1 00:00:00,000 --> 00:00:03,040 Welcome back to the Deep Dive. Today, we are 2 00:00:03,040 --> 00:00:05,799 clearing the table. We are taking all the headlines, 3 00:00:06,139 --> 00:00:09,539 all the hype, all the panic about robots taking 4 00:00:09,539 --> 00:00:11,800 over the world, and we are putting them to the 5 00:00:11,800 --> 00:00:13,779 side for a moment. For just a moment. We are 6 00:00:13,779 --> 00:00:16,719 going to tackle one of the absolute giants of 7 00:00:16,719 --> 00:00:18,879 modern technology, but we're coming at it from 8 00:00:18,879 --> 00:00:20,780 a slightly different angle than you might expect. 9 00:00:20,859 --> 00:00:24,039 We are talking about artificial intelligence. 10 00:00:24,219 --> 00:00:26,640 Indeed. But we aren't just going to marvel at 11 00:00:26,640 --> 00:00:29,519 the latest chat bot or the newest image generator. 12 00:00:29,920 --> 00:00:32,219 We are going to look under the hood. We're looking 13 00:00:32,219 --> 00:00:35,500 at a fundamental critique of the current direction 14 00:00:35,500 --> 00:00:38,420 of AI and a roadmap for where it actually needs 15 00:00:38,420 --> 00:00:41,600 to go if we want it to be, well, Intelligent. 16 00:00:41,820 --> 00:00:44,799 Exactly. We're basing this deep dive on a fascinating 17 00:00:44,799 --> 00:00:47,039 heavyweight paper titled, Building Machines That 18 00:00:47,039 --> 00:00:48,920 Learn and Think Like People. It's by Brendan 19 00:00:48,920 --> 00:00:51,740 Lake, Tomer Ullman, Joshua Tennenbaum, and Samuel 20 00:00:51,740 --> 00:00:54,820 Gershman. These are heavy hitters. MIT, NYU, 21 00:00:55,079 --> 00:00:57,159 Harvard. Right. And the title really says it 22 00:00:57,159 --> 00:00:59,500 all. We have machines that can do amazing things, 23 00:00:59,679 --> 00:01:01,740 right? They can beat us at go. They can recognize 24 00:01:01,740 --> 00:01:04,180 a cat in a photo. They can transcribe our speech 25 00:01:04,180 --> 00:01:07,340 with uncanny accuracy. Correct. We are currently 26 00:01:07,340 --> 00:01:10,329 living through what you might call a boom cycle 27 00:01:10,329 --> 00:01:13,670 in AI history, specifically the boom of deep 28 00:01:13,670 --> 00:01:15,730 learning and neural networks. It's the engine 29 00:01:15,730 --> 00:01:17,829 behind almost everything you see in the news. 30 00:01:18,629 --> 00:01:20,909 But, and here's the hook, the thing that stops 31 00:01:20,909 --> 00:01:24,530 you in your tracks. This paper argues that despite 32 00:01:24,530 --> 00:01:27,609 all those wins, despite the crushed high scores 33 00:01:27,609 --> 00:01:30,709 and the viral videos, there is a massive fundamental 34 00:01:30,709 --> 00:01:33,640 piece missing. A huge piece. The central premise 35 00:01:33,640 --> 00:01:36,540 here is that while biological inspiration, you 36 00:01:36,540 --> 00:01:39,400 know, neurons, brains got us this far, current 37 00:01:39,400 --> 00:01:42,000 AI is actually nothing like human intelligence. 38 00:01:42,280 --> 00:01:44,439 It's a bold claim, but the evidence they present 39 00:01:44,439 --> 00:01:46,840 is compelling. The authors argue that if we really 40 00:01:46,840 --> 00:01:49,180 want machines that think we can't just keep building 41 00:01:49,180 --> 00:01:51,920 bigger chips and feeding them more data, we need 42 00:01:51,920 --> 00:01:54,579 to reverse engineer the human mind. Reverse engineer 43 00:01:54,579 --> 00:01:56,560 the mind. We need to look at the software of 44 00:01:56,560 --> 00:01:58,780 cognition. not just the hardware of neurons. 45 00:01:59,079 --> 00:02:00,980 I love that distinction. And just to set the 46 00:02:00,980 --> 00:02:02,760 stage for everyone listening, the authors break 47 00:02:02,760 --> 00:02:05,180 this down into a battle between two heavyweights, 48 00:02:05,560 --> 00:02:08,300 pattern recognition versus model building. Pattern 49 00:02:08,300 --> 00:02:10,599 recognition is what our current AI does, right? 50 00:02:11,000 --> 00:02:13,539 Precisely. Pattern recognition is mapping inputs 51 00:02:13,539 --> 00:02:16,000 to outputs. You show the machine a million pixels 52 00:02:16,000 --> 00:02:18,620 and it says cat. Wow. It's entirely focused on 53 00:02:18,620 --> 00:02:21,219 prediction. OK. But the authors argue that human 54 00:02:21,219 --> 00:02:23,620 intelligence is about model building. It's not 55 00:02:23,620 --> 00:02:26,099 just predicting, it's explaining. It's constructing 56 00:02:26,099 --> 00:02:29,240 a causal model of how the world actually works. 57 00:02:29,659 --> 00:02:31,860 So our mission for this deep dive is to figure 58 00:02:31,860 --> 00:02:34,259 out what that model building actually looks like. 59 00:02:34,560 --> 00:02:36,500 We're going to look at the startup software we 60 00:02:36,500 --> 00:02:39,120 are all born with, how we learn so much from 61 00:02:39,120 --> 00:02:43,340 so little data, and how to bridge the gap between 62 00:02:43,340 --> 00:02:46,740 fast reactions and deep thinking. It is the roadmap 63 00:02:46,740 --> 00:02:49,639 to the next generation of AI. We are moving from 64 00:02:49,639 --> 00:02:52,610 the how do we make it faster? phase to the, how 65 00:02:52,610 --> 00:02:54,849 do we make it understand phase? OK, let's unpack 66 00:02:54,849 --> 00:02:56,710 the landscape first, because we have to give 67 00:02:56,710 --> 00:02:58,770 credit where credit is due. Deep learning has 68 00:02:58,770 --> 00:03:00,469 been a juggernaut. It's not like these scientists 69 00:03:00,469 --> 00:03:02,889 are just saying AI is useless. Absolutely not. 70 00:03:02,930 --> 00:03:05,849 I mean, if you look at the last decade, the progress 71 00:03:05,849 --> 00:03:08,050 is undeniable. We're talking about convenience, 72 00:03:08,229 --> 00:03:10,750 convolutional neural networks. In tasks like 73 00:03:10,750 --> 00:03:13,909 object recognition, they managed to have the 74 00:03:13,909 --> 00:03:16,449 error rates of previous technologies almost overnight. 75 00:03:16,610 --> 00:03:19,509 Have them? Yeah, that was around 2012. And it 76 00:03:19,509 --> 00:03:21,389 changed everything is the ImageNet moment. And 77 00:03:21,389 --> 00:03:24,650 we've seen DeepMind's DQN, that's the DeepQ network, 78 00:03:24,789 --> 00:03:27,710 playing Atari games at superhuman levels. I remember 79 00:03:27,710 --> 00:03:29,650 watching those clips of the computer playing 80 00:03:29,650 --> 00:03:32,009 Breakout. Oh, yeah. It figured out that if it 81 00:03:32,009 --> 00:03:34,050 tunneled the ball up the side of the wall, the 82 00:03:34,050 --> 00:03:36,030 ball would bounce around the top and destroy 83 00:03:36,030 --> 00:03:38,289 all the bricks without the paddle having to move. 84 00:03:38,430 --> 00:03:41,849 It was genius, in a way. It is impressive. Yeah. 85 00:03:41,889 --> 00:03:44,370 But we need to understand how it does that. At 86 00:03:44,370 --> 00:03:46,810 a very high level, these deep neural networks 87 00:03:46,810 --> 00:03:49,669 are layers of representations. They use an algorithm 88 00:03:49,669 --> 00:03:52,150 called backpropagation. OK, let's slow down and 89 00:03:52,150 --> 00:03:54,110 explain backpropagation, because it's the engine 90 00:03:54,110 --> 00:03:56,490 of the whole industry. How does it actually work? 91 00:03:56,569 --> 00:03:58,629 Can you break that down for us? For sure. Think 92 00:03:58,629 --> 00:04:01,009 of it as a correction loop. The system makes 93 00:04:01,009 --> 00:04:03,509 a guess. It sees an image and says, that's a 94 00:04:03,509 --> 00:04:06,469 toaster. The answer key says, no, that's a cat. 95 00:04:07,449 --> 00:04:10,150 The system then calculates the error mathematically, 96 00:04:10,270 --> 00:04:13,050 how wrong was I? Yeah. Then it propagates that 97 00:04:13,050 --> 00:04:15,629 error backward. through the layers of the network. 98 00:04:15,770 --> 00:04:18,290 Ah, so that's the back part of back propagation. 99 00:04:18,389 --> 00:04:20,569 That's the one. It adjusts the connections, the 100 00:04:20,569 --> 00:04:22,930 weights between the artificial neurons. It essentially 101 00:04:22,930 --> 00:04:25,810 says, hey, next time you see this pattern, don't 102 00:04:25,810 --> 00:04:28,310 fire the toaster neuron so hard. Fire the cat 103 00:04:28,310 --> 00:04:30,810 neuron. And it does this just once. It does this 104 00:04:30,810 --> 00:04:34,709 millions of times, billions even, with millions 105 00:04:34,709 --> 00:04:36,990 of pictures of cats and toasters and everything 106 00:04:36,990 --> 00:04:41,189 else. And that is the catch. These systems are 107 00:04:41,189 --> 00:04:43,750 incredibly data -hungry. Right. To teach a neural 108 00:04:43,750 --> 00:04:46,490 network, to recognize a handwritten letter A, 109 00:04:47,209 --> 00:04:48,910 you might need to show it thousands of different 110 00:04:48,910 --> 00:04:51,050 examples of an A. Whereas if I show my three 111 00:04:51,050 --> 00:04:53,930 -year -old an A once, maybe twice, she's got 112 00:04:53,930 --> 00:04:55,829 it, she can recognize it, she can draw it, she 113 00:04:55,829 --> 00:04:58,149 can probably even spot it in a weird font or 114 00:04:58,149 --> 00:05:00,250 carved into a tree. Exactly. And that brings 115 00:05:00,250 --> 00:05:02,709 us to the core distinction the paper makes. Humans 116 00:05:02,709 --> 00:05:05,970 don't just predict jada, we explain it. We explain 117 00:05:05,970 --> 00:05:09,300 it. The paper uses a beautiful analogy. Think 118 00:05:09,300 --> 00:05:12,189 of a scientist. A scientist doesn't just want 119 00:05:12,189 --> 00:05:14,050 to predict where a planet will be next Tuesday. 120 00:05:14,370 --> 00:05:17,329 They want to understand the laws of gravity that 121 00:05:17,329 --> 00:05:19,709 cause the planet to be there. So human thought 122 00:05:19,709 --> 00:05:22,410 is a model -building activity. We are constantly 123 00:05:22,410 --> 00:05:24,810 creating little theories about the world. If 124 00:05:24,810 --> 00:05:26,670 I know the theory of gravity, I can predict the 125 00:05:26,670 --> 00:05:28,709 planet, but I can also predict what happens if 126 00:05:28,709 --> 00:05:31,569 a meteor hits the planet or if the Sun disappears. 127 00:05:31,790 --> 00:05:34,939 Yes, you have a flexible model. Deep learning, 128 00:05:35,100 --> 00:05:37,540 largely, is just fitting a curve to data points. 129 00:05:37,699 --> 00:05:40,600 It doesn't have the underlying theory. And this 130 00:05:40,600 --> 00:05:43,019 brings us to a historical debate that the paper 131 00:05:43,019 --> 00:05:45,639 touches on, which is the blank slate myth. Oh, 132 00:05:45,660 --> 00:05:48,360 I know this one, the John Locke idea. It's the 133 00:05:48,360 --> 00:05:51,040 idea that we are born knowing absolutely nothing 134 00:05:51,040 --> 00:05:53,800 and experience writes everything onto our minds. 135 00:05:54,500 --> 00:05:56,699 Alan Turing even compared the child's mind to 136 00:05:56,699 --> 00:05:59,439 a notebook with lots of blank sheets. It is a 137 00:05:59,439 --> 00:06:02,579 romantic idea. And it aligns well with how current 138 00:06:02,579 --> 00:06:04,720 neural networks work. They start with random 139 00:06:04,720 --> 00:06:07,120 weights, knowing nothing, and learn everything 140 00:06:07,120 --> 00:06:09,519 from the data. But cognitive science tells us 141 00:06:09,519 --> 00:06:13,129 this is simply not true for humans. If we were 142 00:06:13,129 --> 00:06:15,990 truly blank slates, we would never learn as fast 143 00:06:15,990 --> 00:06:19,129 as we do. The math just doesn't work. There's 144 00:06:19,129 --> 00:06:21,589 too much to learn. Way too much. The authors 145 00:06:21,589 --> 00:06:24,990 argue we possess inductive biases, or what they 146 00:06:24,990 --> 00:06:27,610 call startup software. Startup software? I like 147 00:06:27,610 --> 00:06:30,730 that. It implies we come preloaded with an operating 148 00:06:30,730 --> 00:06:33,290 system that helps us interpret the world immediately. 149 00:06:33,990 --> 00:06:36,149 We aren't just hardware waiting for data. We 150 00:06:36,149 --> 00:06:38,550 have a bootloader. Correct. And before we get 151 00:06:38,550 --> 00:06:40,790 into the specifics of that software, the paper 152 00:06:40,790 --> 00:06:43,410 presents two challenges to illustrate just how 153 00:06:43,410 --> 00:06:45,870 big the gap is between machines and humans. They 154 00:06:45,870 --> 00:06:48,230 call them the Characters Challenge and the Frostbite 155 00:06:48,230 --> 00:06:50,129 Challenge. Let's start with the Characters Challenge. 156 00:06:50,209 --> 00:06:52,639 This is about handwriting, right? It is. The 157 00:06:52,639 --> 00:06:54,980 standard benchmark in AI for this is the MNIST 158 00:06:54,980 --> 00:06:57,540 of data set. It's a massive collection of handwritten 159 00:06:57,540 --> 00:06:59,959 digits, zero through nine. I've seen this one. 160 00:06:59,959 --> 00:07:01,980 It's like the hello world of machine learning. 161 00:07:02,220 --> 00:07:05,379 It basically is. And current deep learning algorithms 162 00:07:05,379 --> 00:07:09,120 can hit an error rate of 0 .2 % on this. They're 163 00:07:09,120 --> 00:07:10,800 practically perfect at looking at a scribble 164 00:07:10,800 --> 00:07:15,009 and saying, that's a seven. So game over. Machines 165 00:07:15,009 --> 00:07:18,769 win. Well, if the game is just classifying an 166 00:07:18,769 --> 00:07:22,089 image, then yes, it's a tie. But the authors 167 00:07:22,089 --> 00:07:24,930 argue that simply correctly labeling an image 168 00:07:24,930 --> 00:07:26,970 isn't the same as understanding the concept. 169 00:07:27,209 --> 00:07:29,629 Okay, so what's the difference? How do we prove 170 00:07:29,629 --> 00:07:31,750 the machine doesn't understand? The difference 171 00:07:31,750 --> 00:07:34,230 is one -shot learning and rich representation. 172 00:07:34,550 --> 00:07:36,889 As you mentioned with your daughter, humans can 173 00:07:36,889 --> 00:07:39,170 learn from a single example. Right. If I show 174 00:07:39,170 --> 00:07:40,750 you a character from the Tibetan alphabet, which 175 00:07:40,750 --> 00:07:42,910 you've likely never seen, you can immediately 176 00:07:42,910 --> 00:07:45,449 recognize it again. But more than that, you can 177 00:07:45,449 --> 00:07:47,529 draw it. Right. Because when I look at a handwritten 178 00:07:47,529 --> 00:07:50,069 character, I'm not just seeing a grid of black 179 00:07:50,069 --> 00:07:53,629 and white pixels. I'm seeing strokes. I'm seeing 180 00:07:53,629 --> 00:07:56,689 the motion. I can see where the pen started and 181 00:07:56,689 --> 00:07:59,329 where it lifted. I'm unconsciously reverse engineering 182 00:07:59,329 --> 00:08:02,089 the hand that wrote it. Exactly. You are seeing 183 00:08:02,089 --> 00:08:05,470 a program. You interpret the image as a series 184 00:08:05,470 --> 00:08:08,509 of causal actions. A neural network sees a texture. 185 00:08:08,990 --> 00:08:11,089 It sees a correlation of pixels. That's a great 186 00:08:11,089 --> 00:08:13,790 way to put it. A program versus a texture. And 187 00:08:13,790 --> 00:08:16,089 this is why humans can pass what the authors 188 00:08:16,089 --> 00:08:19,329 call a visual Turing test. If I showed you a 189 00:08:19,329 --> 00:08:21,350 Japanese character and asked you to invent a 190 00:08:21,350 --> 00:08:23,629 new character that looks like it belongs in the 191 00:08:23,629 --> 00:08:25,709 same alphabet, you could do it. You understand 192 00:08:25,709 --> 00:08:27,850 the style. And a neural network can't do that. 193 00:08:28,069 --> 00:08:30,709 Not easily. Not from one example. It can classify. 194 00:08:30,910 --> 00:08:34,429 But asking it to generate a variation based on 195 00:08:34,429 --> 00:08:36,710 one example is incredibly difficult for a system 196 00:08:36,710 --> 00:08:39,350 that relies on pattern matching thousands of 197 00:08:39,350 --> 00:08:41,870 examples. What would it produce? It creates something 198 00:08:41,870 --> 00:08:44,330 that looks like a blurry average, not a crisp 199 00:08:44,330 --> 00:08:47,610 new logical symbol. It fails to capture the concept 200 00:08:47,610 --> 00:08:49,350 behind the character. That makes sense. It's 201 00:08:49,350 --> 00:08:51,090 the difference between memorizing a painting 202 00:08:51,090 --> 00:08:53,629 and knowing how to paint. Okay, let's move to 203 00:08:53,629 --> 00:08:56,230 the second challenge, the frostbite challenge. 204 00:08:56,409 --> 00:08:59,149 This refers to the Atari game. Yes, Frostbite. 205 00:08:59,289 --> 00:09:01,730 It's a classic from the 80s. You control an agent, 206 00:09:02,029 --> 00:09:04,090 this little guy named Frostbite Bailey, jumping 207 00:09:04,090 --> 00:09:07,009 on ice floes to build an igloo. While also avoiding 208 00:09:07,009 --> 00:09:09,129 polar bears and managing your body temperature. 209 00:09:09,549 --> 00:09:12,669 Sounds stressful. It is. Now, the Deep Q network, 210 00:09:12,909 --> 00:09:14,809 Deep QN we mentioned earlier, the one that crushed 211 00:09:14,809 --> 00:09:16,830 breakout struggles significantly with Frostbite. 212 00:09:17,070 --> 00:09:20,210 Why? Is it just harder? More things on the screen? 213 00:09:20,269 --> 00:09:22,970 It's a different kind of hard. Breakout is very 214 00:09:22,970 --> 00:09:25,389 reactive. Ball comes down, move paddle. It's 215 00:09:25,389 --> 00:09:27,700 a very tight loop. Simple cause and effect. Right. 216 00:09:28,120 --> 00:09:30,179 Frostbite requires temporally extended planning. 217 00:09:30,580 --> 00:09:33,080 You have to jump on flows to get ice, but only 218 00:09:33,080 --> 00:09:35,399 the active white flows count. And then you have 219 00:09:35,399 --> 00:09:37,700 to get back to shore to build the igloo, block 220 00:09:37,700 --> 00:09:39,820 by block. So there are steps. You have to have 221 00:09:39,820 --> 00:09:42,559 a goal in mind that isn't just survive the next 222 00:09:42,559 --> 00:09:45,340 second. Exactly. And to reach human level performance 223 00:09:45,340 --> 00:09:48,620 on Atari games, the DQN often needs thousands 224 00:09:48,620 --> 00:09:52,149 of hours of gameplay. A human can pick up frostbite 225 00:09:52,149 --> 00:09:54,830 and understand the basic goal in minutes. Thousands 226 00:09:54,830 --> 00:09:58,649 of hours. That's years of a full -time job just 227 00:09:58,649 --> 00:10:00,830 to learn to jump on ice floes. It is immensely 228 00:10:00,830 --> 00:10:02,809 inefficient. And the reason for the difference 229 00:10:02,809 --> 00:10:06,110 is, again, representation. The neural network 230 00:10:06,110 --> 00:10:09,710 sees a screen of pixels changing states. It learns 231 00:10:09,710 --> 00:10:12,429 that when this particular pattern of pixels appears, 232 00:10:12,990 --> 00:10:15,230 move joystick left. So it doesn't know what an 233 00:10:15,230 --> 00:10:17,129 ice floe is. Not at all. It doesn't know what 234 00:10:17,129 --> 00:10:18,929 a bear is. It doesn't know it's building an igloo. 235 00:10:19,049 --> 00:10:21,710 It doesn't know that the bear is an enemy or 236 00:10:21,710 --> 00:10:24,509 that the ice is a solid platform that will support 237 00:10:24,509 --> 00:10:27,190 it. No, it just learns statistical correlations 238 00:10:27,190 --> 00:10:29,830 between pixel patterns and high scores. This 239 00:10:29,830 --> 00:10:32,009 is the critical part. If you change the color 240 00:10:32,009 --> 00:10:35,330 of the ice flows from white to, say, green or 241 00:10:35,330 --> 00:10:37,549 move the score counter to the bottom of the screen, 242 00:10:37,710 --> 00:10:40,049 the whole thing breaks. The neural networks might 243 00:10:40,049 --> 00:10:42,169 completely fail because the pattern it memorized 244 00:10:42,169 --> 00:10:45,029 has changed. A human wouldn't blink. We'd say, 245 00:10:45,049 --> 00:10:47,840 oh, the ice is green now. Fine. The physics of 246 00:10:47,840 --> 00:10:49,860 jumping on a platform is still the same. That's 247 00:10:49,860 --> 00:10:52,799 a huge distinction. We see objects and agents. 248 00:10:53,179 --> 00:10:55,840 The machine sees data streams. It's a fragility 249 00:10:55,840 --> 00:10:58,539 problem. The machine is brilliant within the 250 00:10:58,539 --> 00:11:00,820 exact parameters it was trained on. But if you 251 00:11:00,820 --> 00:11:03,259 shift the world slightly, it breaks. And that 252 00:11:03,259 --> 00:11:05,740 leads us directly into section three. What is 253 00:11:05,740 --> 00:11:08,480 this startup software that allows us to see objects 254 00:11:08,480 --> 00:11:11,100 and agents so effortlessly? This is the stuff 255 00:11:11,100 --> 00:11:13,620 we're born with, or at least develop extremely 256 00:11:13,620 --> 00:11:16,549 early. The paper breaks it down into two main 257 00:11:16,549 --> 00:11:19,210 ingredients, intuitive physics and intuitive 258 00:11:19,210 --> 00:11:22,289 psychology. Let's look at ingredient A, intuitive 259 00:11:22,289 --> 00:11:26,129 physics. Even infants have primitive object concepts. 260 00:11:26,730 --> 00:11:28,570 Developmental psychologists have shown this for 261 00:11:28,570 --> 00:11:31,809 decades. Right, like solidity. If I drop my keys, 262 00:11:31,870 --> 00:11:34,389 they don't pass through the floor. They hit the 263 00:11:34,389 --> 00:11:38,259 floor and stop. A baby is surprised when a magician 264 00:11:38,259 --> 00:11:41,389 seems to violate that rule. Exactly. Solidity, 265 00:11:41,750 --> 00:11:44,090 persistence, knowing an object still exists when 266 00:11:44,090 --> 00:11:47,110 it goes behind a couch in continuity. Objects 267 00:11:47,110 --> 00:11:49,470 move on connected paths. They don't just teleport 268 00:11:49,470 --> 00:11:51,190 from one place to another. And we don't have 269 00:11:51,190 --> 00:11:53,149 to learn this every time we play a new video 270 00:11:53,149 --> 00:11:55,490 game. It's just baked in. It's baked in. When 271 00:11:55,490 --> 00:11:58,190 you play Frostbite, you assume the ice flow is 272 00:11:58,190 --> 00:12:00,429 a solid thing that will hold you up. You assume 273 00:12:00,429 --> 00:12:02,330 if the bear goes off the left side of the screen, 274 00:12:02,330 --> 00:12:03,870 it might come back or it's at least somewhere 275 00:12:03,870 --> 00:12:06,470 off screen. It didn't cease to exist. But the 276 00:12:06,470 --> 00:12:08,590 AI has to learn that from scratch. Precisely. 277 00:12:08,700 --> 00:12:11,600 We transfer this vast library of physical intuition 278 00:12:11,600 --> 00:12:14,299 to every new task. The neural network has to 279 00:12:14,299 --> 00:12:17,159 learn the concept of solid platform from scratch, 280 00:12:17,580 --> 00:12:20,200 pixel by pixel, every single time. It is trying 281 00:12:20,200 --> 00:12:22,899 to derive Newton's laws of motion from a video 282 00:12:22,899 --> 00:12:25,259 game score, which is an absurdly hard way to 283 00:12:25,259 --> 00:12:27,299 learn physics. OK, that makes perfect sense. 284 00:12:27,659 --> 00:12:30,500 Then we have ingredient B, intuitive psychology. 285 00:12:30,860 --> 00:12:33,820 This sounds a bit more complex. It's the understanding 286 00:12:33,820 --> 00:12:36,990 of agency. Infants understand that some objects 287 00:12:36,990 --> 00:12:39,809 are agents. They have mental states, goals, and 288 00:12:39,809 --> 00:12:42,549 beliefs. They act efficiently to achieve those 289 00:12:42,549 --> 00:12:45,190 goals. So a rock just falls, but a person walks. 290 00:12:45,350 --> 00:12:47,850 There's intent there. There's intent. And we 291 00:12:47,850 --> 00:12:50,309 are wired to look for it. The Pac -Man example 292 00:12:50,309 --> 00:12:52,909 in the outline is great for this. If you see 293 00:12:52,909 --> 00:12:55,009 Pac -Man running away from a ghost, you don't 294 00:12:55,009 --> 00:12:57,210 just see two -colored blobs moving on a screen. 295 00:12:57,309 --> 00:12:59,789 No, I infer a story. The ghost is bad. Pac -Man 296 00:12:59,789 --> 00:13:02,789 doesn't want to die. I immediately get the relationship. 297 00:13:03,009 --> 00:13:05,940 And you infer that efficiently. You assume the 298 00:13:05,940 --> 00:13:08,019 agent is rational. This is called the efficiency 299 00:13:08,019 --> 00:13:10,360 principle. If the agent wants the food, it will 300 00:13:10,360 --> 00:13:12,659 go directly to the food. If it runs away, there 301 00:13:12,659 --> 00:13:14,960 must be a reason. So this is a huge learning 302 00:13:14,960 --> 00:13:18,200 shortcut. A massive one. It allows us to learn 303 00:13:18,200 --> 00:13:21,279 from observation incredibly fast. If you watch 304 00:13:21,279 --> 00:13:24,820 a pro play frostbite, you infer their goal. Oh, 305 00:13:24,820 --> 00:13:26,899 they're trying to jump on the white ice. You 306 00:13:26,899 --> 00:13:29,220 don't need to try and fail a thousand times to 307 00:13:29,220 --> 00:13:31,519 figure out that the white ice is good. You infer 308 00:13:31,519 --> 00:13:34,460 it through intuitive psychology. So we're basically 309 00:13:34,460 --> 00:13:36,740 mind reading the player to understand the game. 310 00:13:36,919 --> 00:13:39,299 We are building a model of the player's intent. 311 00:13:39,519 --> 00:13:42,120 In a sense, yes. We are building a causal model 312 00:13:42,120 --> 00:13:44,980 of the player's mind. They want points, therefore 313 00:13:44,980 --> 00:13:47,659 jumping on white O -seals points. The machine 314 00:13:47,659 --> 00:13:50,620 just sees pixels changing color. It misses the 315 00:13:50,620 --> 00:13:53,360 Y. It has no model of other minds. Okay, so we 316 00:13:53,360 --> 00:13:55,820 have this startup software physics and psychology. 317 00:13:55,980 --> 00:13:58,659 It's the foundation. But how do we actually learn 318 00:13:58,659 --> 00:14:01,220 new things using it? That brings us to section 319 00:14:01,220 --> 00:14:03,600 four, learning as rapid model building. Right. 320 00:14:03,740 --> 00:14:06,039 And here the paper proposes a framework called 321 00:14:06,039 --> 00:14:09,679 Bayesian Program Learning, or BPL. Bayesian Program 322 00:14:09,679 --> 00:14:11,639 Learning. That sounds intense. Let's break that 323 00:14:11,639 --> 00:14:14,860 down. It's not as scary as it sounds. The core 324 00:14:14,860 --> 00:14:18,399 idea is the program analogy. Instead of thinking 325 00:14:18,399 --> 00:14:20,639 of the mind as a network of weights being adjusted, 326 00:14:20,940 --> 00:14:23,019 like a giant mixing board in a music studio, 327 00:14:23,600 --> 00:14:25,879 think of the mind as a software developer writing 328 00:14:25,879 --> 00:14:28,429 code. So when I learn a new concept, I'm writing 329 00:14:28,429 --> 00:14:30,970 a little computer program in my head. Metaphorically, 330 00:14:31,149 --> 00:14:34,710 yes. To understand the letter A, your mind writes 331 00:14:34,710 --> 00:14:38,210 a script. Draw an angled line left, draw an angled 332 00:14:38,210 --> 00:14:40,929 line right, cross in the middle. That is a generative 333 00:14:40,929 --> 00:14:43,330 model. It's a set of instructions that can produce 334 00:14:43,330 --> 00:14:45,450 the data. It's not a picture of an A, it's the 335 00:14:45,450 --> 00:14:48,029 recipe for an A. Perfect analogy. And because 336 00:14:48,029 --> 00:14:50,470 it's a program, it's flexible. You can run that 337 00:14:50,470 --> 00:14:53,870 program with a thick pen, a thin pen, on a napkin, 338 00:14:54,009 --> 00:14:56,200 on a whiteboard. the underlying code is the same. 339 00:14:56,480 --> 00:14:58,940 OK, I get it. And the Bayesian part, what's that 340 00:14:58,940 --> 00:15:00,879 about? The Bayesian part just means it deals 341 00:15:00,879 --> 00:15:03,460 with probability. The model is uncertain. Maybe 342 00:15:03,460 --> 00:15:05,759 the line is a little wobbly. Maybe the crossbar 343 00:15:05,759 --> 00:15:09,039 is high or low. The model represents all these 344 00:15:09,039 --> 00:15:11,759 possibilities and which are more likely. I see. 345 00:15:11,759 --> 00:15:14,799 So it's a probabilistic program. Exactly. And 346 00:15:14,799 --> 00:15:17,200 this brings in three sub -ingredients that make 347 00:15:17,200 --> 00:15:20,759 this mental programming so powerful. The first 348 00:15:20,759 --> 00:15:23,129 is compositionality. This is one of my favorite 349 00:15:23,129 --> 00:15:24,929 concepts. It's basically the LEGO principle, 350 00:15:25,090 --> 00:15:27,809 right? It is. Compositionality is the ability 351 00:15:27,809 --> 00:15:30,870 to build complex ideas out of simpler, reusable 352 00:15:30,870 --> 00:15:34,450 parts. Take the visual example of vehicles. Right. 353 00:15:34,730 --> 00:15:36,950 I know what a wheel is. I know what a handlebar 354 00:15:36,950 --> 00:15:39,970 is. I know what a chassis is. If I see a motorcycle, 355 00:15:40,090 --> 00:15:42,309 I don't memorize the whole image as a unique 356 00:15:42,309 --> 00:15:46,419 blob. I see Wheels plus chassis plus handlebars 357 00:15:46,419 --> 00:15:49,360 plus engine. And the power comes from combinatorics 358 00:15:49,360 --> 00:15:52,039 rearranging those parts. If I show you a segue, 359 00:15:52,639 --> 00:15:55,740 you've maybe never seen one before, but you recognize 360 00:15:55,740 --> 00:15:58,480 wheels and platform and handle. You can construct 361 00:15:58,480 --> 00:16:01,039 a model of this new object instantly by recombining 362 00:16:01,039 --> 00:16:03,379 old parts you already know. And deep networks 363 00:16:03,379 --> 00:16:05,720 struggle with this. They often struggle to reconfigure 364 00:16:05,720 --> 00:16:08,299 parts they haven't seen together before. If a 365 00:16:08,299 --> 00:16:10,600 network is trained on cars and bicycles, and 366 00:16:10,600 --> 00:16:13,179 then you show it a unicycle, it might be completely 367 00:16:13,179 --> 00:16:15,750 baffled. Because the pattern of a single wheel 368 00:16:15,750 --> 00:16:18,870 with a seat doesn't match its statistical database. 369 00:16:19,509 --> 00:16:22,370 It looks for the bike pattern and fails. But 370 00:16:22,370 --> 00:16:24,629 a human says, oh, it's like a bike, but with 371 00:16:24,629 --> 00:16:28,669 only one wheel. We compose and decompose. That 372 00:16:28,669 --> 00:16:31,029 leads to the second sub ingredient, causality. 373 00:16:31,269 --> 00:16:34,029 We touched on this with handwriting. Yes. Causality 374 00:16:34,029 --> 00:16:36,450 is understanding the process that generates the 375 00:16:36,450 --> 00:16:38,529 data. It's not just the shape of the letter. 376 00:16:38,870 --> 00:16:40,929 It's the motor program of the hand that drew 377 00:16:40,929 --> 00:16:43,169 it. This is why we can spot a fake signature, 378 00:16:43,330 --> 00:16:46,090 right? Even if it looks good superficially. Often, 379 00:16:46,309 --> 00:16:49,049 yes. You can tell if a line was drawn slowly 380 00:16:49,049 --> 00:16:52,029 and shakily or quickly and confidently. You are 381 00:16:52,029 --> 00:16:54,990 reverse -engineering the cause. The paper argues 382 00:16:54,990 --> 00:16:57,210 that this causal understanding is what allows 383 00:16:57,210 --> 00:16:59,350 us to pass that visual Turing test we mentioned. 384 00:16:59,450 --> 00:17:01,629 Because I know how it's made. If you know how 385 00:17:01,629 --> 00:17:04,009 a character is drawn, you can invent a new one 386 00:17:04,009 --> 00:17:06,269 that follows the same rules of construction. 387 00:17:06,750 --> 00:17:08,430 You're running your little mental program with 388 00:17:08,430 --> 00:17:11,390 new parameters. And finally, the third sub -ingredient. 389 00:17:12,130 --> 00:17:14,589 Learning to learn. This sounds like meta -learning. 390 00:17:14,970 --> 00:17:17,529 This is crucial. It's how we get so good at learning. 391 00:17:17,849 --> 00:17:20,170 We use knowledge from previous classes of objects 392 00:17:20,170 --> 00:17:23,470 to learn new ones faster. So if I learn the Greek 393 00:17:23,470 --> 00:17:26,130 alphabet, it helps me learn the Hebrew alphabet, 394 00:17:26,150 --> 00:17:28,210 even though the letters are different. In a way. 395 00:17:28,609 --> 00:17:30,529 You learn what an alphabet is. You learn the 396 00:17:30,529 --> 00:17:31,950 characters are usually made of simple strokes. 397 00:17:31,990 --> 00:17:33,549 They have standard size. They are distinct from 398 00:17:33,549 --> 00:17:36,210 each other. You build a higher level model. The 399 00:17:36,210 --> 00:17:38,910 paper calls this hierarchical knowledge. Right. 400 00:17:39,049 --> 00:17:41,539 Level one is learning a specific letter. like 401 00:17:41,539 --> 00:17:44,400 alpha, level two is learning what do letters 402 00:17:44,400 --> 00:17:47,420 look like in general. Neural networks typically 403 00:17:47,420 --> 00:17:49,519 start from scratch every time. They don't have 404 00:17:49,519 --> 00:17:51,700 that level two knowledge saying, hey, this is 405 00:17:51,700 --> 00:17:53,799 a handwritten character, so expect strokes and 406 00:17:53,799 --> 00:17:56,660 lines. They don't learn the context. BPL and 407 00:17:56,660 --> 00:18:00,140 humans transfer that high level knowledge to 408 00:18:00,140 --> 00:18:02,180 speed up the specific learning. It's like, I 409 00:18:02,180 --> 00:18:04,559 know how distinct strokes usually connect. So 410 00:18:04,559 --> 00:18:06,240 when I see a new character, I can immediately 411 00:18:06,240 --> 00:18:08,259 start breaking it down into parts I understand. 412 00:18:08,539 --> 00:18:13,130 So if we have BPL. this amazing, flexible, programmable 413 00:18:13,130 --> 00:18:16,519 way of thinking, why do we need neural networks 414 00:18:16,519 --> 00:18:20,799 at all? Or rather, why is the brain not just 415 00:18:20,799 --> 00:18:23,819 a slow, logical computer? Because sometimes I 416 00:18:23,819 --> 00:18:26,259 don't reason things out, sometimes I just react. 417 00:18:26,460 --> 00:18:29,200 That is the perfect segue to section five, real 418 00:18:29,200 --> 00:18:31,619 -time thinking. The problem with Bayesian inference 419 00:18:31,619 --> 00:18:33,960 and model building is that it is computationally 420 00:18:33,960 --> 00:18:37,279 expensive. It's slow. It's very slow. If a tiger 421 00:18:37,279 --> 00:18:39,599 jumps out at you, you don't have time to construct 422 00:18:39,599 --> 00:18:42,380 a causal model of the tiger's biology and for 423 00:18:42,380 --> 00:18:45,140 its goals using intuitive psychology and calculate 424 00:18:45,140 --> 00:18:47,160 the physics of its trajectory. You just run. 425 00:18:47,279 --> 00:18:48,930 You need to run. That's pattern recognition. 426 00:18:49,170 --> 00:18:51,789 A big, stripy, toothy pattern means danger. So 427 00:18:51,789 --> 00:18:54,329 it's a trade -off. Exactly. This is the classic 428 00:18:54,329 --> 00:18:56,910 system one and system two idea from Daniel Kahneman. 429 00:18:57,109 --> 00:18:59,250 System one is fast, intuitive pattern matching. 430 00:18:59,769 --> 00:19:02,710 System two is slow, deliberative, model -based 431 00:19:02,710 --> 00:19:04,609 reasoning. And the argument in the paper is that 432 00:19:04,609 --> 00:19:06,309 the brain uses both, and they work together. 433 00:19:06,430 --> 00:19:09,130 Yes. And they introduce a fascinating concept 434 00:19:09,130 --> 00:19:12,170 called amortized inference. Amortized inference. 435 00:19:12,269 --> 00:19:13,910 OK, unpack that. It sounds like something from 436 00:19:13,910 --> 00:19:16,569 an accounting class. It sort of is. Think of 437 00:19:16,569 --> 00:19:19,549 it as muscle memory for thought. Let's say you 438 00:19:19,549 --> 00:19:21,690 are learning to solve a specific type of math 439 00:19:21,690 --> 00:19:25,289 problem. At first, you use model -based thinking, 440 00:19:25,650 --> 00:19:29,809 system two, step by step, logical, slow. You 441 00:19:29,809 --> 00:19:32,109 are running the program. Painfully slow. Like 442 00:19:32,109 --> 00:19:33,730 when you first learn to drive a stick shift, 443 00:19:33,849 --> 00:19:35,710 you're thinking, OK, clutch in, move the stick, 444 00:19:35,849 --> 00:19:39,210 gas down, clutch out. Exactly. But after you 445 00:19:39,210 --> 00:19:42,450 have done it 100 times, you start to just see 446 00:19:42,450 --> 00:19:45,529 the answer or feel the gear shift. You recognize 447 00:19:45,529 --> 00:19:47,029 the pattern. You don't think about it anymore. 448 00:19:47,150 --> 00:19:48,869 You are no longer running the full causal model. 449 00:19:49,150 --> 00:19:51,670 You are simply mapping the input to the output 450 00:19:51,670 --> 00:19:54,690 based on past experience. You have amortized 451 00:19:54,690 --> 00:19:57,369 the cost of that thinking into a fast neural 452 00:19:57,369 --> 00:20:00,470 reflex. System 1 takes over. So the neural network, 453 00:20:00,609 --> 00:20:02,490 the pattern recognizer, is actually trained by 454 00:20:02,490 --> 00:20:04,549 the model builder. The slow, thoughtful part 455 00:20:04,549 --> 00:20:06,910 of me teaches the fast, reactive part of me. 456 00:20:07,069 --> 00:20:09,849 That is the theory. The paper mentions the Helmholtz 457 00:20:09,849 --> 00:20:12,339 machine idea from the 90s. Imagine your brain 458 00:20:12,339 --> 00:20:15,640 uses its causal models to dream or imagine situations. 459 00:20:16,019 --> 00:20:18,319 It generates its own data and then trains its 460 00:20:18,319 --> 00:20:20,799 fast pattern recognition networks on that imagined 461 00:20:20,799 --> 00:20:23,279 data. Wow, that sounds like what happens when 462 00:20:23,279 --> 00:20:26,799 we sleep. Consolidation. We replay the day's 463 00:20:26,799 --> 00:20:29,460 events. It connects beautifully to sleep research. 464 00:20:30,460 --> 00:20:32,900 Offline processing might be the brain transferring 465 00:20:32,900 --> 00:20:36,180 knowledge from the slow, flexible, model -based 466 00:20:36,180 --> 00:20:39,559 systems to the fast, efficient reflex systems. 467 00:20:39,779 --> 00:20:42,480 That is a light bulb moment for me. It's a powerful 468 00:20:42,480 --> 00:20:44,920 idea. So the authors aren't saying neural networks 469 00:20:44,920 --> 00:20:47,740 are useless. They are saying neural networks 470 00:20:47,740 --> 00:20:50,039 are the output of a deeper learning process, 471 00:20:50,160 --> 00:20:52,480 not the whole process itself. Exactly. If you 472 00:20:52,480 --> 00:20:55,220 only build the fast reactions, like current AI, 473 00:20:55,900 --> 00:20:58,140 you get brittle systems that break when the game 474 00:20:58,140 --> 00:21:00,039 changes. What if you have the model? If you have 475 00:21:00,039 --> 00:21:01,880 the model, you can retrain the reflex. You can 476 00:21:01,880 --> 00:21:04,859 adapt. OK. We've covered a lot of ground. We've 477 00:21:04,859 --> 00:21:06,920 bashed deep learning a little bit, praised the 478 00:21:06,920 --> 00:21:09,829 human mind, and built a roadmap. But surely there 479 00:21:09,829 --> 00:21:11,769 are people screaming at their speakers right 480 00:21:11,769 --> 00:21:14,569 now with objections. Section 6 addresses the 481 00:21:14,569 --> 00:21:16,890 critiques. Oh, for sure. And the most common 482 00:21:16,890 --> 00:21:19,529 critique is the unfair comparison. Right. The 483 00:21:19,529 --> 00:21:22,150 argument goes, hey, you're comparing a newborn 484 00:21:22,150 --> 00:21:25,569 AI to an adult human. Humans have years of experience. 485 00:21:25,910 --> 00:21:27,990 The neural net starts from nothing. If the net 486 00:21:27,990 --> 00:21:30,190 had 20 years of pre -training, it would match 487 00:21:30,190 --> 00:21:33,250 the human. It's a fair observation. But the authors 488 00:21:33,250 --> 00:21:36,210 have a really sharp rebuttal. I say. That is 489 00:21:36,210 --> 00:21:38,509 exactly the point. That's the feature, not the 490 00:21:38,509 --> 00:21:40,549 bub. Right. The challenge isn't just more data. 491 00:21:41,089 --> 00:21:43,309 The challenge is how to represent that prior 492 00:21:43,309 --> 00:21:46,230 experience so it can be used effectively. Humans 493 00:21:46,230 --> 00:21:48,930 don't just have a hard drive full of random video 494 00:21:48,930 --> 00:21:52,380 clips from their childhood. They have structured 495 00:21:52,380 --> 00:21:54,680 compositional causal knowledge, that startup 496 00:21:54,680 --> 00:21:56,779 software we talked about. The critique that humans 497 00:21:56,779 --> 00:21:59,119 have more prior knowledge just reinforces the 498 00:21:59,119 --> 00:22:01,440 argument that we need to figure out how to give 499 00:22:01,440 --> 00:22:03,940 machines that same kind of rich prior knowledge. 500 00:22:04,059 --> 00:22:05,920 It's not about the quantity of the data. It's 501 00:22:05,920 --> 00:22:07,920 about the quality of the structure holding it. 502 00:22:07,980 --> 00:22:11,720 Perfectly said. OK. Critique number two, biological 503 00:22:11,720 --> 00:22:14,720 plausibility. This one comes from the neuroscientists, 504 00:22:14,839 --> 00:22:17,380 right? The brain is a neural network, so we should 505 00:22:17,380 --> 00:22:19,960 only use neural networks to model it. This is 506 00:22:19,960 --> 00:22:23,180 the hardware fallacy. And the authors use the 507 00:22:23,180 --> 00:22:25,759 aerodynamics versus imitating a bird analogy, 508 00:22:25,859 --> 00:22:28,500 which is just fantastic. I love this one. The 509 00:22:28,500 --> 00:22:30,519 Wright brothers didn't succeed by building a 510 00:22:30,519 --> 00:22:32,680 machine that flapped its wings and had feathers. 511 00:22:32,920 --> 00:22:36,779 No, they tried that and it failed. They succeeded 512 00:22:36,779 --> 00:22:39,059 by understanding the principles of aerodynamics, 513 00:22:39,339 --> 00:22:42,319 lift, drag, thrust. They understood the physics, 514 00:22:42,680 --> 00:22:45,859 not just the surface level biology. OK. The argument 515 00:22:45,859 --> 00:22:49,539 here is that we need to understand The software 516 00:22:49,539 --> 00:22:52,720 algorithms of intelligent symbolic processing, 517 00:22:53,180 --> 00:22:56,339 causal models, before we obsess over the hardware 518 00:22:56,339 --> 00:22:59,440 implementation of neurons. Just because the brain 519 00:22:59,440 --> 00:23:02,079 uses neurons doesn't mean the best way to explain 520 00:23:02,079 --> 00:23:04,480 intelligence is to simulate neurons one by one 521 00:23:04,480 --> 00:23:06,700 without a higher level theory. We need to understand 522 00:23:06,700 --> 00:23:09,019 the code, not just the transistors it runs on. 523 00:23:09,160 --> 00:23:11,319 Exactly. What computations are those neurons 524 00:23:11,319 --> 00:23:13,460 performing? That's the key question. Okay, final 525 00:23:13,460 --> 00:23:16,099 critique. What about language? The paper doesn't 526 00:23:16,099 --> 00:23:18,079 talk much about language and isn't that the pinnacle 527 00:23:18,079 --> 00:23:20,599 of human intelligence. You see things like GPT 528 00:23:20,599 --> 00:23:23,059 -3 and it feels like we're close. Language is 529 00:23:23,059 --> 00:23:26,180 huge, but the authors argue that it sits on top 530 00:23:26,180 --> 00:23:28,359 of the startup software. It's built upon it. 531 00:23:28,440 --> 00:23:31,339 How so? You can't learn the word ball until you 532 00:23:31,339 --> 00:23:33,920 understand the concept of an object and the physics 533 00:23:33,920 --> 00:23:36,859 of solidity. You can't learn the word give until 534 00:23:36,859 --> 00:23:39,480 you understand agency and goals and other minds. 535 00:23:39,599 --> 00:23:42,460 So language is the walls and the roof. But intuitive 536 00:23:42,460 --> 00:23:45,099 physics and intuitive psychology are the foundation. 537 00:23:45,480 --> 00:23:48,039 You need the foundation first. Yes. You need 538 00:23:48,039 --> 00:23:51,240 to understand the world before you can talk about 539 00:23:51,240 --> 00:23:54,759 it. Language maps onto these pre -existing conceptual 540 00:23:54,759 --> 00:23:57,720 structures. This brings us to the outro. What 541 00:23:57,720 --> 00:24:00,119 is the synthesis here? It feels like we are moving 542 00:24:00,119 --> 00:24:02,819 towards a hybrid model, not a competition. That 543 00:24:02,819 --> 00:24:05,660 is the vision. The future isn't neural nets versus 544 00:24:05,660 --> 00:24:08,619 cognitive science. It's the integration. The 545 00:24:08,619 --> 00:24:11,099 paper points towards something they call differentiable 546 00:24:11,099 --> 00:24:13,079 programming. Which sounds like the best of both 547 00:24:13,079 --> 00:24:15,980 worlds. It is. It's about combining the powerful 548 00:24:15,980 --> 00:24:18,259 gradient -based learning of deep learning with 549 00:24:18,259 --> 00:24:20,740 the structure, compositionality, and reasoning 550 00:24:20,740 --> 00:24:23,460 of program induction. We want machines that can 551 00:24:23,460 --> 00:24:25,700 learn from massive data and reason from first 552 00:24:25,700 --> 00:24:27,980 principles. And why does this matter to the listener? 553 00:24:28,240 --> 00:24:30,599 Why should we care if machines think like this? 554 00:24:31,039 --> 00:24:34,180 What's the practical application? Imagine robotics. 555 00:24:34,819 --> 00:24:37,140 Right now, if you want a robot to fold laundry, 556 00:24:37,759 --> 00:24:39,680 you might have to train it on millions of towels. 557 00:24:40,160 --> 00:24:43,200 If you give it a shirt, it fails. It has no concept 558 00:24:43,200 --> 00:24:45,900 of laundry. Exactly. We want robots that can 559 00:24:45,900 --> 00:24:48,240 learn a new tool or a new environment instantly, 560 00:24:48,339 --> 00:24:50,400 like a human using a new drill for the first 561 00:24:50,400 --> 00:24:53,500 time. We want data efficiency. We want systems 562 00:24:53,500 --> 00:24:55,240 that don't need the entire internet to learn 563 00:24:55,240 --> 00:24:58,079 what a cat is. That's the dream. A robot you 564 00:24:58,079 --> 00:25:00,559 can show once, here's how you load the dishwasher, 565 00:25:00,859 --> 00:25:03,579 and it gets it. Because it understands dishes, 566 00:25:04,000 --> 00:25:06,759 water, And cleaning, not just pixel coordinates 567 00:25:06,759 --> 00:25:09,119 and motor commands, it has a model of the world. 568 00:25:09,460 --> 00:25:12,240 Before we wrap up, there is one final slightly 569 00:25:12,240 --> 00:25:13,759 provocative thought at the end of the paper that 570 00:25:13,759 --> 00:25:17,579 I found fascinating. It's about evolution. Ah, 571 00:25:17,720 --> 00:25:21,039 yes, the stochastic hill climbing concept. The 572 00:25:21,039 --> 00:25:23,099 authors suggest that if we can't code these complex 573 00:25:23,099 --> 00:25:25,799 structures by hand, this startup software, and 574 00:25:25,799 --> 00:25:27,900 if gradient descent, which is standard deep learning, 575 00:25:28,099 --> 00:25:30,599 can't find them, then maybe we are missing the 576 00:25:30,599 --> 00:25:33,430 biggest algorithm of all, evolution. It's a bit 577 00:25:33,430 --> 00:25:36,589 of a mic drop moment. Evolution is a messy, slow, 578 00:25:36,869 --> 00:25:39,589 random search process. It took millions of years 579 00:25:39,589 --> 00:25:42,109 of trial and error to build the human brain's 580 00:25:42,109 --> 00:25:45,670 architecture. So, what are they suggesting? They 581 00:25:45,670 --> 00:25:48,960 ask a profound question. Are we trying to skip 582 00:25:48,960 --> 00:25:51,160 the millions of years of evolution that gave 583 00:25:51,160 --> 00:25:54,059 us our startup software in the first place? We 584 00:25:54,059 --> 00:25:58,059 want the result. The baby's mind. Without the 585 00:25:58,059 --> 00:26:01,569 process. without the millions of years of death 586 00:26:01,569 --> 00:26:04,190 and selection that forged it. It suggests that 587 00:26:04,190 --> 00:26:06,869 maybe the real path forward is to simulate that 588 00:26:06,869 --> 00:26:09,410 evolutionary process to build the right architectures, 589 00:26:09,710 --> 00:26:12,210 which then can learn rapidly like a child. We 590 00:26:12,210 --> 00:26:13,769 might be trying to make the child learn things 591 00:26:13,769 --> 00:26:16,779 that the species learned over eons. That is a 592 00:26:16,779 --> 00:26:19,039 heavy thought. We might be asking our AI to do 593 00:26:19,039 --> 00:26:21,119 too much learning and not enough inheriting. 594 00:26:21,500 --> 00:26:23,259 We're giving it a blank slate when it needs a 595 00:26:23,259 --> 00:26:25,319 genetic history. It's a humbling perspective 596 00:26:25,319 --> 00:26:27,299 on the challenge ahead. This has been a massive 597 00:26:27,299 --> 00:26:29,740 deep dive. We've gone from Atari to handwritten 598 00:26:29,740 --> 00:26:32,640 digits, from neurons to Bayesian programs. It 599 00:26:32,640 --> 00:26:34,960 really changes how you look at the AI revolution, 600 00:26:35,119 --> 00:26:36,900 doesn't it? It feels like we are just at the 601 00:26:36,900 --> 00:26:39,039 beginning of the real challenge. I think so. 602 00:26:39,660 --> 00:26:44,000 We are past... the magic trick phase of AI. Now 603 00:26:44,000 --> 00:26:46,079 we have to do the hard work of understanding 604 00:26:46,079 --> 00:26:48,259 the mind. Thank you so much for breaking this 605 00:26:48,259 --> 00:26:50,299 down with me. This was incredible. And to our 606 00:26:50,299 --> 00:26:52,799 listeners, thanks for tuning in. As always, keep 607 00:26:52,799 --> 00:26:55,279 asking the hard questions. Keep thinking. We'll 608 00:26:55,279 --> 00:26:56,779 see you next time on the Deep Dive.