WEBVTT 00:00:00.000 --> 00:00:02.839 It used to be that when you picked up a tool, 00:00:03.520 --> 00:00:05.719 you understood exactly how it worked just by 00:00:05.719 --> 00:00:07.940 looking at it. Yeah, the mechanics were just 00:00:07.940 --> 00:00:10.320 obvious. Exactly. You pick up a hammer, you look 00:00:10.320 --> 00:00:11.919 at the heavy metal head, you look at the wooden 00:00:11.919 --> 00:00:14.699 handle, and the physics of the tool just make 00:00:14.699 --> 00:00:17.420 perfect intuitive sense to your brain. You can 00:00:17.420 --> 00:00:20.519 see the cause and effect. Right. And even with 00:00:20.519 --> 00:00:23.420 early computers, I mean, as unimaginably complex 00:00:23.420 --> 00:00:27.280 as they were, they were still just incredibly 00:00:27.280 --> 00:00:30.839 fast rule followers. You type a command and the 00:00:30.839 --> 00:00:33.479 machine executes it exactly as written the mechanism 00:00:33.479 --> 00:00:35.820 was entirely visible to the programmer Yeah, 00:00:35.920 --> 00:00:38.560 but today you pull out your smartphone You open 00:00:38.560 --> 00:00:41.399 an app and suddenly you're interacting with a 00:00:41.399 --> 00:00:44.219 tool where that mechanism is completely invisible 00:00:44.219 --> 00:00:47.820 It's hidden behind glass and code it it almost 00:00:47.820 --> 00:00:49.880 acts like it has a mind of its own Well, we've 00:00:49.880 --> 00:00:52.140 transitioned from tools that simply follow our 00:00:52.140 --> 00:00:54.700 explicit instructions to tools that actively 00:00:54.700 --> 00:00:57.359 anticipate our needs. And sometimes they do that 00:00:57.359 --> 00:00:59.740 before we even consciously register those needs 00:00:59.740 --> 00:01:02.079 ourselves. Which is a little unsettling. It is. 00:01:02.340 --> 00:01:04.819 That loss of visibility fundamentally changes 00:01:04.819 --> 00:01:07.000 our relationship with technology. I mean, we're 00:01:07.000 --> 00:01:09.319 no longer just operators. We are participants 00:01:09.319 --> 00:01:11.760 in a system we can't fully see. Welcome to this 00:01:11.760 --> 00:01:15.500 deep dive. Today we are looking at a really comprehensive 00:01:15.500 --> 00:01:18.620 foundational source, a highly detailed Wikipedia 00:01:18.620 --> 00:01:21.280 article on machine learning. We're trying to 00:01:21.280 --> 00:01:23.400 understand how these invisible tools actually 00:01:23.400 --> 00:01:25.879 work. And we are focusing on how we lost that 00:01:25.879 --> 00:01:28.319 visibility into our own technology. Right. The 00:01:28.319 --> 00:01:30.420 mission today is to cut through all the relentless 00:01:30.420 --> 00:01:33.620 modern hype and buzzwords. How did we go from 00:01:33.620 --> 00:01:36.939 machines that just followed our rigid rules to 00:01:36.939 --> 00:01:39.439 these opaque black boxes that learn on their 00:01:39.439 --> 00:01:42.560 own? And, you know, what happens when those mathematically 00:01:42.560 --> 00:01:46.040 perfect black boxes collide with our incredibly 00:01:46.040 --> 00:01:48.900 messy, biased human reality? The Culloden that's 00:01:48.900 --> 00:01:50.819 happening everywhere right now. It really is. 00:01:51.260 --> 00:01:53.799 So, whether you are actively working in tech 00:01:53.799 --> 00:01:56.359 or you're just trying to navigate a world that 00:01:56.359 --> 00:01:59.900 is increasingly run by algorithms, this deep 00:01:59.900 --> 00:02:01.980 dive is going to give you the ultimate foundational 00:02:01.980 --> 00:02:06.099 map. So, okay, let's unpack this. To really understand 00:02:06.099 --> 00:02:08.300 these sophisticated algorithms running our lives 00:02:08.300 --> 00:02:11.719 right now, We have to start by tracing a massive 00:02:11.719 --> 00:02:14.960 historical paradigm shift. The shift from programming 00:02:14.960 --> 00:02:17.979 a machine to teaching a machine. Exactly. And 00:02:17.979 --> 00:02:20.180 that shift actually started much earlier than 00:02:20.180 --> 00:02:22.599 most people realize. I mean, the term machine 00:02:22.599 --> 00:02:24.800 learning wasn't coined in the last decade during 00:02:24.800 --> 00:02:27.240 this current AI boom. It goes all the way back 00:02:27.240 --> 00:02:31.710 to 1959. Wait, 1959? Yeah. An IBM researcher 00:02:31.710 --> 00:02:34.150 named Arthur Samuel coined the term while he 00:02:34.150 --> 00:02:36.310 was working on a checkers program. A checkers 00:02:36.310 --> 00:02:39.169 program? Right. He realized that manually coding 00:02:39.169 --> 00:02:41.650 every possible board state and every possible 00:02:41.650 --> 00:02:44.150 strategy into a computer was just computationally 00:02:44.150 --> 00:02:47.569 impossible. There were just too many permutations. 00:02:47.789 --> 00:02:50.229 So what did he do? Well, he built a program that 00:02:50.229 --> 00:02:52.810 could evaluate the board and mathematically adjust 00:02:52.810 --> 00:02:55.030 its own internal parameters based on whether 00:02:55.030 --> 00:02:58.909 it won or lost a game. It is wild to think. that 00:02:58.909 --> 00:03:02.729 the origins of the systems currently driving 00:03:02.729 --> 00:03:04.810 our cars on the highway and diagnosing complex 00:03:04.810 --> 00:03:07.629 diseases started with a guy trying to get a computer 00:03:07.629 --> 00:03:09.969 to calculate winning chances in a board game. 00:03:10.129 --> 00:03:12.770 It's a very humble beginning. And the evolution 00:03:12.770 --> 00:03:15.270 from there is just fascinating. Our source points 00:03:15.270 --> 00:03:18.349 out that by the early 1960s, the Raytheon company 00:03:18.349 --> 00:03:20.729 had built an experimental learning machine called 00:03:20.729 --> 00:03:23.789 Cybertron. Cybertron, yeah. That sounds incredibly 00:03:23.789 --> 00:03:26.090 sci -fi. It really does. It analyzed complex 00:03:26.090 --> 00:03:30.250 data like sonar signals, electrocardiograms, 00:03:30.270 --> 00:03:32.509 and speech patterns using punch tape memory. 00:03:33.069 --> 00:03:35.289 But the absolute best detail here is that the 00:03:35.289 --> 00:03:37.810 machine was equipped with a literal physical 00:03:37.810 --> 00:03:40.129 goof button. I love the goof button. A human 00:03:40.129 --> 00:03:42.189 operator would literally sit at the console and 00:03:42.189 --> 00:03:44.169 act as the teacher. Just waiting for it to mess 00:03:44.169 --> 00:03:47.090 up. Exactly. Whenever the machine made an incorrect 00:03:47.090 --> 00:03:49.830 decision in its pattern recognition, the human 00:03:49.830 --> 00:03:52.250 operator would physically slam the goof button. 00:03:52.830 --> 00:03:55.449 Hitting that button forced the machine to mathematically 00:03:55.449 --> 00:03:58.430 penalize the logic path it had just used, so 00:03:58.430 --> 00:04:01.129 it had to reevaluate its approach. That is amazing. 00:04:01.270 --> 00:04:04.849 It's a highly rudimentary, tactile form of reinforcement 00:04:04.849 --> 00:04:06.969 learning. Yeah, instead of giving the computer 00:04:06.969 --> 00:04:09.990 a strict step -by -step logical flowchart, which 00:04:09.990 --> 00:04:13.270 is how traditional programming works. The human 00:04:13.270 --> 00:04:15.689 operator was just evaluating the final output. 00:04:16.329 --> 00:04:18.790 If the machine's output was wrong, you hit the 00:04:18.790 --> 00:04:21.430 button. It's like traditional programming is 00:04:21.430 --> 00:04:23.829 like giving a computer a strict recipe to bake 00:04:23.829 --> 00:04:26.230 a cake. Right. But machine learning is like feeding 00:04:26.230 --> 00:04:29.149 the computer 100 different cakes and asking it 00:04:29.149 --> 00:04:32.009 to figure out the recipe itself, with the human 00:04:32.009 --> 00:04:33.870 occasionally slamming the goof button when it 00:04:33.870 --> 00:04:35.889 bakes something awful. That's a great way to 00:04:35.889 --> 00:04:38.209 look at it. And what's fascinating here is the 00:04:38.209 --> 00:04:40.829 underlying philosophy of this shift, which really 00:04:40.829 --> 00:04:42.829 began to accelerate and take over the industry 00:04:42.829 --> 00:04:45.509 in the 1990s. The shift away from logic -based 00:04:45.509 --> 00:04:48.870 AI. Exactly. For a long time, artificial intelligence 00:04:48.870 --> 00:04:52.269 was dominated by a symbolic, logic -based approach. 00:04:52.689 --> 00:04:54.670 Researchers trying to program human knowledge 00:04:54.670 --> 00:04:57.610 and rigid cognitive rules into the machine. Like 00:04:57.610 --> 00:05:00.399 endless lists of ups and statements. Yeah, but 00:05:00.399 --> 00:05:03.120 the real world is incredibly complex and noisy. 00:05:03.600 --> 00:05:06.399 You simply cannot write enough if -then statements 00:05:06.399 --> 00:05:09.379 to cover, say, every possible lighting condition 00:05:09.379 --> 00:05:11.220 a self -driving car might encounter on a rainy 00:05:11.220 --> 00:05:13.800 night. The physical world is just too messy for 00:05:13.800 --> 00:05:15.779 a perfect set of written rules. So the field 00:05:15.779 --> 00:05:18.579 pivoted. They embraced a concept proposed much 00:05:18.579 --> 00:05:21.639 earlier by the mathematician Alan Turing. Right, 00:05:21.639 --> 00:05:24.279 the Turing test guy. Him, yeah. He suggested 00:05:24.279 --> 00:05:26.860 we stop asking the philosophical, unanswerable 00:05:26.860 --> 00:05:29.839 question, can machines think? and replace it 00:05:29.839 --> 00:05:32.319 with a pragmatic, testable question, which is, 00:05:32.540 --> 00:05:35.160 can machines do what we, as thinking entities, 00:05:35.279 --> 00:05:38.560 can do? Just focus on the output. Yes. The focus 00:05:38.560 --> 00:05:41.120 moved entirely away from trying to replicate 00:05:41.120 --> 00:05:43.899 human consciousness or cognitive thought. Instead, 00:05:44.019 --> 00:05:46.459 the goal became achieving practical, solvable 00:05:46.459 --> 00:05:49.259 tasks based purely on statistics, probability, 00:05:49.699 --> 00:05:51.899 and optimization. Okay, so if we are no longer 00:05:51.899 --> 00:05:54.560 writing rigid rules, how do we actually set up 00:05:54.560 --> 00:05:56.879 the classroom for the machine today? Like how 00:05:56.879 --> 00:05:58.879 do we feed it the cakes? Well modern machine 00:05:58.879 --> 00:06:01.360 learning generally falls into three main paradigms 00:06:01.360 --> 00:06:03.500 based on how the data is presented to the algorithm. 00:06:03.819 --> 00:06:05.879 Let's walk through those. The first is supervised 00:06:05.879 --> 00:06:08.459 learning, right? Yeah, supervised learning. This 00:06:08.459 --> 00:06:10.939 is where the machine is given a data set that 00:06:10.939 --> 00:06:13.899 has already been explicitly labeled by a human 00:06:13.899 --> 00:06:17.139 teacher. You provide the system with thousands, 00:06:17.379 --> 00:06:20.639 or maybe millions, of input -output pairs. Like 00:06:20.639 --> 00:06:23.360 classifying emails into a spam folder. Exactly. 00:06:23.480 --> 00:06:26.139 The algorithm's job is to calculate the mathematical 00:06:26.139 --> 00:06:28.839 function that best maps those specific inputs 00:06:28.839 --> 00:06:31.500 to the correct outputs. And it's used heavily 00:06:31.500 --> 00:06:33.779 for classification, but also for something called 00:06:33.779 --> 00:06:36.060 regression analysis. Regression is a bit different, 00:06:36.240 --> 00:06:38.699 right? Yeah, in regression the output isn't a 00:06:38.699 --> 00:06:41.920 category like spam or not spam. It's a continuous 00:06:41.920 --> 00:06:45.420 numerical value. For example, predicting future 00:06:45.420 --> 00:06:47.560 temperature fluctuations based on historical 00:06:47.560 --> 00:06:50.879 atmospheric data. Oh, okay. So the machine calculates 00:06:50.879 --> 00:06:53.040 the mathematical trajectory based on the labeled 00:06:53.040 --> 00:06:55.079 history. You got it. Then you have the second 00:06:55.079 --> 00:06:58.180 paradigm, which is unsupervised learning. This 00:06:58.180 --> 00:06:59.839 is where the training wheels completely come 00:06:59.839 --> 00:07:04.360 off. The machine is fed a massive pile of unlabeled 00:07:04.360 --> 00:07:07.079 raw data, and it is left completely on its own 00:07:07.079 --> 00:07:09.720 to find hidden structures or commonalities. Right, 00:07:09.879 --> 00:07:12.480 no labels to guide it at all. So the algorithm 00:07:12.480 --> 00:07:15.759 relies on mathematical techniques like dimensionality 00:07:15.759 --> 00:07:18.540 reduction. Dimensionality reduction. That sounds 00:07:18.540 --> 00:07:21.519 intense. It's just a way of dealing with massive 00:07:21.519 --> 00:07:24.699 datasets. It works by mapping data points in 00:07:24.699 --> 00:07:26.839 a high -dimensional space to find which variables 00:07:26.839 --> 00:07:29.879 correlate the most. It essentially merges redundant 00:07:29.879 --> 00:07:32.759 features. So if you have a dataset with a thousand 00:07:32.759 --> 00:07:35.720 variables... The machine figures out which handful 00:07:35.720 --> 00:07:38.699 of variables actually drive the variance. It 00:07:38.699 --> 00:07:41.339 strips away the noise so it can process the true 00:07:41.339 --> 00:07:44.160 signal efficiently without needing a human to 00:07:44.160 --> 00:07:46.040 tell it what to look for. Okay, but here's where 00:07:46.040 --> 00:07:48.740 it gets really interesting. Because if I dump 00:07:48.740 --> 00:07:51.199 a million unlabeled data points into an algorithm, 00:07:51.560 --> 00:07:53.560 and it groups them based on its own math, how 00:07:53.560 --> 00:07:55.740 does the machine know the pattern it found is 00:07:55.740 --> 00:07:58.000 actually useful? What do you mean? Well, what 00:07:58.000 --> 00:07:59.860 if it just groups all the photos in a database 00:07:59.860 --> 00:08:02.819 that happen to have exactly 42 red pixels in 00:08:02.819 --> 00:08:05.560 the top left corner? I mean, that is a mathematically 00:08:05.560 --> 00:08:08.319 valid pattern, but it's a completely random, 00:08:08.860 --> 00:08:11.269 meaningless coincidence in the real world. That 00:08:11.269 --> 00:08:14.430 is a very real, very crucial tension in the field, 00:08:14.649 --> 00:08:16.810 and it's exactly why we have to distinguish between 00:08:16.810 --> 00:08:19.069 data mining and machine learning. OK, what's 00:08:19.069 --> 00:08:21.769 the difference? Data mining is exploratory. It's 00:08:21.769 --> 00:08:24.290 looking for previously unknown properties or, 00:08:24.290 --> 00:08:27.029 as you said, random coincidences in a static 00:08:27.029 --> 00:08:30.550 database. Machine learning, however, is intensely 00:08:30.550 --> 00:08:33.009 focused on prediction. It has to actually work 00:08:33.009 --> 00:08:36.190 going forward. Yes. To ensure a pattern isn't 00:08:36.190 --> 00:08:38.929 just a meaningless coincidence like your 42 red 00:08:38.929 --> 00:08:41.929 pixels, we test the model through a concept called 00:08:41.929 --> 00:08:44.919 generalization. Generalization, so testing it 00:08:44.919 --> 00:08:48.000 against reality. Exactly. The true test of an 00:08:48.000 --> 00:08:50.320 algorithm isn't how well it memorizes its training 00:08:50.320 --> 00:08:53.179 data. I mean, memorization is computationally 00:08:53.179 --> 00:08:56.080 easy. Generalization is the machine's ability 00:08:56.080 --> 00:08:58.759 to apply the patterns it found to completely 00:08:58.759 --> 00:09:01.740 new, unseen examples. Oh, I see. So if its weird 00:09:01.740 --> 00:09:04.299 red pixel rule fails to predict anything useful 00:09:04.299 --> 00:09:06.639 when exposed to new photos. The model's internal 00:09:06.639 --> 00:09:09.440 math is penalized, and it adjusts to discard 00:09:09.440 --> 00:09:11.820 that useless rule. It has to survive outside 00:09:11.820 --> 00:09:14.360 the classroom. Which leads us organically to 00:09:14.360 --> 00:09:17.179 the third paradigm, reinforcement learning. Right. 00:09:17.419 --> 00:09:19.480 This is where the program interacts directly 00:09:19.480 --> 00:09:21.960 with a dynamic environment to achieve a goal. 00:09:22.480 --> 00:09:24.919 So it's not given a static data set to analyze 00:09:24.919 --> 00:09:27.659 at all? No. It's given a world state and a set 00:09:27.659 --> 00:09:31.139 of possible actions. As it takes actions, it 00:09:31.139 --> 00:09:33.259 receives feedback from the environment in the 00:09:33.259 --> 00:09:36.539 form of numerical rewards or punishments. Kind 00:09:36.539 --> 00:09:38.200 of like the goof button again, but automated. 00:09:38.409 --> 00:09:42.049 Very much so. Its sole objective is to maximize 00:09:42.049 --> 00:09:45.370 that cumulative reward over time. This is how 00:09:45.370 --> 00:09:48.029 DeepMind's AlphaGo system famously mastered the 00:09:48.029 --> 00:09:51.669 board game. Go back in 2016. It didn't just study 00:09:51.669 --> 00:09:53.950 historical human games, right? Right. It played 00:09:53.950 --> 00:09:56.210 millions of matches against itself, constantly 00:09:56.210 --> 00:09:58.490 tweaking its mathematical strategy to maximize 00:09:58.490 --> 00:10:01.029 the reward function of winning. Okay, so once 00:10:01.029 --> 00:10:03.960 the machine extracts these generalized patterns 00:10:03.960 --> 00:10:06.159 from the environment, whether that's supervised, 00:10:06.480 --> 00:10:09.700 unsupervised, or reinforcement learning, how 00:10:09.700 --> 00:10:12.000 is that knowledge physically structured inside 00:10:12.000 --> 00:10:13.980 the model? That's the big question. Because if 00:10:13.980 --> 00:10:16.600 it isn't a list of if -then rules like old programming, 00:10:17.120 --> 00:10:18.860 what does the resulting model actually look like 00:10:18.860 --> 00:10:21.399 under the hood? Well, the dominant architecture 00:10:21.399 --> 00:10:24.320 driving today's advancements is the artificial 00:10:24.320 --> 00:10:27.460 neural network. Right. These are loosely modeled 00:10:27.460 --> 00:10:29.940 on the biological structure of animal brains. 00:10:30.240 --> 00:10:32.440 So instead of traditional lines of sequential 00:10:32.440 --> 00:10:36.500 code, you have a vast web of nodes, which act 00:10:36.500 --> 00:10:39.259 as artificial neurons arranged in multiple layers. 00:10:39.379 --> 00:10:42.039 And these nodes are connected by edges. Yes, 00:10:42.139 --> 00:10:45.139 edges, which function like biological synapses. 00:10:45.340 --> 00:10:47.539 And the magic really happens in those connections, 00:10:47.620 --> 00:10:49.639 doesn't it? Every single edge has a specific 00:10:49.639 --> 00:10:52.360 mathematical way to assign to it. Exactly. As 00:10:52.360 --> 00:10:54.679 the machine learns, it uses a process called 00:10:54.679 --> 00:10:57.399 backpropagation. Backpropagation. Yeah. So when 00:10:57.399 --> 00:10:59.679 the network makes an error, the mathematical 00:10:59.679 --> 00:11:01.500 difference between its guess and the correct 00:11:01.500 --> 00:11:04.940 answer is calculated. That error signal is then 00:11:04.940 --> 00:11:07.220 propagated backward through the entire network. 00:11:07.399 --> 00:11:10.419 Like a ripple effect. Exactly. The system automatically 00:11:10.419 --> 00:11:12.259 adjusts the weights of all those connections, 00:11:12.759 --> 00:11:15.679 strengthening some and weakening others, to minimize 00:11:15.679 --> 00:11:18.360 the error next time. So the knowledge of the 00:11:18.360 --> 00:11:20.759 system doesn't live in any single line of code. 00:11:21.100 --> 00:11:23.879 It is distributed across millions or billions 00:11:23.879 --> 00:11:26.500 of these highly specific weighted probabilities. 00:11:26.940 --> 00:11:29.639 Wow. And our sources draw a parallel here that 00:11:29.639 --> 00:11:31.700 fundamentally changes how you look at artificial 00:11:31.700 --> 00:11:34.240 intelligence. It states that machine learning 00:11:34.240 --> 00:11:37.039 is deeply intimately tied to data compression. 00:11:37.440 --> 00:11:41.159 The AIXI theory. Yes. This theoretical framework 00:11:41.159 --> 00:11:43.299 suggests that the absolute best compression of 00:11:43.299 --> 00:11:46.419 a data file is simply the smallest possible software 00:11:46.419 --> 00:11:48.419 program that can mathematically generate that 00:11:48.419 --> 00:11:51.820 file. It is a brilliant conceptual lens. Think 00:11:51.820 --> 00:11:54.440 of a massive large language model developed by 00:11:54.440 --> 00:11:57.360 DeepMind called Chinchilla 7DB. What did Chinchilla 00:11:57.360 --> 00:12:00.320 do? Well, in research tests, this AI language 00:12:00.320 --> 00:12:03.519 model actually outperformed traditional dedicated 00:12:03.519 --> 00:12:06.039 lossless compression tools. Wait, a language 00:12:06.039 --> 00:12:09.159 model beat compression software? Yes. When fed 00:12:09.159 --> 00:12:12.799 raw image data, it compressed it down to 43 .4 00:12:12.799 --> 00:12:15.639 % of its original size, beating the standard 00:12:15.639 --> 00:12:18.600 PNG format. That's insane. And for audio, it 00:12:18.600 --> 00:12:21.899 compressed data down to 16 .4%. massively outperforming 00:12:21.899 --> 00:12:25.220 the FLAC format. So if I am grasping the magnitude 00:12:25.220 --> 00:12:28.460 of this, learning at its core is just the ultimate 00:12:28.460 --> 00:12:30.740 form of data compression. It really is. We tend 00:12:30.740 --> 00:12:33.460 to anthropomorphize AI. You know, we imagine 00:12:33.460 --> 00:12:35.879 this expansive, almost conscious brain growing 00:12:35.879 --> 00:12:39.200 inside a server rack, but maybe it's just squishing 00:12:39.200 --> 00:12:42.480 the absolute chaos of our reality into a highly 00:12:42.480 --> 00:12:45.820 efficient, mathematically predictable ZIP file. 00:12:45.919 --> 00:12:48.340 That's exactly what it's doing. It finds the 00:12:48.340 --> 00:12:51.440 shortest statistical path to represent a staggeringly 00:12:51.440 --> 00:12:53.620 complex world. But if we connect this to the 00:12:53.620 --> 00:12:57.100 bigger picture, that relentless drive to perfectly 00:12:57.100 --> 00:13:00.379 compress reality is exactly what causes one of 00:13:00.379 --> 00:13:02.600 the greatest technical failures. in machine learning, 00:13:03.059 --> 00:13:05.940 which is overfitting. Yes. Overfitting is when 00:13:05.940 --> 00:13:08.740 the compression goes too far. How so? Imagine 00:13:08.740 --> 00:13:11.340 a machine trying to draw a predictive line through 00:13:11.340 --> 00:13:14.320 a graph of scattered data points. If the model 00:13:14.320 --> 00:13:16.600 is too complex, it essentially gerrymanders its 00:13:16.600 --> 00:13:18.759 own mathematical theory. It twists itself into 00:13:18.759 --> 00:13:21.259 knots. Right. It draws this wildly complicated 00:13:21.259 --> 00:13:24.220 chaotic line that perfectly touches every single 00:13:24.220 --> 00:13:26.899 outlier and anomaly in the training data. It 00:13:26.899 --> 00:13:29.059 achieves perfect compression of the past. But 00:13:29.059 --> 00:13:32.230 the past has noise in it. Exactly. Because it 00:13:32.230 --> 00:13:35.230 has molded itself so rigidly to the specific 00:13:35.230 --> 00:13:38.389 noise of that data set, it utterly fails when 00:13:38.389 --> 00:13:41.929 deployed against new messy data. It has high 00:13:41.929 --> 00:13:45.009 variance, but zero predictive value. It's like 00:13:45.009 --> 00:13:48.909 a student who memorizes the exact answers to 00:13:48.909 --> 00:13:51.330 a practice test, but because they didn't actually 00:13:51.330 --> 00:13:53.690 learn the underlying concepts. they completely 00:13:53.690 --> 00:13:56.690 bomb the real exam. That is the perfect analogy. 00:13:57.029 --> 00:13:59.149 The model didn't learn the physics of the data. 00:13:59.370 --> 00:14:01.649 It just memorized the noise. And when we take 00:14:01.649 --> 00:14:03.789 these highly compressed pattern matchers, especially 00:14:03.789 --> 00:14:06.450 the overfitted ones, and deploy them into the 00:14:06.450 --> 00:14:08.950 unpredictable human environment, the collisions 00:14:08.950 --> 00:14:11.549 can be spectacular. Spectacular and sometimes 00:14:11.549 --> 00:14:14.720 very dangerous. Yeah. The sources highlight some 00:14:14.720 --> 00:14:17.580 massive high -profile failures that really pull 00:14:17.580 --> 00:14:19.720 back the curtain on the technology's limitations. 00:14:20.059 --> 00:14:22.279 Like, take the Netflix prize. Oh, that's a classic 00:14:22.279 --> 00:14:25.679 case study. Right. In 2006, Netflix offered a 00:14:25.679 --> 00:14:28.340 $1 million prize to anyone who could improve 00:14:28.340 --> 00:14:31.279 their cinematic recommendation algorithm by 10%. 00:14:31.279 --> 00:14:33.940 A million dollars just for a 10 % bump. Yeah. 00:14:34.299 --> 00:14:37.539 And a team finally won it in 2009, using this 00:14:37.539 --> 00:14:40.879 complex ensemble model. But Netflix quickly realized 00:14:40.879 --> 00:14:43.100 the model was completely useless for their actual 00:14:43.100 --> 00:14:45.399 business. Because the algorithm was mathematically 00:14:45.399 --> 00:14:49.360 sound, but the proxy data it trained on was fundamentally 00:14:49.360 --> 00:14:52.480 flawed. The human star ratings. Right. Human 00:14:52.480 --> 00:14:54.679 star ratings turned out to be terrible indicators 00:14:54.679 --> 00:14:58.200 of actual viewing behavior. Users would rate 00:14:58.200 --> 00:15:01.299 acclaimed documentaries five stars because, you 00:15:01.299 --> 00:15:02.919 know, they wanted to project an intellectual 00:15:02.919 --> 00:15:06.039 persona. But their actual behavior showed they 00:15:06.039 --> 00:15:08.679 spent their weekends binge watching reality television. 00:15:08.649 --> 00:15:12.549 Exactly. The machine optimized for the lie, not 00:15:12.549 --> 00:15:15.370 the reality. But the stakes are often much higher 00:15:15.370 --> 00:15:18.110 than movie recommendations, unfortunately. The 00:15:18.110 --> 00:15:20.409 source notes the 2018 incident where an Uber 00:15:20.409 --> 00:15:22.730 self -driving car failed to properly classify 00:15:22.730 --> 00:15:25.389 a pedestrian with a bicycle, which resulted in 00:15:25.389 --> 00:15:27.730 a fatal collision. A tragedy that shows how brittle 00:15:27.730 --> 00:15:31.129 these systems can be. It also details IBM Watson's 00:15:31.129 --> 00:15:34.110 failure to deliver reliable healthcare diagnostics 00:15:34.110 --> 00:15:37.370 despite billions of dollars invested. And it 00:15:37.370 --> 00:15:39.690 highlights language models like Microsoft's Bing 00:15:39.690 --> 00:15:43.309 Chat, producing hostile, erratic responses because 00:15:43.309 --> 00:15:45.110 they couldn't mathematically handle the edge 00:15:45.110 --> 00:15:48.549 cases of human conversation. Well, sometimes 00:15:48.549 --> 00:15:51.149 the machine just learns a functionally correct 00:15:51.149 --> 00:15:53.529 but utterly useless lesson. What do you mean 00:15:53.529 --> 00:15:55.970 by that? The source gives a great toy example. 00:15:56.389 --> 00:15:59.250 Imagine an image classifier trained entirely 00:15:59.250 --> 00:16:02.450 on pictures of brown horses and black cats. The 00:16:02.450 --> 00:16:04.970 network might optimize its weights to simply 00:16:04.970 --> 00:16:07.750 conclude that any brown patch of pixels in an 00:16:07.750 --> 00:16:10.070 image is a horse. Oh wow. So it didn't learn 00:16:10.070 --> 00:16:12.629 the shape or the legs or the context? Nope. It 00:16:12.629 --> 00:16:15.049 just found a lazy statistical shortcut. Brown 00:16:15.049 --> 00:16:17.570 equals horse. And because these are non -linear 00:16:17.570 --> 00:16:20.330 systems evaluating pixel relationships that human 00:16:20.330 --> 00:16:23.190 eyes don't even perceive, they have massive adversarial 00:16:23.190 --> 00:16:25.519 vulnerabilities. They really do. A researcher 00:16:25.519 --> 00:16:28.159 can change a single adversarially calculated 00:16:28.159 --> 00:16:30.740 pixel in an image, something completely invisible 00:16:30.740 --> 00:16:34.500 to you and me. Just one pixel. Just one. By adding 00:16:34.500 --> 00:16:37.820 a specific noise vector to the image, that tiny 00:16:37.820 --> 00:16:40.440 mathematical shift cascades through the network's 00:16:40.440 --> 00:16:43.200 weights, completely fooling the machine into 00:16:43.200 --> 00:16:45.980 misclassifying a stop sign as a speed limit sign. 00:16:46.220 --> 00:16:49.440 That is terrifying. But wait, if we know a machine 00:16:49.440 --> 00:16:52.100 learned the brown horse shortcut, or if we know 00:16:52.100 --> 00:16:54.620 it's being fooled by specific adversarial noise, 00:16:55.039 --> 00:16:57.940 how do developers fix it? Can't a software engineer 00:16:57.940 --> 00:17:00.100 just open up the code, find the line that says 00:17:00.100 --> 00:17:02.340 brown equals horse, and just delete it? And here 00:17:02.340 --> 00:17:05.000 we hit the terrifying core of modern artificial 00:17:05.000 --> 00:17:08.049 intelligence. The black box theory. The black 00:17:08.049 --> 00:17:10.390 box. The answer is no, they cannot just go in 00:17:10.390 --> 00:17:12.250 and tweak the rule. Because there are no explicit 00:17:12.250 --> 00:17:14.410 rules. As we discussed with neural networks, 00:17:14.569 --> 00:17:17.250 the logic is distributed across billions of minute 00:17:17.250 --> 00:17:19.089 weight adjustments. It's all just math. It is 00:17:19.089 --> 00:17:21.829 a vast matrix of floating -point numbers. It 00:17:21.829 --> 00:17:24.309 is so mathematically opaque that even the original 00:17:24.309 --> 00:17:26.849 coders who architected the system cannot audit 00:17:26.849 --> 00:17:29.410 the specific pattern the machine extracted. Wait, 00:17:29.410 --> 00:17:31.809 really? The people who built the system cannot 00:17:31.809 --> 00:17:34.509 trace its logic. They know the initial architecture 00:17:34.509 --> 00:17:37.769 and they know the data they fed it, but the specific 00:17:37.769 --> 00:17:41.490 pathway of logic for any given decision is an 00:17:41.490 --> 00:17:44.789 impenetrable black box. That is wild. And this 00:17:44.789 --> 00:17:46.910 opaqueness is driving a critical subfield right 00:17:46.910 --> 00:17:50.730 now called explainable AI or XAI. Because we 00:17:50.730 --> 00:17:53.150 urgently need systems where humans can actually 00:17:53.150 --> 00:17:55.690 decipher the prediction. Exactly. Especially 00:17:55.690 --> 00:17:58.990 in high stakes areas like medicine. Imagine a 00:17:58.990 --> 00:18:01.950 proprietary algorithm recommending unnecessary 00:18:02.220 --> 00:18:04.900 expensive medical tests, because the company 00:18:04.900 --> 00:18:07.960 that owns it trained the model on data that subtly 00:18:07.960 --> 00:18:10.880 rewards financial incentives. If the system is 00:18:10.880 --> 00:18:13.720 a black box, it's nearly impossible for regulators 00:18:13.720 --> 00:18:16.240 to prove that mathematical intent. Exactly. So 00:18:16.240 --> 00:18:19.000 if we can't audit the black box's logic, we are 00:18:19.000 --> 00:18:20.900 entirely at the mercy of the data we fed it. 00:18:21.150 --> 00:18:23.490 And historically, the data we've fed is incredibly 00:18:23.490 --> 00:18:26.630 flawed. Machine learning essentially digitizes 00:18:26.630 --> 00:18:29.490 and scales historical human behavior. It is a 00:18:29.490 --> 00:18:32.250 pure mathematical reflection. And a stark example 00:18:32.250 --> 00:18:35.529 of this occurred in 1988 at the UK's St. George's 00:18:35.529 --> 00:18:37.750 Medical School. What happened there? They implemented 00:18:37.750 --> 00:18:40.950 a computer program to screen initial applicants. 00:18:41.250 --> 00:18:43.549 And the program ended up automatically denying 00:18:43.549 --> 00:18:46.509 nearly 60 qualified candidates simply because 00:18:46.509 --> 00:18:49.029 they were women or because they had non -European 00:18:49.029 --> 00:18:51.319 sounding names. Wow. But the algorithm wasn't 00:18:51.319 --> 00:18:53.519 coded to be prejudiced. Yeah, it wasn't coded 00:18:53.519 --> 00:18:56.339 with malicious intent. It was trained on the 00:18:56.339 --> 00:18:58.960 historical behavior of the human admissions staff 00:18:58.960 --> 00:19:01.619 from previous years. It found the statistical 00:19:01.619 --> 00:19:04.160 pattern in the human decisions and ruthlessly 00:19:04.160 --> 00:19:06.900 codified it into a mathematical rule. We see 00:19:06.900 --> 00:19:09.319 this mechanical echo repeatedly. The investigative 00:19:09.319 --> 00:19:12.200 organization ProPublica found that risk assessment 00:19:12.200 --> 00:19:14.539 algorithms used in the criminal justice system 00:19:14.539 --> 00:19:17.519 falsely flagged black defendants as high risk 00:19:17.519 --> 00:19:20.549 for recidivism twice as often as white defendants, 00:19:20.950 --> 00:19:23.009 again based on historical arrest data. It just 00:19:23.009 --> 00:19:25.849 absorbs the history. In 2015, Google Photos' 00:19:26.049 --> 00:19:29.150 classifier tagged humans as gorillas due to massive 00:19:29.150 --> 00:19:31.210 gaps in its training data. And language models 00:19:31.210 --> 00:19:33.430 inherently absorb these claws, too, because human 00:19:33.430 --> 00:19:35.670 language corpora contain them. When Microsoft 00:19:35.670 --> 00:19:38.569 released its TeyChat bot on Twitter, it adopted 00:19:38.569 --> 00:19:41.029 inflammatory racist language almost immediately. 00:19:41.250 --> 00:19:43.309 Because it was learning from Twitter users. It 00:19:43.309 --> 00:19:47.259 simply ran the statistics on the text the crowd 00:19:47.259 --> 00:19:50.180 was feeding it, optimizing for engagement based 00:19:50.180 --> 00:19:53.559 on what it saw in real time. So the tech industry 00:19:53.559 --> 00:19:55.980 is currently attempting to address some of the 00:19:55.980 --> 00:19:58.380 data flow issues, particularly around privacy, 00:19:58.880 --> 00:20:01.660 by moving toward what is called embedded machine 00:20:01.660 --> 00:20:04.059 learning. Right, moving to the edge. Yeah. Instead 00:20:04.059 --> 00:20:06.420 of sending all your personal biometric or text 00:20:06.420 --> 00:20:09.440 data to a giant server in the cloud to be processed, 00:20:10.019 --> 00:20:12.259 models are being deployed directly onto edge 00:20:12.259 --> 00:20:15.000 devices, like your smartphone, your smartwatch, 00:20:15.339 --> 00:20:18.109 or local industrial sensors. But this requires 00:20:18.109 --> 00:20:20.890 specialized hardware. I mean, a standard CPU 00:20:20.890 --> 00:20:23.109 isn't efficient enough for the massive matrix 00:20:23.109 --> 00:20:25.490 multiplication required by neural networks. So 00:20:25.490 --> 00:20:27.809 what do they use? Devices are now equipped with 00:20:27.809 --> 00:20:31.250 tensor processing units, or TPUs, which are microchips 00:20:31.250 --> 00:20:33.869 designed specifically to run complex AI math 00:20:33.869 --> 00:20:36.430 locally, right there in your hand, without needing 00:20:36.430 --> 00:20:38.470 an internet connection. So what does this all 00:20:38.470 --> 00:20:41.460 mean? The push to the edge solves the latency 00:20:41.460 --> 00:20:44.380 issue, and it solves the massive privacy issue 00:20:44.380 --> 00:20:47.700 of broadcasting your life to a server farm. But 00:20:47.700 --> 00:20:50.099 if we step back and look at the fundamental mechanics, 00:20:50.500 --> 00:20:53.319 does it actually solve the bias? That's the million 00:20:53.319 --> 00:20:56.200 dollar question. Right. If an algorithm requires 00:20:56.200 --> 00:20:59.000 vast amounts of human data to learn, and that 00:20:59.000 --> 00:21:01.640 data is inherently skewed by history, aren't 00:21:01.640 --> 00:21:05.200 we just building incredibly fast, highly efficient 00:21:05.200 --> 00:21:08.099 digital mirrors of our own flaws? This raises 00:21:08.099 --> 00:21:09.880 an important question about the trade -offs of 00:21:09.880 --> 00:21:13.059 localized AI. A compounding factor in the data 00:21:13.059 --> 00:21:15.920 pipeline is the lack of diversity among the engineers 00:21:15.920 --> 00:21:18.420 curating it. Who is actually building these systems? 00:21:18.839 --> 00:21:22.240 Well, globally, only about 16 .1 % of faculty 00:21:22.240 --> 00:21:24.880 members focusing on AI are female, and the broader 00:21:24.880 --> 00:21:27.559 demographics of AI PhD graduates are similarly 00:21:27.559 --> 00:21:30.180 skewed. So the people defining the reward functions 00:21:30.180 --> 00:21:32.700 and selecting the training sets naturally represent 00:21:32.700 --> 00:21:36.000 a very narrow slice of humanity. Exactly. The 00:21:36.000 --> 00:21:38.539 hardware is getting exponentially better, but 00:21:38.539 --> 00:21:40.980 the baseline data remains a structural issue. 00:21:41.319 --> 00:21:45.279 The AI scientist Fei -Fei Lai summarized it perfectly 00:21:45.279 --> 00:21:47.680 when she noted, there's nothing artificial about 00:21:47.680 --> 00:21:50.319 AI. It's inspired by people, it's created by 00:21:50.319 --> 00:21:53.019 people, and it impacts people. That is a powerful 00:21:53.019 --> 00:21:55.920 quote. It is. Moving to embedded machine learning 00:21:55.920 --> 00:21:58.519 secures your data privacy because your personal 00:21:58.519 --> 00:22:01.680 data never leaves your device. But it also means 00:22:01.680 --> 00:22:04.220 that the mathematically opaque, biased black 00:22:04.220 --> 00:22:06.819 box we just discussed isn't living in a remote 00:22:06.819 --> 00:22:09.240 server anymore. It's living locally. It is now 00:22:09.240 --> 00:22:11.759 living right in your pocket. Wow. Wow, we have 00:22:11.759 --> 00:22:13.500 covered a massive amount of technical ground 00:22:13.500 --> 00:22:16.299 today. We traced the history from Arthur Samuel's 00:22:16.299 --> 00:22:19.240 checkers to the cybernetic goof button. We mapped 00:22:19.240 --> 00:22:22.559 the three learning paradigms, unpacked how backpropagation 00:22:22.559 --> 00:22:24.680 adjusts the microscopic weights of neural networks, 00:22:25.039 --> 00:22:27.559 and looked at how AI is essentially the ultimate 00:22:27.559 --> 00:22:30.079 zip file of reality. That's a lot to take in. 00:22:30.259 --> 00:22:32.019 And finally, we confronted what happens when 00:22:32.019 --> 00:22:34.619 those unauditable, overfitted models collide 00:22:34.619 --> 00:22:37.109 with our inherently flawed historical data. It 00:22:37.109 --> 00:22:39.609 truly is a profound shift in how humans interact 00:22:39.609 --> 00:22:43.049 with computation. We are moving from tools we 00:22:43.049 --> 00:22:46.130 explicitly instruct to tools we merely influence. 00:22:46.430 --> 00:22:48.890 I want to leave you, our listener, with a final 00:22:48.890 --> 00:22:51.690 scenario to mull over, building on the mechanics 00:22:51.690 --> 00:22:53.990 of embedded machine learning and the black box. 00:22:54.650 --> 00:22:57.369 Imagine a near future where an embedded machine 00:22:57.369 --> 00:23:00.230 learning model is running entirely locally on 00:23:00.230 --> 00:23:02.990 a life -critical wearable inside your body. Let's 00:23:02.990 --> 00:23:05.960 say a smart pacemaker. Okay. It uses unsupervised 00:23:05.960 --> 00:23:08.539 learning to continuously map your unique biological 00:23:08.539 --> 00:23:10.460 patterns. It monitors the minute variance in 00:23:10.460 --> 00:23:12.299 your heart rate, your stress biomarkers, your 00:23:12.299 --> 00:23:15.259 sleep cycles. Then one day it makes a split second 00:23:15.259 --> 00:23:17.700 life altering decision about how to regulate 00:23:17.700 --> 00:23:19.960 your heart during a severe medical crisis. A 00:23:19.960 --> 00:23:22.960 completely autonomous decision. Right. But because 00:23:22.960 --> 00:23:25.099 the neural network's logic is distributed across 00:23:25.099 --> 00:23:27.500 millions of unauditable mathematical weights, 00:23:28.119 --> 00:23:30.160 neither your doctors nor the original hardware 00:23:30.160 --> 00:23:33.039 developers can explain exactly why it made that 00:23:33.039 --> 00:23:36.380 specific choice. It saved you, or maybe it didn't, 00:23:36.480 --> 00:23:38.599 but no human can read the mathematical path it 00:23:38.599 --> 00:23:42.000 took. It is a very unsettling reality to consider. 00:23:42.500 --> 00:23:45.000 Who is ultimately responsible for the machine 00:23:45.000 --> 00:23:48.640 when it has become an un -auditable black box 00:23:48.640 --> 00:23:51.700 of you? Exactly. Pay attention to the tools you 00:23:51.700 --> 00:23:54.019 use today, the recommendations you get, the way 00:23:54.019 --> 00:23:56.660 your phone auto -completes your thoughts, the 00:23:56.660 --> 00:23:58.960 subtle nudges from your apps, how much of your 00:23:58.960 --> 00:24:01.440 own daily reality is already being shaped by 00:24:01.440 --> 00:24:03.819 these invisible, mathematically opaque patterns. 00:24:04.140 --> 00:24:06.460 It's something we all need to be aware of. We 00:24:06.460 --> 00:24:08.099 started by talking about the comfort of holding 00:24:08.099 --> 00:24:10.319 a hammer, seeing exactly how the physics of the 00:24:10.319 --> 00:24:12.440 tool work. We are now living in a world where 00:24:12.440 --> 00:24:14.759 the hammer decides which nail to hit, and the 00:24:14.759 --> 00:24:17.519 math won't tell us why. Thanks for joining us 00:24:17.519 --> 00:24:19.000 on this deep dive. Stay curious.