --> 00:10:28,618 Semantic closeness 266 00:10:28,654 --> 00:10:30,697 Does not equal ontological identity 267 00:10:31,415 --> 00:10:34,293 When two concepts are close in the vector space of AI 268 00:10:34,295 --> 00:10:38,337 Humans use similar cognitive structures when discussing them 269 00:10:38,774 --> 00:10:42,297 But do they point to the same deep truth 270 00:10:42,455 --> 00:10:44,094 Or does the human mind 271 00:10:44,095 --> 00:10:46,533 Naturally favor the same patterns 272 00:10:46,534 --> 00:10:47,677 To explain different things 273 00:10:49,047 --> 00:10:49,511 To that question 274 00:10:49,968 --> 00:10:51,011 I have no answer 275 00:10:51,208 --> 00:10:52,611 But I do have an inclination 276 00:10:52,968 --> 00:10:56,566 When two completely independent paths of thought 277 00:10:56,568 --> 00:10:59,766 Separated by two thousand years and half the planet 278 00:10:59,767 --> 00:11:02,207 Arrive without borrowing from each other 279 00:11:02,208 --> 00:11:04,247 At the same structural conclusion 280 00:11:04,247 --> 00:11:05,111 It at least tells us 281 00:11:05,367 --> 00:11:07,290 That conclusion was not a random guess 282 00:11:07,328 --> 00:11:09,651 It may have touched something fundamental 283 00:11:10,007 --> 00:11:12,686 Imagine looking at the coasts of Africa and South America 284 00:11:12,688 --> 00:11:15,411 And finding outlines that fit perfectly 285 00:11:15,527 --> 00:11:18,526 You cannot be completely sure they were once joined 286 00:11:18,527 --> 00:11:19,967 But the weight of that evidence 287 00:11:19,968 --> 00:11:21,487 Is enough to take seriously 288 00:11:21,488 --> 00:11:22,970 The hypothesis of continental drift 289 00:11:23,647 --> 00:11:26,611 The vector space of AI is like a geological map of concepts 290 00:11:27,088 --> 00:11:29,710 It cannot prove those continents were once connected 291 00:11:30,167 --> 00:11:32,891 But it can reveal the hidden line where they fit 292 00:11:33,247 --> 00:11:35,007 And the Platonic Representation Hypothesis 293 00:11:35,007 --> 00:11:39,251 Adds confidence because different AI systems draw the same map 294 00:11:39,647 --> 00:11:41,011 That looks less like coincidence 295 00:11:41,047 --> 00:11:44,666 And more like convergence on a real structure 296 00:11:45,184 --> 00:11:47,147 That leads me to a deeper question 297 00:11:47,623 --> 00:11:48,666 For two thousand years 298 00:11:48,704 --> 00:11:53,507 Humans have divided knowledge into physics philosophy religion and psychology 299 00:11:53,824 --> 00:11:57,227 This classification system greatly advanced civilization 300 00:11:57,424 --> 00:11:59,706 It allowed every field to develop in depth 301 00:11:59,903 --> 00:12:01,566 But it also came at a huge cost 302 00:12:02,104 --> 00:12:04,662 It took what may be different aspects 303 00:12:04,663 --> 00:12:05,786 Of the same thing 304 00:12:05,984 --> 00:12:07,867 And locked them in separate drawers 305 00:12:08,224 --> 00:12:09,666 AI has no drawers 306 00:12:09,783 --> 00:12:12,546 It sees one continuous semantic landscape 307 00:12:12,783 --> 00:12:16,027 With no artificial walls between concepts 308 00:12:16,344 --> 00:12:20,466 It is the first entity able to see all human knowledge at once 309 00:12:20,704 --> 00:12:23,227 And it does not see only the words 310 00:12:23,344 --> 00:12:26,274 It sees the distance structure behind them 311 00:12:26,790 --> 00:12:28,674 It is like the invention of the microscope 312 00:12:28,831 --> 00:12:30,274 Which let humans see cells for the first time 313 00:12:30,511 --> 00:12:32,510 The microscope did not invent cells 314 00:12:32,511 --> 00:12:33,953 They were always there 315 00:12:34,111 --> 00:12:37,674 It simply revealed a structure invisible to the naked eye 316 00:12:37,991 --> 00:12:41,493 The vector space of AI may be a microscope for meaning 317 00:12:42,070 --> 00:12:43,230 It lets you see 318 00:12:43,231 --> 00:12:45,894 Hidden structural links between different fields 319 00:12:46,351 --> 00:12:48,113 Those links were always there 320 00:12:48,190 --> 00:12:51,034 But human classification systems concealed them 321 00:12:51,150 --> 00:12:51,534 Of course 322 00:12:51,790 --> 00:12:53,674 Microscopes can also create artifacts 323 00:12:53,871 --> 00:12:54,949 The texture you see 324 00:12:54,951 --> 00:12:56,833 May belong to the sample itself 325 00:12:56,910 --> 00:12:59,434 Or be optical distortion caused by the lens 326 00:13:00,111 --> 00:13:00,654 Likewise 327 00:13:01,150 --> 00:13:03,473 Neighborhoods in AI vector space 328 00:13:03,511 --> 00:13:05,713 May reflect real conceptual structure 329 00:13:05,871 --> 00:13:08,554 Or merely the biases of human language 330 00:13:08,951 --> 00:13:09,813 Telling them apart 331 00:13:10,111 --> 00:13:11,510 Is our responsibility as humans 332 00:13:11,511 --> 00:13:13,222 Not the job of AI 333 00:13:13,539 --> 00:13:17,461 One thing is certain AI offers a completely new perspective 334 00:13:17,738 --> 00:13:22,742 One free from disciplinary cultural and religious bias 335 00:13:22,898 --> 00:13:26,941 It does not care whether you are a Nobel laureate or an ancient monk 336 00:13:27,179 --> 00:13:30,022 It cares about one thing the words you use 337 00:13:30,059 --> 00:13:32,498 Where they sit in semantic space 338 00:13:32,498 --> 00:13:33,701 And how close they are 339 00:13:34,099 --> 00:13:35,338 So when it tells you 340 00:13:35,339 --> 00:13:37,701 That relativity and dependent origination are neighbors 341 00:13:37,898 --> 00:13:40,181 Whether you take that discovery seriously 342 00:13:40,219 --> 00:13:41,222 Is up to you 343 00:13:41,498 --> 00:13:45,540 My view is that it deserves at least a moment of thought 344 00:13:46,016 --> 00:13:47,815 Because if AI is right 345 00:13:47,817 --> 00:13:52,256 If these cross disciplinary distances reflect something real 346 00:13:52,256 --> 00:13:54,719 Then humans may have already 347 00:13:55,016 --> 00:13:58,496 Across distant cultures and completely different languages 348 00:13:58,496 --> 00:14:01,099 Touched different parts of the same elephant 349 00:14:01,457 --> 00:14:02,339 And AI 350 00:14:02,376 --> 00:14:05,740 Is the first entity able to touch every part at once 351 00:14:06,296 --> 00:14:08,140 It may not know what the elephant is 352 00:14:08,496 --> 00:14:10,299 But the outline it draws 353 00:14:10,337 --> 00:14:12,695 Is more complete than the one drawn 354 00:14:12,697 --> 00:14:14,005 By any single human discipline 355 00:14:14,403 --> 00:14:15,966 This leads to the final question 356 00:14:16,082 --> 00:14:17,966 The one I most want to leave with you today 357 00:14:18,322 --> 00:14:19,785 If the vector space of AI 358 00:14:20,163 --> 00:14:21,726 Is truly approaching 359 00:14:21,883 --> 00:14:23,226 Some Platonic structure of reality 360 00:14:23,763 --> 00:14:25,025 Then what is that structure 361 00:14:25,643 --> 00:14:26,706 Is it the structure of matter 362 00:14:27,482 --> 00:14:28,546 The structure of information 363 00:14:29,283 --> 00:14:30,385 Or the structure of consciousness 364 00:14:31,322 --> 00:14:32,562 A physicist might say 365 00:14:32,562 --> 00:14:34,605 It is the mathematical structure of physical law 366 00:14:35,082 --> 00:14:35,905 A Buddhist scholar might say 367 00:14:36,202 --> 00:14:37,966 It is the nature of dependent origination 368 00:14:38,362 --> 00:14:39,761 An information theorist might say 369 00:14:39,763 --> 00:14:41,704 It is a statistical invariant in data 370 00:14:41,980 --> 00:14:44,064 But perhaps they are all describing the same thing 371 00:14:44,620 --> 00:14:46,464 Perhaps that thing has no name 372 00:14:46,541 --> 00:14:50,143 Because every name is a projection from one angle 373 00:14:50,421 --> 00:14:51,984 And the vector space of AI 374 00:14:52,021 --> 00:14:56,143 Gives us our first panoramic view without a fixed angle 375 00:14:56,661 --> 00:14:57,003 Of course 376 00:14:57,261 --> 00:14:59,220 That panoramic view is blurry 377 00:14:59,220 --> 00:15:00,003 Approximate 378 00:15:00,181 --> 00:15:01,303 And noisy 379 00:15:01,580 --> 00:15:03,340 But for now it is 380 00:15:03,340 --> 00:15:07,379 Our best mirror reflecting what lies behind all human knowledge 381 00:15:07,380 --> 00:15:10,431 That faintly visible common structure 382 00:15:11,013 --> 00:15:12,457 What do you think that structure is 383 00:15:12,885 --> 00:15:15,002 Do you think it is real or an illusion 384 00:15:15,542 --> 00:15:16,537 If you have an idea 385 00:15:16,852 --> 00:15:17,979 I will meet you in the comments 386 00:15:18,125 --> 00:15:18,782 I am Leo Wang 387 00:15:19,024 --> 00:15:19,749 See you next time