WEBVTT 00:00:00.000 --> 00:00:02.620 Imagine you build a machine to find like the 00:00:02.620 --> 00:00:04.919 absolute best software engineers in the country. 00:00:05.240 --> 00:00:07.740 You feed it this mountain of resumes. You turn 00:00:07.740 --> 00:00:10.519 it on and it instantly starts filtering the top 00:00:10.519 --> 00:00:13.400 candidates with, you know, just incredible efficiency. 00:00:13.699 --> 00:00:15.740 Right. Sounds like the dream scenario for any 00:00:15.740 --> 00:00:19.059 HR department. Exactly. But then you look a little 00:00:19.059 --> 00:00:21.800 closer at the math and you realize the machine 00:00:21.800 --> 00:00:25.620 has secretly taught itself to reject any applicant 00:00:25.620 --> 00:00:28.519 whose resume contains the word woman. Oh, wow. 00:00:28.760 --> 00:00:30.679 Yeah. And this isn't some hypothetical thought 00:00:30.679 --> 00:00:32.679 experiment. It actually happened at Amazon in 00:00:32.679 --> 00:00:35.520 2018. They had to scrap the entire project. Wow. 00:00:35.780 --> 00:00:39.060 Because the machine wasn't programmed to be biased. 00:00:39.240 --> 00:00:42.479 It wasn't harboring a grudge. It was just ruthlessly 00:00:42.479 --> 00:00:45.079 executing a mathematical equation based on the 00:00:45.079 --> 00:00:47.250 data it was given. Which is exactly why we need 00:00:47.250 --> 00:00:49.270 to understand the underlying mechanics of these 00:00:49.270 --> 00:00:51.509 systems. I mean, as a society, we tend to talk 00:00:51.509 --> 00:00:54.390 about artificial intelligence as if it's, well, 00:00:54.509 --> 00:00:57.149 like it's magic. We throw around terms like neural 00:00:57.149 --> 00:01:00.409 networks and we picture this glowing thinking 00:01:00.409 --> 00:01:03.909 brain in a jar. But once you peel back the layers 00:01:03.909 --> 00:01:06.989 and look at the actual architecture, the magic 00:01:06.989 --> 00:01:08.930 kind of fades away. Yeah, the illusion breaks. 00:01:09.209 --> 00:01:12.370 Exactly. What you're left with is a fascinating 00:01:12.730 --> 00:01:15.950 highly flawed, and honestly incredibly powerful 00:01:15.950 --> 00:01:18.569 statistical tool. And that is our mission for 00:01:18.569 --> 00:01:20.670 you today. If you're listening to this, you're 00:01:20.670 --> 00:01:22.849 likely someone who wants to cut through all the 00:01:22.849 --> 00:01:28.430 endless hype and the dense Silicon Valley jargon. 00:01:28.609 --> 00:01:30.150 Right, you want to know how it actually works. 00:01:30.290 --> 00:01:32.469 Yeah, how the invisible systems making decisions 00:01:32.469 --> 00:01:35.370 about your life actually function. So we're doing 00:01:35.370 --> 00:01:37.569 a deep dive into the foundational source material 00:01:37.569 --> 00:01:41.250 of the modern AI boom. We've gathered the comprehensive 00:01:41.250 --> 00:01:43.390 documentation on artificial neural networks, 00:01:43.469 --> 00:01:45.150 and we're going to break down what they are. 00:01:45.290 --> 00:01:47.329 We're going to look at their shockingly old origins, 00:01:47.650 --> 00:01:50.069 too. Yes, and explain the mathematical tricks 00:01:50.069 --> 00:01:53.290 that make them work and shine a bright light 00:01:53.290 --> 00:01:55.290 on why they fail. So to start, we really have 00:01:55.290 --> 00:01:58.709 to completely abandon that brain in a jar metaphor. 00:01:58.890 --> 00:02:01.379 OK, throw it out. Yeah, throw it away. If we 00:02:01.379 --> 00:02:04.159 look at the foundational architecture, an artificial 00:02:04.159 --> 00:02:06.719 neural network is simply a computational model 00:02:06.719 --> 00:02:09.500 made up of connected nodes. The data enters through 00:02:09.500 --> 00:02:12.139 an input layer, passes through a series of what 00:02:12.139 --> 00:02:14.840 are called hidden layers, and eventually produces 00:02:14.840 --> 00:02:17.199 a result at the output layer. OK, so instead 00:02:17.199 --> 00:02:20.219 of a brain, I actually think it's much more accurate 00:02:20.219 --> 00:02:24.759 to visualize a giant, incredibly complex recipe. 00:02:24.979 --> 00:02:27.400 I like that. Yeah, imagine you're making a massive 00:02:27.400 --> 00:02:30.789 batch of soup. The ingredients are your data. 00:02:31.409 --> 00:02:33.990 The nodes or the artificial neurons are the different 00:02:33.990 --> 00:02:36.409 stages of tasting and adjusting. And they're 00:02:36.409 --> 00:02:38.189 all connected by edges, which, you know, you 00:02:38.189 --> 00:02:40.129 can think of as the specific measurements for 00:02:40.129 --> 00:02:42.569 each ingredient. That recipe analogy holds up 00:02:42.569 --> 00:02:45.250 remarkably well mathematically, actually. Every 00:02:45.250 --> 00:02:47.469 connection between these nodes has a weight. 00:02:47.830 --> 00:02:50.030 A weight. Yeah, a weight is really just a number 00:02:50.030 --> 00:02:53.289 that determines how much influence one node has 00:02:53.289 --> 00:02:56.210 on the next. So if your soup needs a lot of salt, 00:02:56.610 --> 00:02:59.229 the salt node has a very heavy weight. The node 00:02:59.229 --> 00:03:02.150 takes all the inputs it receives, multiplies 00:03:02.150 --> 00:03:03.969 them by their specific weights, adds them all 00:03:03.969 --> 00:03:06.949 together, and then passes that total sum through 00:03:06.949 --> 00:03:09.129 something called an activation function. OK, 00:03:09.150 --> 00:03:11.889 let's slow down on that term, an activation function. 00:03:11.909 --> 00:03:14.229 Sure. If we stick to the kitchen analogy. Is 00:03:14.229 --> 00:03:17.169 that essentially a threshold? Like a chef tasting 00:03:17.169 --> 00:03:19.770 the soup and deciding, if this isn't spicy enough, 00:03:19.909 --> 00:03:22.110 I don't pass it to the next station. That is 00:03:22.110 --> 00:03:24.810 a perfect way to visualize it. Mathematically, 00:03:24.969 --> 00:03:27.669 it's a nonlinear function acting as a gatekeeper. 00:03:28.210 --> 00:03:30.930 If the sum of those weighted inputs hits a certain 00:03:30.930 --> 00:03:34.870 threshold, the neuron fires and passes the signal 00:03:34.870 --> 00:03:36.770 forward. And if it doesn't? If it doesn't hit 00:03:36.770 --> 00:03:38.669 the threshold, the signal just dies right there. 00:03:38.789 --> 00:03:42.389 Wow, OK. Yeah. And by stacking hundreds, thousands, 00:03:42.629 --> 00:03:45.490 or I mean millions of these nodes and constantly 00:03:45.490 --> 00:03:47.509 adjusting the weights between them, the network 00:03:47.509 --> 00:03:50.430 can model insanely complex relationships and 00:03:50.430 --> 00:03:52.930 recognize patterns in massive amounts of data. 00:03:53.129 --> 00:03:56.250 Which sounds incredibly modern, right? But to 00:03:56.250 --> 00:03:58.789 really grasp how these networks are shaping our 00:03:58.789 --> 00:04:01.229 current reality, we have to look at the history. 00:04:01.669 --> 00:04:03.710 Because we aren't starting in a shiny Silicon 00:04:03.710 --> 00:04:06.789 Valley garage in the 2000s. Not at all. The foundational 00:04:06.789 --> 00:04:09.949 math behind this goes back over 200 years. We 00:04:09.949 --> 00:04:12.969 are starting in the late 1700s. Yeah, the simplest 00:04:12.969 --> 00:04:15.349 form of a neural network is essentially a linear 00:04:15.349 --> 00:04:17.949 network. which relies on the method of least 00:04:17.949 --> 00:04:21.490 squares or linear regression. This mathematical 00:04:21.490 --> 00:04:23.589 technique was actually used by the mathematician 00:04:23.589 --> 00:04:27.490 Carl Friedrich Gauss in 1795 and then later published 00:04:27.490 --> 00:04:31.329 by Legendre in 1805 to predict planetary movement. 00:04:31.709 --> 00:04:34.930 That is wild. They were taking a set of astronomical 00:04:34.930 --> 00:04:37.670 data points and finding a line of best fit to 00:04:37.670 --> 00:04:39.850 predict where a planet would appear next in the 00:04:39.850 --> 00:04:42.089 sky. Wait, wait. So early AI was essentially 00:04:42.089 --> 00:04:45.230 just spicy 18th century statistics. Basically, 00:04:45.490 --> 00:04:47.949 yeah. Gauss was just drawing a statistical line 00:04:47.949 --> 00:04:50.230 through data points, and that same foundational 00:04:50.230 --> 00:04:52.250 logic is what powers machine learning today. 00:04:52.490 --> 00:04:55.730 At its most basic mathematical level, yes. You 00:04:55.730 --> 00:04:58.850 are finding the optimal line or curve to fit 00:04:58.850 --> 00:05:02.810 a set of data. That's crazy. It is. Now the leap 00:05:02.810 --> 00:05:06.069 toward computing happened much later, in 1943. 00:05:06.560 --> 00:05:09.339 Warren McCulloch and Walter Pitts created the 00:05:09.339 --> 00:05:12.399 first conceptual model for neural networks. Their 00:05:12.399 --> 00:05:14.600 model couldn't actually learn or adjust its own 00:05:14.600 --> 00:05:17.379 weights, but it proved that a network of artificial 00:05:17.379 --> 00:05:20.819 neurons could compute logical functions. And 00:05:20.819 --> 00:05:24.439 then in 1958, a psychologist named Frank Rosenblatt 00:05:24.439 --> 00:05:27.540 took that concept and invented the perceptron. 00:05:27.740 --> 00:05:29.560 And this is a detail from the history that always 00:05:29.560 --> 00:05:32.100 blows my mind. The perceptron wasn't just some 00:05:32.100 --> 00:05:34.660 academic side project. It was funded by the U 00:05:34.660 --> 00:05:37.290 .S. Office of Naval Research. Right, big military 00:05:37.290 --> 00:05:39.110 money. Yeah, the military was pouring money into 00:05:39.110 --> 00:05:41.810 this, and it kicked off what scientists at the 00:05:41.810 --> 00:05:44.329 time genuinely believed was the golden age of 00:05:44.329 --> 00:05:47.829 AI. There was this massive wave of optimism that 00:05:47.829 --> 00:05:50.439 perceptrons were going to like... quickly emulate 00:05:50.439 --> 00:05:52.980 human intelligence. The optimism was staggering, 00:05:53.600 --> 00:05:55.920 but it was also built on a very fragile foundation. 00:05:56.199 --> 00:05:58.660 The early perceptrons were incredibly shallow. 00:05:58.800 --> 00:06:00.399 Meaning what, exactly? Well, they essentially 00:06:00.399 --> 00:06:02.579 lacked those hidden layers we discussed earlier, 00:06:02.600 --> 00:06:04.860 meaning they could only solve what mathematicians 00:06:04.860 --> 00:06:08.399 call linearly separable problems. OK, we definitely 00:06:08.399 --> 00:06:12.540 need to unpack linearly separable, because that 00:06:12.540 --> 00:06:15.259 concept is the exact thing that caused the entire 00:06:15.259 --> 00:06:18.199 field of AI to just crash and burn in the 1960s. 00:06:18.199 --> 00:06:22.240 Yeah. How should we visualize a linearly separable 00:06:22.240 --> 00:06:24.819 problem? All right, think about a map of a field. 00:06:24.920 --> 00:06:26.839 On this field, you have a bunch of apple trees 00:06:26.839 --> 00:06:29.439 clustered on the left side and a bunch of orange 00:06:29.439 --> 00:06:31.680 trees clustered on the right. OK, I'm picturing 00:06:31.680 --> 00:06:34.660 it. If you can take a ruler and draw a single 00:06:34.660 --> 00:06:36.620 straight line right down the middle of the field 00:06:36.620 --> 00:06:38.819 to perfectly separate the apples from the oranges, 00:06:39.339 --> 00:06:42.240 that problem is linearly separable. Got it. The 00:06:42.240 --> 00:06:44.160 early perceptron could solve that easily. It 00:06:44.160 --> 00:06:46.259 could draw that single straight line. But the 00:06:46.259 --> 00:06:48.620 real world rarely lets you draw one straight 00:06:48.620 --> 00:06:51.720 line. Exactly. And that was the fatal flaw. In 00:06:51.720 --> 00:06:55.139 1969, Marvin Minsky and Seymour Papert published 00:06:55.139 --> 00:06:57.720 a highly influential book titled Perceptrons. 00:06:58.319 --> 00:07:00.620 They mathematically proved that these early networks 00:07:00.620 --> 00:07:03.600 couldn't solve basic logic puzzles. Like what? 00:07:03.860 --> 00:07:06.259 Most notably, something called the Exclusive 00:07:06.259 --> 00:07:10.230 or Circuit, or XOR. Let's apply that XOR problem 00:07:10.230 --> 00:07:13.550 to our field of trees. Sure. So in an XOR scenario, 00:07:14.189 --> 00:07:16.569 the trees aren't neatly clustered left and right. 00:07:17.290 --> 00:07:19.990 Imagine the field is divided into a grid of four 00:07:19.990 --> 00:07:22.649 squares. You have apple trees in the top left 00:07:22.649 --> 00:07:25.389 square and the bottom right square. And you have 00:07:25.389 --> 00:07:27.129 orange trees in the top right and the bottom 00:07:27.129 --> 00:07:29.350 left. So they're situated diagonally from each 00:07:29.350 --> 00:07:32.069 other. Oh, I see. So try taking a ruler and drawing 00:07:32.069 --> 00:07:34.649 a single straight line to separate all the apples 00:07:34.649 --> 00:07:36.810 from all the oranges now. Right. You can't do 00:07:36.810 --> 00:07:40.029 it. You would need to draw a circle or two intersecting 00:07:40.029 --> 00:07:43.310 lines or a curve. You would need a nonlinear 00:07:43.310 --> 00:07:46.500 boundary. But because the early perceptrons lacked 00:07:46.500 --> 00:07:48.920 those hidden layers, they could only ever draw 00:07:48.920 --> 00:07:51.959 one straight line. Minsky and Papert exposed 00:07:51.959 --> 00:07:54.480 this fundamental limitation, proving the machine 00:07:54.480 --> 00:07:57.220 was completely stumped by anything even slightly 00:07:57.220 --> 00:08:00.480 more complex than a basic, perfectly segregated 00:08:00.480 --> 00:08:02.660 data set. It's an incredible piece of history. 00:08:02.980 --> 00:08:05.040 The very technology we are currently worried 00:08:05.040 --> 00:08:07.360 might upend the global economy was, you know, 00:08:07.800 --> 00:08:10.800 in the late 60s, dismissed by the scientific 00:08:10.800 --> 00:08:13.220 community as a dead end. because it couldn't 00:08:13.220 --> 00:08:15.720 solve a diagonal puzzle. Yeah, it's pretty ironic. 00:08:16.240 --> 00:08:19.220 Funding completely dried up, research stagnated, 00:08:19.560 --> 00:08:22.579 and this ushered in an era known as the AI winter. 00:08:22.910 --> 00:08:26.149 It perfectly illustrates how dependent technological 00:08:26.149 --> 00:08:28.949 progress is on overcoming specific mathematical 00:08:28.949 --> 00:08:32.250 bottlenecks. To get out of the AI winter, researchers 00:08:32.250 --> 00:08:35.149 needed a way to add multiple hidden layers to 00:08:35.149 --> 00:08:37.269 the network. And more importantly, they needed 00:08:37.269 --> 00:08:39.629 a way for those hidden layers to actually learn 00:08:39.629 --> 00:08:42.330 from their mistakes. Which brings us to the breakthrough. 00:08:42.639 --> 00:08:45.639 If neural networks effectively died in the 70s, 00:08:45.960 --> 00:08:48.259 the only way we ended up with the modern AI of 00:08:48.259 --> 00:08:50.759 today is through a mathematical trick popularized 00:08:50.759 --> 00:08:54.100 in the 1980s. Right, a concept called backpropagation. 00:08:54.320 --> 00:08:56.419 And once again, to understand the future we have 00:08:56.419 --> 00:08:59.240 to look to the distant past. Yeah, backpropagation 00:08:59.240 --> 00:09:01.700 relies heavily on the chain rule from calculus, 00:09:01.899 --> 00:09:04.100 which was derived by Gottfried Wilhelm Leibniz 00:09:04.100 --> 00:09:08.779 all the way back in 1673. 1673? I know. Backpropagation 00:09:08.779 --> 00:09:10.960 is essentially an algorithm that calculates an 00:09:10.960 --> 00:09:13.240 error gradient and propagates it backward through 00:09:13.240 --> 00:09:15.460 the network. Let's break down error gradient 00:09:15.460 --> 00:09:19.159 using our recipe analogy. You've made your massive 00:09:19.159 --> 00:09:22.720 batch of soup. The data move forward. through 00:09:22.720 --> 00:09:25.179 the network, from the input ingredients, through 00:09:25.179 --> 00:09:27.720 the hidden layers of tasting and mixing, to the 00:09:27.720 --> 00:09:29.879 final output. That's your forward pass, yeah. 00:09:30.139 --> 00:09:33.440 Right. You taste the final soup and it is terrible. 00:09:33.799 --> 00:09:35.919 It's too salty and there's not enough garlic. 00:09:36.399 --> 00:09:38.679 That gap between the terrible soup you have and 00:09:38.679 --> 00:09:41.110 the perfect soup you want. That is your error. 00:09:41.389 --> 00:09:43.649 Exactly. Now you have to fix it. But you don't 00:09:43.649 --> 00:09:45.970 just randomly throw different amounts of ingredients 00:09:45.970 --> 00:09:48.049 into the next batch and hope for the best. No. 00:09:48.269 --> 00:09:51.350 You use calculus. Right. The gradient is essentially 00:09:51.350 --> 00:09:53.730 a mathematical slope. It tells you the exact 00:09:53.730 --> 00:09:55.990 direction and magnitude of the changes you need 00:09:55.990 --> 00:09:57.990 to make to reach the bottom of that slope, which 00:09:57.990 --> 00:10:01.389 represents zero error. It calculates the mathematical 00:10:01.389 --> 00:10:03.629 path of least resistance to a better tasting 00:10:03.629 --> 00:10:06.190 soup. Yes. It tells the system, OK, trace this 00:10:06.190 --> 00:10:08.350 error backward. lower the weight of the salt 00:10:08.350 --> 00:10:11.809 node by exactly 2 .4 % and increase the weight 00:10:11.809 --> 00:10:15.190 of the garlic node by 1 .8%. And the network 00:10:15.190 --> 00:10:17.870 literally learns by tracing its failures back 00:10:17.870 --> 00:10:20.190 to the source and tweaking the ingredients for 00:10:20.190 --> 00:10:22.710 the next forward pass. It's brilliant. But this 00:10:22.710 --> 00:10:24.789 learning process takes massive amounts of data 00:10:24.789 --> 00:10:27.600 to execute. When we look at how this is applied 00:10:27.600 --> 00:10:30.559 in practice, object detection provides a brilliant 00:10:30.559 --> 00:10:33.019 example of how these weighted adjustments actually 00:10:33.019 --> 00:10:35.840 manifest. Let's say we want to train our network 00:10:35.840 --> 00:10:39.039 to recognize a starfish versus a sea urchin. 00:10:39.179 --> 00:10:41.940 Okay, we feed the network thousands of images. 00:10:42.659 --> 00:10:45.039 Through constant back propagation and adjusting 00:10:45.039 --> 00:10:47.580 those mathematical weights, the network eventually 00:10:47.580 --> 00:10:50.460 learns that a starfish highly correlates with 00:10:50.460 --> 00:10:52.879 a star -shaped outline and a ringed texture. 00:10:53.080 --> 00:10:55.480 Right. And a sea urchin highly correlates with 00:10:55.480 --> 00:10:58.559 an oval shape and a striped texture. But consider 00:10:58.559 --> 00:11:01.159 what happens when the network encounters an anomaly 00:11:01.159 --> 00:11:04.360 in the training data. Say, a rare species of 00:11:04.360 --> 00:11:06.220 sea urchin that happens to have a ringed texture 00:11:06.220 --> 00:11:08.700 instead of stripes. The network processes that 00:11:08.700 --> 00:11:11.399 image. And because it's told this is a sea urchin, 00:11:11.639 --> 00:11:14.720 it creates a very weak, subtle mathematical association 00:11:14.720 --> 00:11:17.659 between the concept of a ringed texture and the 00:11:17.659 --> 00:11:19.919 output of sea urchin. Exactly. It adjusts the 00:11:19.919 --> 00:11:22.860 weights just a tiny bit. Right. Now later on, 00:11:23.159 --> 00:11:25.759 you show the network a picture of a normal starfish 00:11:25.759 --> 00:11:28.899 sitting on the ocean floor. The network sees 00:11:28.899 --> 00:11:31.500 the ringed texture of the starfish, and while 00:11:31.500 --> 00:11:34.000 it mostly sends a strong signal for starfish, 00:11:34.659 --> 00:11:37.179 that weak association from earlier also sends 00:11:37.179 --> 00:11:40.000 a tiny signal to the sea urchin output. Okay, 00:11:40.059 --> 00:11:42.120 I see where this is going. Now imagine there 00:11:42.120 --> 00:11:45.039 is a smooth oval -shaped rock sitting in the 00:11:45.039 --> 00:11:48.519 background of the photo. Oh. The oval rock triggers 00:11:48.519 --> 00:11:51.500 the oval shape criteria for a sea urchin. Centrally, 00:11:51.820 --> 00:11:54.279 the weak signal from the ring texture combines 00:11:54.279 --> 00:11:56.740 with the weak signal from the oval rock. The 00:11:56.740 --> 00:11:59.220 activation function's threshold is crossed. And 00:11:59.220 --> 00:12:01.399 fired. The network outputs a false positive, 00:12:02.080 --> 00:12:03.799 confidently declaring there's a sea urchin in 00:12:03.799 --> 00:12:06.220 the photo when there isn't. That's fascinating. 00:12:06.340 --> 00:12:08.340 It's like a toddler pointing at a cow and calling 00:12:08.340 --> 00:12:10.340 it a dog simply because it has four legs and 00:12:10.340 --> 00:12:13.019 a tail. Yes. The machine isn't thinking, I see 00:12:13.019 --> 00:12:16.200 a sea urchin. It's just blindly following weighted 00:12:16.200 --> 00:12:19.419 visual breadcrumbs. And how you lay out those 00:12:19.419 --> 00:12:21.620 breadcrumbs dictates everything. The training 00:12:21.620 --> 00:12:24.620 paradigm is everything. The starfish example 00:12:24.620 --> 00:12:27.210 is what we call supervised learning. You have 00:12:27.210 --> 00:12:29.629 a teacher providing the network with labeled 00:12:29.629 --> 00:12:33.009 data. You explicitly tell it, this image is a 00:12:33.009 --> 00:12:36.149 starfish, this one is an urchin, so it can calculate 00:12:36.149 --> 00:12:39.409 its error. But labeling millions of images is 00:12:39.409 --> 00:12:42.190 incredibly time consuming. Which leads to unsupervised 00:12:42.190 --> 00:12:44.470 learning. Exactly. In unsupervised learning, 00:12:44.490 --> 00:12:47.730 you just feed the network raw, unlabeled data. 00:12:47.830 --> 00:12:50.210 You don't tell it what it's looking at. The network's 00:12:50.210 --> 00:12:52.990 only job is to analyze the data and find hidden 00:12:52.990 --> 00:12:55.669 structures or clusters on its own. So it's organizing 00:12:55.669 --> 00:12:58.389 on its own. Yeah, it might group all the spiky 00:12:58.389 --> 00:13:01.230 objects into one pile and all the smooth objects 00:13:01.230 --> 00:13:04.230 into another without ever knowing the word starfish 00:13:04.230 --> 00:13:06.590 or rock. And then there is the paradigm that 00:13:06.590 --> 00:13:09.029 usually powers the really flashy headlines, right? 00:13:09.470 --> 00:13:12.190 Like AI mastering complex video games or beating 00:13:12.190 --> 00:13:14.850 grandmasters at chess, reinforcement learning. 00:13:15.149 --> 00:13:17.149 Reinforcement learning operates on a completely 00:13:17.149 --> 00:13:19.809 different framework. The environment the AI operates 00:13:19.809 --> 00:13:22.970 in is modeled as a Markov decision process. OK, 00:13:22.990 --> 00:13:26.049 let's define a Markov decision process. That 00:13:26.049 --> 00:13:29.029 sounds intimidating, but it's actually very intuitive. 00:13:29.330 --> 00:13:31.409 It essentially means the AI is only concerned 00:13:31.409 --> 00:13:33.690 with its current state and the immediate actions 00:13:33.690 --> 00:13:36.049 available to it rather than memorizing the entire 00:13:36.049 --> 00:13:38.210 history of how it got there. Like a mouse in 00:13:38.210 --> 00:13:41.049 a maze. Yes. Imagine a mouse navigating a maze. 00:13:41.419 --> 00:13:45.200 The AI is the mouse. It takes an action, like 00:13:45.200 --> 00:13:47.840 turning left or right. The environment then gives 00:13:47.840 --> 00:13:50.759 it a response, either a reward, like cheese, 00:13:50.860 --> 00:13:53.200 or a penalty, like hitting a dead end. Got it. 00:13:53.360 --> 00:13:55.460 Through millions of cycles of trial and error, 00:13:55.700 --> 00:13:58.100 the network continually updates its policy to 00:13:58.100 --> 00:14:00.600 maximize its reward score. It's just a hyperfast 00:14:00.600 --> 00:14:03.440 version of playing a video game, dying, learning 00:14:03.440 --> 00:14:06.019 where the trap is, and trying again. Precisely. 00:14:06.159 --> 00:14:08.340 But because these networks are teaching themselves 00:14:08.340 --> 00:14:10.559 through billions of trial and error adjustments 00:14:10.559 --> 00:14:13.940 or finding hidden clusters in massive unlabeled 00:14:13.940 --> 00:14:16.840 data sets, we run into a monumental problem. 00:14:17.159 --> 00:14:19.039 We don't actually know what the machine has learned 00:14:19.039 --> 00:14:21.470 until it makes a mistake. This is universally 00:14:21.470 --> 00:14:24.330 known as the black box problem, and it is arguably 00:14:24.330 --> 00:14:27.309 the most critical vulnerability in modern AI 00:14:27.309 --> 00:14:29.389 deployment. Which brings us right back to the 00:14:29.389 --> 00:14:31.309 Amazon recruiting tool from the start of our 00:14:31.309 --> 00:14:33.909 deep dive. Yes, perfect example. That entire 00:14:33.909 --> 00:14:36.889 failure was a textbook example of data set bias, 00:14:37.590 --> 00:14:39.929 which is a direct consequence of the black box. 00:14:41.169 --> 00:14:44.590 Amazon fed the AI 10 years of historical hiring 00:14:44.590 --> 00:14:48.350 data. The problem? The tech industry has historically 00:14:48.350 --> 00:14:52.090 been male -dominated. The majority of successful 00:14:52.090 --> 00:14:55.389 resumes in that old data belonged to men. And 00:14:55.389 --> 00:14:57.649 the neural network wasn't explicitly programmed 00:14:57.649 --> 00:15:00.450 to filter by gender. It was simply told to find 00:15:00.450 --> 00:15:02.750 the statistical patterns of a successful hire 00:15:02.750 --> 00:15:06.029 based on the past data. The math found an incredibly 00:15:06.029 --> 00:15:08.370 strong correlation between male applicants and 00:15:08.370 --> 00:15:11.269 getting the job. Wow. It optimized for that pattern 00:15:11.269 --> 00:15:14.049 ruthlessly to the point of actively penalizing 00:15:14.049 --> 00:15:16.730 resumes that included the word women or listed 00:15:16.730 --> 00:15:19.730 a women's college. It drew a line of best fit 00:15:19.730 --> 00:15:22.480 through a flawed reality. And even if you manage 00:15:22.480 --> 00:15:24.879 to scrub all the historical bias out of your 00:15:24.879 --> 00:15:27.340 data, you still have to deal with what the documentation 00:15:27.340 --> 00:15:31.179 calls concept drift. Concept drift, or non -stationarity, 00:15:31.480 --> 00:15:34.240 occurs when the real world evolves, but the model's 00:15:34.240 --> 00:15:36.460 training data remains static. What does that 00:15:36.460 --> 00:15:39.059 look like in practice? Well... You deploy a highly 00:15:39.059 --> 00:15:41.539 accurate model today, but over the next few months, 00:15:42.000 --> 00:15:44.259 consumer habits change or the economy shifts. 00:15:45.059 --> 00:15:47.419 The statistical properties of the incoming real 00:15:47.419 --> 00:15:49.639 -world data drift away from the baseline the 00:15:49.639 --> 00:15:52.279 model was trained on. Oh, I see. If you aren't 00:15:52.279 --> 00:15:54.399 constantly monitoring and updating the weights, 00:15:55.059 --> 00:15:57.960 the model's predictive accuracy quietly degrades, 00:15:58.360 --> 00:16:00.960 leading to catastrophic decisions. And when those 00:16:00.960 --> 00:16:03.850 catastrophic decisions happen... we hit the core 00:16:03.850 --> 00:16:08.070 of the black box critique, a total lack of interpretability. 00:16:08.169 --> 00:16:10.610 Yeah, that's the scary part. If a deep learning 00:16:10.610 --> 00:16:14.490 model denies you a mortgage or a medical AI misdiagnoses 00:16:14.490 --> 00:16:17.429 your lung scan, it cannot explain why. It cannot 00:16:17.429 --> 00:16:19.529 walk you through its reasoning. It just spits 00:16:19.529 --> 00:16:21.909 out the final result of an unbelievably complex 00:16:21.909 --> 00:16:24.480 math equation. This lack of transparency has 00:16:24.480 --> 00:16:26.879 actually sparked fierce philosophical debates. 00:16:27.320 --> 00:16:30.159 In 1997, the mathematician and writer Alexander 00:16:30.159 --> 00:16:32.820 Dudney published a blistering critique of the 00:16:32.820 --> 00:16:35.659 entire field. He labeled artificial neural networks 00:16:35.659 --> 00:16:39.200 a lazy science. Right. Dudney argued that these 00:16:39.200 --> 00:16:42.580 networks possess a something for nothing quality. 00:16:43.200 --> 00:16:45.740 You throw raw data at a wall, the machine churns 00:16:45.740 --> 00:16:48.080 through its hidden layers, and a solution appears 00:16:48.080 --> 00:16:51.159 almost by magic. No human mind intervenes in 00:16:51.159 --> 00:16:53.460 the logic. Exactly. And more importantly, no 00:16:53.460 --> 00:16:56.159 one actually learns anything about the underlying 00:16:56.159 --> 00:16:59.059 principles of the problem. If we use a neural 00:16:59.059 --> 00:17:01.460 network to predict weather patterns, and it works, 00:17:01.940 --> 00:17:04.940 we still haven't learned any new laws of meteorology. 00:17:05.180 --> 00:17:07.819 We just have an opaque table of numbers. It forces 00:17:07.819 --> 00:17:11.160 us to ask. Are we even doing science anymore, 00:17:11.380 --> 00:17:14.160 or are we just brute forcing statistical correlations? 00:17:14.640 --> 00:17:17.059 Yeah. But, you know, there is very strong counterargument 00:17:17.059 --> 00:17:19.460 to Doudney's critique. Technology writer Roger 00:17:19.460 --> 00:17:22.000 Bridgman defended neural networks by drawing 00:17:22.000 --> 00:17:24.119 a hard line between science and engineering. 00:17:24.500 --> 00:17:26.799 Bridgman essentially argued that it doesn't matter 00:17:26.799 --> 00:17:29.000 if the giant table of numbers is scientifically 00:17:29.000 --> 00:17:31.720 useless. If that unreadable table of numbers 00:17:31.720 --> 00:17:34.240 can consistently and safely steer a car down 00:17:34.240 --> 00:17:37.579 a highway, it is a monumental engineering triumph. 00:17:37.839 --> 00:17:40.559 Exactly. The utility of the machine outweighs 00:17:40.559 --> 00:17:42.599 the scientific headache of not understanding 00:17:42.599 --> 00:17:44.940 its inner workings. We don't know how it works, 00:17:44.960 --> 00:17:48.220 but it drives the car. And engineers wholly embraced 00:17:48.220 --> 00:17:50.500 this philosophy. Rather than giving up because 00:17:50.500 --> 00:17:53.380 the networks were opaque, they decided the theoretical 00:17:53.380 --> 00:17:55.940 math in the 1980s was fundamentally sound. They 00:17:55.940 --> 00:17:58.359 just lacked the raw computing power to fully 00:17:58.359 --> 00:18:01.140 unleash it. which ushers in the modern era, the 00:18:01.140 --> 00:18:03.779 hardware renaissance. Because the reason we are 00:18:03.779 --> 00:18:06.359 suddenly surrounded by AI today isn't because 00:18:06.359 --> 00:18:09.039 someone invented a magical new math formula recently. 00:18:09.599 --> 00:18:11.880 It's because engineers finally threw enough raw 00:18:11.880 --> 00:18:14.299 computing power at these ancient algorithms. 00:18:14.460 --> 00:18:17.599 Yeah, between 1991 and 2015, computing power 00:18:17.599 --> 00:18:20.680 increased roughly a million fold. And the primary 00:18:20.680 --> 00:18:23.799 driver of this explosion was the GPU, the graphics 00:18:23.799 --> 00:18:26.630 processing unit. the chips specifically designed 00:18:26.630 --> 00:18:29.130 to render high -end video games. The synergy 00:18:29.130 --> 00:18:32.250 is fascinating. Traditional CPUs are incredibly 00:18:32.250 --> 00:18:34.390 smart, designed to handle complex sequential 00:18:34.390 --> 00:18:37.390 tasks one at a time. GPUs, on the other hand, 00:18:37.690 --> 00:18:40.470 are designed to do thousands of very simple math 00:18:40.470 --> 00:18:42.869 operations simultaneously. Like rendering millions 00:18:42.869 --> 00:18:45.980 of individual pixels on a screen. Exactly, and 00:18:45.980 --> 00:18:48.599 it turns out calculating the weights of thousands 00:18:48.599 --> 00:18:51.380 of interconnected nodes in a neural network is 00:18:51.380 --> 00:18:55.220 the exact same type of massive parallel mathematical 00:18:55.220 --> 00:18:58.559 operation. Harnessing GPU power allowed networks 00:18:58.559 --> 00:19:01.119 to finally become deep. Deep learning just means 00:19:01.119 --> 00:19:03.599 you have the computing muscle to run dozens or 00:19:03.599 --> 00:19:06.700 even hundreds of hidden layers. Right. The network 00:19:06.700 --> 00:19:09.420 can suddenly learn highly abstract hierarchical 00:19:09.420 --> 00:19:12.339 representations of data and the milestones started 00:19:12.339 --> 00:19:15.920 falling rapidly. In 2012, a deep neural network 00:19:15.920 --> 00:19:19.140 called AlexNet entered a massive global image 00:19:19.140 --> 00:19:22.279 recognition competition and absolutely obliterated 00:19:22.279 --> 00:19:24.619 the traditional machine learning models. That 00:19:24.619 --> 00:19:27.240 victory proved the deep learning concept. But 00:19:27.240 --> 00:19:29.160 the true watershed moment for the world we live 00:19:29.160 --> 00:19:31.700 in right now came in 2017 with the publication 00:19:31.700 --> 00:19:34.019 of a landmark paper titled Attention Is All You 00:19:34.019 --> 00:19:36.759 Need. Oh, man. This paper introduced the transformer 00:19:36.759 --> 00:19:38.460 architecture. And this is the technology that 00:19:38.460 --> 00:19:40.460 powers large language models. This is the T in 00:19:40.460 --> 00:19:43.000 chat GPT. It is. So how does a transformer act? 00:19:42.859 --> 00:19:45.940 actually work? Well, prior to transformers, networks 00:19:45.940 --> 00:19:49.779 process text stodily, word by word. The transformer 00:19:49.779 --> 00:19:52.779 introduced attention mechanisms. It allows the 00:19:52.779 --> 00:19:55.299 network to process an entire sentence at once 00:19:55.299 --> 00:19:57.980 and weigh the contextual importance of every 00:19:57.980 --> 00:20:00.440 word against every other word. Can you give an 00:20:00.440 --> 00:20:03.180 example? Sure. For example, if a sentence says, 00:20:03.299 --> 00:20:06.099 the bank of the river, the attention mechanism 00:20:06.099 --> 00:20:08.579 mathematically forces the network to associate 00:20:08.579 --> 00:20:12.009 the word bank heavily with River, understanding 00:20:12.009 --> 00:20:13.930 it means a shoreline rather than a financial 00:20:13.930 --> 00:20:15.750 institution. Well, that makes sense. Yeah, it 00:20:15.750 --> 00:20:18.450 models long -range dependencies and data incredibly 00:20:18.450 --> 00:20:21.369 efficiently. Armed with deep layers, attention 00:20:21.369 --> 00:20:24.950 mechanisms, and endless GPU power, the real -world 00:20:24.950 --> 00:20:28.069 applications have absolutely exploded. We have 00:20:28.069 --> 00:20:30.710 generative AI models trained on hundreds of millions 00:20:30.710 --> 00:20:33.289 of image text pairs that can create original 00:20:33.289 --> 00:20:35.490 art from a simple prompt. It's amazing. We have 00:20:35.490 --> 00:20:38.109 networks diagnosing colorectal cancer by distinguishing 00:20:38.109 --> 00:20:40.869 highly invasive cancer cells based purely on 00:20:40.869 --> 00:20:42.890 subtle cellular shapes that human eyes miss. 00:20:43.210 --> 00:20:44.950 In the realm of biology, they are predicting 00:20:44.950 --> 00:20:47.250 three -dimensional protein structures, which 00:20:47.250 --> 00:20:50.329 is revolutionizing drug discovery. In finance, 00:20:50.569 --> 00:20:52.890 they're forecasting market trends and automating 00:20:52.890 --> 00:20:55.480 credit scoring. They filter out malicious spam. 00:20:55.680 --> 00:20:58.180 They translate obscure languages in real time. 00:20:58.400 --> 00:21:00.880 They detect cybersecurity inclusions. And they 00:21:00.880 --> 00:21:03.359 remain the foundational brains behind the push 00:21:03.359 --> 00:21:06.299 for autonomous vehicles. So as a learner trying 00:21:06.299 --> 00:21:08.519 to make sense of this rapidly changing world, 00:21:09.119 --> 00:21:11.579 what is the ultimate takeaway? It is the realization 00:21:11.579 --> 00:21:14.440 that the ancient math gauss used to track planets 00:21:14.440 --> 00:21:17.980 in the 1700s, running on modern video game hardware, 00:21:18.539 --> 00:21:21.059 is now actively woven into the very fabric of 00:21:21.059 --> 00:21:23.420 your life. It really is. It is deciding your 00:21:23.420 --> 00:21:26.099 credit worthiness, filtering your daily communications, 00:21:26.680 --> 00:21:28.940 and quite possibly diagnosing your medical scans. 00:21:29.319 --> 00:21:31.859 We have achieved a profound integration of technology 00:21:31.859 --> 00:21:34.799 into daily life. driven entirely by systems that 00:21:34.799 --> 00:21:37.539 we engineer, but still do not completely comprehend. 00:21:37.859 --> 00:21:39.380 It has been an unbelievable journey. I mean, 00:21:39.440 --> 00:21:41.400 we started with predicting planetary motion, 00:21:41.779 --> 00:21:43.880 moved to Navy -funded perceptrons that were declared 00:21:43.880 --> 00:21:46.480 completely dead in the 1970s. Survived the AI 00:21:46.480 --> 00:21:50.039 winter. Survived the AI winter, utilized the 00:21:50.039 --> 00:21:52.859 17th century calculus rule to teach machines 00:21:52.859 --> 00:21:55.400 how to learn from their mistakes, and rode a 00:21:55.400 --> 00:21:58.839 wave of GPU gaming chips to arrive at the generative 00:21:58.839 --> 00:22:01.880 AI explosion of today. But you know there is 00:22:01.880 --> 00:22:04.559 a final concept in the documentation that perfectly 00:22:04.559 --> 00:22:06.819 encapsulates the paradox we're currently living 00:22:06.819 --> 00:22:09.460 in. The research highlights the emergence of 00:22:09.460 --> 00:22:12.259 a relatively new field called topological deep 00:22:12.259 --> 00:22:15.180 learning, along with massive academic efforts 00:22:15.180 --> 00:22:18.079 trying to map the biology of large language models, 00:22:18.200 --> 00:22:21.220 the biology of a math model. Yeah. Because these 00:22:21.220 --> 00:22:23.980 networks are opaque black boxes, researchers 00:22:23.980 --> 00:22:26.480 are now having to invent entirely new fields 00:22:26.480 --> 00:22:29.259 of science just to visualize and decompose how 00:22:29.259 --> 00:22:31.539 these circuits reach their decisions. Wait, so 00:22:31.539 --> 00:22:33.559 we built the machine, but we don't understand 00:22:33.559 --> 00:22:35.480 it. So now we have to invent a completely new 00:22:35.480 --> 00:22:38.599 science just to observe. own creation. Yes, it 00:22:38.599 --> 00:22:41.240 leads to a deeply provocative question. If we 00:22:41.240 --> 00:22:43.480 are reaching a point where debugging a computer 00:22:43.480 --> 00:22:46.079 program looks less like traditional computer 00:22:46.079 --> 00:22:48.859 science and more like human psychology and neuroscience, 00:22:49.380 --> 00:22:52.740 if we have to actively psychoanalyze our own 00:22:52.740 --> 00:22:56.299 code to uncover its hidden associations and structural 00:22:56.299 --> 00:22:59.799 biases, Who is really teaching whom at this point? 00:23:00.059 --> 00:23:02.539 Are we the teacher or are we the subject? Exactly. 00:23:02.740 --> 00:23:04.460 That is a phenomenal thought to leave you with. 00:23:04.619 --> 00:23:06.839 Thank you for joining us on this deep dive. Our 00:23:06.839 --> 00:23:09.339 goal is always to help you question the invisible 00:23:09.339 --> 00:23:11.619 systems around you. So the next time you ask 00:23:11.619 --> 00:23:14.319 an AI a question or marvel at a piece of generated 00:23:14.319 --> 00:23:17.220 art, just remember, you aren't talking to a glowing 00:23:17.220 --> 00:23:19.339 brain. Not at all. You're talking to a giant 00:23:19.339 --> 00:23:21.680 200 -year -old mathematical game of trial and 00:23:21.680 --> 00:23:23.880 error that is still desperately trying to perfect 00:23:23.880 --> 00:23:26.420 its own recipe. We will catch you next time.