WEBVTT 00:00:00.000 --> 00:00:03.680 Imagine trying to teach a machine to drive a 00:00:03.680 --> 00:00:06.559 car. It seems pretty straightforward at first, 00:00:06.660 --> 00:00:08.800 right? You program a rule that says stop at a 00:00:08.800 --> 00:00:10.500 red light. Right. And you write another rule 00:00:10.500 --> 00:00:12.800 that says stay between the white lines. Exactly. 00:00:13.419 --> 00:00:16.000 But how do you program a rule that tells the 00:00:16.000 --> 00:00:18.859 car what to do if a clown on a unicycle juggling 00:00:18.859 --> 00:00:21.100 flaming torches suddenly crosses the street? 00:00:21.219 --> 00:00:24.480 Oh, man. Or what if a prankster shoves a banana 00:00:24.480 --> 00:00:27.609 into the car's tailpipe? Yes. In the world of 00:00:27.609 --> 00:00:29.530 artificial intelligence, this is actually known 00:00:29.530 --> 00:00:33.049 as the banana in the tailpipe problem. It sounds 00:00:33.049 --> 00:00:35.530 like a joke, but it's genuinely the central dilemma 00:00:35.530 --> 00:00:38.189 that has plagued computer scientists for, like, 00:00:38.469 --> 00:00:41.439 over 60 years. It really is. I mean... It is 00:00:41.439 --> 00:00:44.420 the ultimate bottleneck of pure logic. We're 00:00:44.420 --> 00:00:47.039 so used to seeing AI today, the kind that writes 00:00:47.039 --> 00:00:49.740 essays or generates photorealistic art. And we 00:00:49.740 --> 00:00:51.899 kind of forget that for decades, researchers 00:00:51.899 --> 00:00:54.320 tried to build intelligence in a completely different 00:00:54.320 --> 00:00:56.140 way. Right. They didn't want a mysterious black 00:00:56.140 --> 00:00:58.640 box that just guesses the next word. Exactly. 00:00:58.740 --> 00:01:01.060 They wanted a perfectly transparent, flawless 00:01:01.060 --> 00:01:04.040 logic machine. They basically wanted an electronic 00:01:04.040 --> 00:01:06.379 mathematician. Well, welcome to the Deep Dive. 00:01:06.540 --> 00:01:08.680 I'm your host, and today I'm joined by our resident 00:01:08.680 --> 00:01:11.760 expert to unpack a really comprehensive Wikipedia 00:01:11.760 --> 00:01:14.519 article on symbolic artificial intelligence. 00:01:15.379 --> 00:01:18.200 Our mission today is to explore how early AI 00:01:18.200 --> 00:01:20.780 tried to think exactly like a human reasoning 00:01:20.780 --> 00:01:25.540 through a math problem. Why that approach repeatedly 00:01:25.540 --> 00:01:28.079 crashed into periods of devastating funding loss, 00:01:28.540 --> 00:01:31.200 which researchers call AI winters. Yeah, and 00:01:31.200 --> 00:01:34.439 why this classic old school logic based approach 00:01:34.439 --> 00:01:36.900 might actually be the missing puzzle piece for 00:01:36.900 --> 00:01:39.040 the future of artificial general intelligence. 00:01:39.299 --> 00:01:41.120 The history of this field really gets to the 00:01:41.120 --> 00:01:44.250 core of how we define intelligence itself. Like, 00:01:44.370 --> 00:01:46.430 is intelligence just a neat set of mathematical 00:01:46.430 --> 00:01:49.590 rules, or is it a messy statistical guess based 00:01:49.590 --> 00:01:51.670 on billions of data points? And if you're listening 00:01:51.670 --> 00:01:53.670 to this right now, our favorite ever curious 00:01:53.670 --> 00:01:55.689 learner, I promise you that by the end of this 00:01:55.689 --> 00:01:57.650 conversation, you're going to understand the 00:01:57.650 --> 00:02:00.370 historical grudges, the multimillion dollar successes, 00:02:00.709 --> 00:02:03.329 and the profound philosophical battles that define 00:02:03.329 --> 00:02:06.530 how machines learn. Absolutely. No overwhelming 00:02:06.530 --> 00:02:09.229 jargon today, just the fascinating aha moments 00:02:09.229 --> 00:02:11.830 of AI's evolution. OK, let's unpack this. To 00:02:11.830 --> 00:02:14.830 really understand where AI is going, today. We 00:02:14.830 --> 00:02:17.289 first have to step back to the late 1940s and 00:02:17.289 --> 00:02:20.599 50s. Long before algorithms were just scraping 00:02:20.599 --> 00:02:23.020 the internet for data, researchers believed they 00:02:23.020 --> 00:02:26.340 could build a functioning brain using pure, readable 00:02:26.340 --> 00:02:29.580 logic. So this is the era of symbolic AI. Right. 00:02:29.639 --> 00:02:32.659 Also known as classical or logic -based AI. I 00:02:32.659 --> 00:02:34.780 see. The source material mentions these methods 00:02:34.780 --> 00:02:38.360 were based on high -level, human -readable representations 00:02:38.360 --> 00:02:41.979 of problems. It relies on things called production 00:02:41.979 --> 00:02:44.900 rules and semantic nets. But for someone who 00:02:44.900 --> 00:02:47.500 isn't a computer scientist, What does a symbolic 00:02:47.500 --> 00:02:50.400 AI actually look like under the hood? Well, think 00:02:50.400 --> 00:02:52.780 of it as a massive, intricate web of if -then 00:02:52.780 --> 00:02:54.860 statements. Like, if condition A is true, then 00:02:54.860 --> 00:02:57.580 execute action B. The beauty of this approach 00:02:57.580 --> 00:02:59.560 is its transparency. You could literally read 00:02:59.560 --> 00:03:01.879 the code and understand the machine's exact thought 00:03:01.879 --> 00:03:04.259 process. So there's no mystery. If it messes 00:03:04.259 --> 00:03:07.099 up, you know why. Exactly. If the AI made a mistake, 00:03:07.539 --> 00:03:09.699 a programmer could just go into the code, trace 00:03:09.699 --> 00:03:12.259 the logic step by step, and fix the specific 00:03:12.259 --> 00:03:15.000 rule that failed. It sounds like early AI was 00:03:15.000 --> 00:03:17.460 basically following a strict step -by -step baking 00:03:17.460 --> 00:03:20.860 recipe, whereas today's AI just tastes the soup 00:03:20.860 --> 00:03:23.400 and tries to blindly guess the ingredients based 00:03:23.400 --> 00:03:26.340 on pattern recognition. That is a perfect analogy. 00:03:26.840 --> 00:03:29.180 And what's fascinating here is the sheer love 00:03:29.180 --> 00:03:31.460 of the confidence those early researchers had 00:03:31.460 --> 00:03:34.280 in their recipe. During what we call the first 00:03:34.280 --> 00:03:37.699 AI summer, stretching roughly from the late 40s 00:03:37.699 --> 00:03:40.520 through the mid 60s, the exuberance was just 00:03:40.680 --> 00:03:42.759 off the charts. They thought they had it all 00:03:42.759 --> 00:03:45.000 figured out. They really did. They genuinely 00:03:45.000 --> 00:03:46.939 believed they would achieve artificial general 00:03:46.939 --> 00:03:49.379 intelligence in just a handful of years. Like 00:03:49.379 --> 00:03:51.879 a machine as smart as a human across the board. 00:03:52.159 --> 00:03:54.560 Looking at their early wins, you can almost forgive 00:03:54.560 --> 00:03:56.680 them for being so arrogant about it. The notes 00:03:56.680 --> 00:03:59.300 mention a robotic turtle built all the way back 00:03:59.300 --> 00:04:02.759 in 1948. Oh yeah. It used just seven vacuum tubes, 00:04:02.819 --> 00:04:04.639 yet it could successfully navigate a physical 00:04:04.639 --> 00:04:07.580 environment. And then in 1955, there was a program 00:04:07.580 --> 00:04:10.319 called the Logic Theorist. This thing successfully 00:04:10.319 --> 00:04:12.900 proved 38 mathematical theorems from Principia 00:04:12.900 --> 00:04:15.840 Mathematica, which is a notoriously dense foundational 00:04:15.840 --> 00:04:18.379 math text. But proving those theorems wasn't 00:04:18.379 --> 00:04:20.980 just a party trick. It required the invention 00:04:20.980 --> 00:04:23.350 of something incredibly important in computer 00:04:23.350 --> 00:04:26.509 science called heuristics. Okay, unpack that 00:04:26.509 --> 00:04:29.209 term for us. The source calls them rules of thumb, 00:04:29.569 --> 00:04:32.829 but how does a computer use a rule of thumb? 00:04:33.490 --> 00:04:36.009 Well, imagine you are dropped into the center 00:04:36.009 --> 00:04:39.490 of a massive physical maze and you need to find 00:04:39.490 --> 00:04:43.149 the exit. If you try to solve it by walking down 00:04:43.149 --> 00:04:46.889 every single possible path, checking every dead 00:04:46.889 --> 00:04:49.250 end until you find the right one. Which would 00:04:49.250 --> 00:04:51.269 take forever. Right. Well, computer scientists 00:04:51.269 --> 00:04:54.149 call an enumerative search. You will eventually 00:04:54.149 --> 00:04:56.649 find the exit. But as the maze gets bigger, the 00:04:56.649 --> 00:04:59.209 time it takes grows exponentially. A computer 00:04:59.209 --> 00:05:01.449 runs out of processing power almost immediately. 00:05:01.750 --> 00:05:04.449 So you can't just brute force a complex problem. 00:05:04.670 --> 00:05:07.490 Exactly. Researchers like Herbert Simon and Alan 00:05:07.490 --> 00:05:09.769 Newell realize they needed to give the computer 00:05:09.769 --> 00:05:12.889 a compass. A heuristic is that compass. It's 00:05:12.889 --> 00:05:15.410 a fast algorithm that guides a search in a promising 00:05:15.410 --> 00:05:17.860 direction. So it might occasionally guide you 00:05:17.860 --> 00:05:20.259 down a suboptimal path, but it cuts through all 00:05:20.259 --> 00:05:23.660 the noise. Yes. This led to breakthroughs like 00:05:23.660 --> 00:05:26.600 the ASTAR algorithm. Instead of checking every 00:05:26.600 --> 00:05:29.139 leaf in the forest, ASTAR constantly calculates 00:05:29.139 --> 00:05:31.560 the distance it has traveled against an estimate 00:05:31.560 --> 00:05:34.579 of the distance left to go. Oh, wow. Yeah, allowing 00:05:34.579 --> 00:05:36.639 the computer to find a guaranteed solution without 00:05:36.639 --> 00:05:39.389 checking every single dead end. Even within this 00:05:39.389 --> 00:05:42.730 highly structured, logic -based world, the source 00:05:42.730 --> 00:05:45.529 material points out a massive philosophical split 00:05:45.529 --> 00:05:48.550 right at the beginning. The researchers actually 00:05:48.550 --> 00:05:51.170 divided themselves into two camps, the Neats 00:05:51.170 --> 00:05:53.569 and the Scruffys. Oh, the Neats and the Scruffys. 00:05:53.689 --> 00:05:56.129 The Neats were championed by figures like John 00:05:56.129 --> 00:05:59.240 McCarthy at Stanford. They believed that AI should 00:05:59.240 --> 00:06:02.079 be built on neat, formal mathematical logic. 00:06:02.160 --> 00:06:05.000 Right. To them, it didn't matter if the machine's 00:06:05.000 --> 00:06:07.379 process perfectly mimicked human psychology or 00:06:07.379 --> 00:06:09.660 biology. They just wanted to capture the pure 00:06:09.660 --> 00:06:12.319 essence of abstract reasoning. And the Scruffys 00:06:12.319 --> 00:06:15.139 were the rebels at MIT. Led by people like Marvin 00:06:15.139 --> 00:06:18.220 Minsky, yeah. The Scruffys argued that human 00:06:18.220 --> 00:06:20.870 intelligence is inherently messy. I mean, we 00:06:20.870 --> 00:06:22.610 don't walk through the world solving mathematical 00:06:22.610 --> 00:06:24.829 proofs to decide when to cross the street. No, 00:06:24.949 --> 00:06:27.629 we just kind of look and go. Exactly. We rely 00:06:27.629 --> 00:06:31.329 on ad hoc, hand -built, scruffy knowledge to 00:06:31.329 --> 00:06:34.350 understand complex things like vision or natural 00:06:34.350 --> 00:06:37.329 language. You simply cannot solve the real world 00:06:37.329 --> 00:06:39.980 with a neat logical formula. The scruffies were 00:06:39.980 --> 00:06:43.300 incredibly prescient because that strict, neat 00:06:43.300 --> 00:06:46.759 recipe approach hit a catastrophic brick wall 00:06:46.759 --> 00:06:48.779 the second it stepped out of the laboratory. 00:06:49.120 --> 00:06:51.319 The real world is relentlessly unpredictable. 00:06:51.720 --> 00:06:54.420 And that unpredictability ushered in the first 00:06:54.420 --> 00:06:58.420 A .I. winter in the late 1960s and 70s. The funding 00:06:58.420 --> 00:07:00.740 just completely evaporated. The promises made 00:07:00.740 --> 00:07:03.339 during that first summer were wild. The U .S. 00:07:03.620 --> 00:07:05.920 military's research arm, DARPA, poured millions 00:07:05.920 --> 00:07:08.439 into projects expecting like autonomous tech 00:07:08.240 --> 00:07:10.699 that could navigate battlefields. And a flawless 00:07:10.699 --> 00:07:13.300 Russian to English translation for Cold War intelligence 00:07:13.300 --> 00:07:15.660 operations. But translating a language isn't 00:07:15.660 --> 00:07:17.699 just about swapping out one symbol for another 00:07:17.699 --> 00:07:20.259 in a dictionary. It requires cultural context 00:07:20.259 --> 00:07:22.899 and nuance, which pure logic just couldn't grasp. 00:07:23.040 --> 00:07:26.300 And the backlash was severe. In the UK, a mathematician 00:07:26.300 --> 00:07:29.000 named Sir James Lighthill wrote a devastating 00:07:29.000 --> 00:07:31.560 report for Parliament. He essentially called 00:07:31.560 --> 00:07:34.879 AI researchers charlatans. That's what it was. 00:07:35.379 --> 00:07:37.759 Lighthill pointed out. that while these logic 00:07:37.759 --> 00:07:40.899 systems work beautifully on controlled toy problems 00:07:40.899 --> 00:07:43.420 in the lab, they would never scale up to the 00:07:43.420 --> 00:07:45.759 real world because of combinatorial explosion. 00:07:46.180 --> 00:07:48.759 Let's visualize that for the listener. Combinatorial 00:07:48.759 --> 00:07:51.040 explosion is kind of like the old riddle of putting 00:07:51.040 --> 00:07:53.199 a grain of rice on a chessboard and doubling 00:07:53.199 --> 00:07:55.980 it on every square. Right. By the time an AI 00:07:55.980 --> 00:07:58.540 looks three moves ahead in a real -world scenario, 00:07:58.970 --> 00:08:02.029 It has 10 options. By move 4, it has 1 ,000. 00:08:02.230 --> 00:08:04.529 In seconds, the computer's entirely paralyzed, 00:08:04.730 --> 00:08:06.769 trying to calculate more branching possibilities 00:08:06.769 --> 00:08:09.029 than there are atoms in the universe. The sheer 00:08:09.029 --> 00:08:12.009 volume of possibilities just crushes the if -then 00:08:12.009 --> 00:08:15.529 rules. But the field didn't die completely. It 00:08:15.529 --> 00:08:18.430 pivoted. And this ushered in the second AI summer 00:08:18.430 --> 00:08:21.029 in the late 70s and 80s. And this was driven 00:08:21.029 --> 00:08:23.009 by a new realization from a researcher named 00:08:23.009 --> 00:08:26.230 Ed Feigenbaum, right? He coined the phrase, in 00:08:26.230 --> 00:08:28.829 the knowledge lies the power. meaning general 00:08:28.829 --> 00:08:31.470 logic wasn't enough and AI couldn't just be generally 00:08:31.470 --> 00:08:34.169 smart. It needed highly specific domain level 00:08:34.169 --> 00:08:36.470 knowledge. And this is where we see the boom 00:08:36.470 --> 00:08:38.690 of something called expert systems. I'm looking 00:08:38.690 --> 00:08:40.350 at the notes on this and it sounds like they 00:08:40.350 --> 00:08:43.129 were trying to like digitize the brains of actual 00:08:43.129 --> 00:08:46.409 human specialists. That was the exact goal. They 00:08:46.409 --> 00:08:49.710 would interview a world -class expert, say a 00:08:49.710 --> 00:08:52.840 chemist, and systematically extract all their 00:08:52.840 --> 00:08:55.720 knowledge, encoding it into thousands of highly 00:08:55.720 --> 00:08:58.080 specific rules. Right, like the Dendrel system. 00:08:58.179 --> 00:09:01.120 Yeah. Dendrel was designed to deduce the structure 00:09:01.120 --> 00:09:03.600 of organic molecules. The medical example here 00:09:03.600 --> 00:09:05.039 is what's blowing my mind, though. There was 00:09:05.039 --> 00:09:08.299 an expert system called MYCIN designed to diagnose 00:09:08.299 --> 00:09:11.879 blood infections like bacteremia. How did a doctor 00:09:11.879 --> 00:09:15.100 actually use that in the 1970s? So a doctor would 00:09:15.100 --> 00:09:17.419 sit at a computer terminal and the system would 00:09:17.419 --> 00:09:19.600 prompt them with questions. It would ask, is 00:09:19.600 --> 00:09:22.460 the patient's blood culture Positive. The doctor 00:09:22.460 --> 00:09:24.960 typed in the answer. Okay. Then my ask for the 00:09:24.960 --> 00:09:27.440 morphology of the organism. Behind the scenes, 00:09:27.659 --> 00:09:30.460 MYCIN was navigating a massive decision tree 00:09:30.460 --> 00:09:33.740 built from about 450 handwritten rules extracted 00:09:33.740 --> 00:09:36.279 from top infectious disease specialists. The 00:09:36.279 --> 00:09:38.980 text notes that MYCIN actually performed as well 00:09:38.980 --> 00:09:41.360 as some leading medical experts and considerably 00:09:41.360 --> 00:09:43.620 better than junior doctors. It was incredibly 00:09:43.620 --> 00:09:45.860 powerful. Okay, but I have to push back here. 00:09:46.240 --> 00:09:49.399 If we had software outperforming human doctors 00:09:49.399 --> 00:09:52.639 in the 1970s, Why aren't we all being diagnosed 00:09:52.639 --> 00:09:56.620 by descendants of MYCIN today? I mean, why did 00:09:56.620 --> 00:10:00.559 this lead to a second AI winter? Because maintaining 00:10:00.559 --> 00:10:03.840 those systems became an absolute nightmare. Expert 00:10:03.840 --> 00:10:05.860 systems worked beautifully because their domain 00:10:05.860 --> 00:10:08.519 was incredibly narrow. But imagine trying to 00:10:08.519 --> 00:10:10.639 keep thousands of interconnected handwritten 00:10:10.639 --> 00:10:13.840 rules perfectly updated as medical science evolves. 00:10:14.179 --> 00:10:16.639 Oh, I see. If a new drug is invented or a new 00:10:16.639 --> 00:10:19.019 symptom is discovered, a human programmer has 00:10:19.019 --> 00:10:21.299 to manually integrate that into the delicate 00:10:21.299 --> 00:10:24.960 web of 450 rules without breaking the logic of 00:10:24.960 --> 00:10:27.700 all the others. It became an impossible bottleneck. 00:10:27.840 --> 00:10:29.700 Plus, I imagine doctors weren't exactly thrilled 00:10:29.700 --> 00:10:32.279 about taking orders from a clunky 1970s computer 00:10:32.279 --> 00:10:34.960 terminal. Not at all. There was immense reluctance 00:10:34.960 --> 00:10:37.480 to trust a machine over a human's gut instinct. 00:10:37.860 --> 00:10:39.539 And we haven't even mentioned the corporate side 00:10:39.539 --> 00:10:41.559 yet. Oh, right. The corporate boom. The source 00:10:41.559 --> 00:10:44.240 mentions XCon. Yeah. XCon saved the computer 00:10:44.240 --> 00:10:47.460 company DEC millions of dollars. It reduced their 00:10:47.460 --> 00:10:49.820 computer configuration times from 90 days down 00:10:49.820 --> 00:10:53.500 to just 90 minutes. 90 days to 90 minutes? That's 00:10:53.500 --> 00:10:56.200 insane. It was a massive success. But eventually, 00:10:56.460 --> 00:10:59.110 the same maintenance bottleneck killed it. These 00:10:59.110 --> 00:11:01.330 systems were expensive to build, a nightmare 00:11:01.330 --> 00:11:04.250 to maintain, and culturally difficult to deploy. 00:11:04.429 --> 00:11:06.470 And the specialized hardware companies that built 00:11:06.470 --> 00:11:08.710 the machines for these systems went bankrupt 00:11:08.710 --> 00:11:11.809 as regular desktop computers just got faster 00:11:11.809 --> 00:11:14.990 and cheaper. The boom collapsed. The second AI 00:11:14.990 --> 00:11:19.269 winter set in from about 1988 to 1993. The very 00:11:19.269 --> 00:11:22.070 thing that made them powerful, those hyper -specific 00:11:22.070 --> 00:11:25.190 handwritten rules, became their fatal flaw. And 00:11:25.190 --> 00:11:27.570 this raises an important question. How do you 00:11:27.570 --> 00:11:30.529 program a machine to know what doesn't change 00:11:30.529 --> 00:11:33.470 when an action occurs? This gets to the deepest 00:11:33.470 --> 00:11:36.330 philosophical flaw of symbolic logic, known as 00:11:36.330 --> 00:11:38.779 the frame problem. Right, it's easy to program 00:11:38.779 --> 00:11:41.259 a rule that says if I drop this glass, it shatters. 00:11:41.559 --> 00:11:43.320 But you also have to program the machine to know 00:11:43.320 --> 00:11:45.340 that dropping the glass doesn't change the color 00:11:45.340 --> 00:11:47.059 of the walls or the temperature of the room or 00:11:47.059 --> 00:11:49.379 the fact that it's Tuesday. Exactly. And the 00:11:49.379 --> 00:11:51.480 companion to this is the qualification problem. 00:11:52.039 --> 00:11:55.120 You cannot possibly enumerate every single precondition 00:11:55.120 --> 00:11:57.690 required for an action to succeed. Like the car 00:11:57.690 --> 00:12:00.029 engine. Yeah, if an AI understands the logical 00:12:00.029 --> 00:12:03.190 mechanics of how a car engine works, it still 00:12:03.190 --> 00:12:05.730 wouldn't know that a banana shoved into the tailpipe 00:12:05.730 --> 00:12:08.350 would prevent the car from starting. A symbolic 00:12:08.350 --> 00:12:11.370 system can't anticipate the banana because nobody 00:12:11.370 --> 00:12:14.690 sat down and wrote an if -then rule for tropical 00:12:14.690 --> 00:12:17.250 fruit in the exhaust system. You simply cannot 00:12:17.250 --> 00:12:20.769 write a rule for every bizarre, unexpected variable 00:12:20.769 --> 00:12:23.029 the physical universe might throw at you. Here's 00:12:23.029 --> 00:12:25.500 where it gets really interesting. Think about 00:12:25.500 --> 00:12:28.240 your own daily commute. When you ride a bike, 00:12:28.620 --> 00:12:30.940 you don't consciously calculate the exact wind 00:12:30.940 --> 00:12:33.679 resistance, the pull of gravity, and the friction 00:12:33.679 --> 00:12:36.240 of every single pebble on the asphalt. No, you 00:12:36.240 --> 00:12:39.080 just ride. You use intuition. You let your body 00:12:39.080 --> 00:12:41.259 feel the balance. And according to the source 00:12:41.259 --> 00:12:44.250 material... That is exactly the realization that 00:12:44.250 --> 00:12:47.009 shattered the symbolic AI paradigm. A researcher 00:12:47.009 --> 00:12:49.730 at MIT named Rodney Brooks realized the only 00:12:49.730 --> 00:12:51.950 way to solve the problem was to just throw the 00:12:51.950 --> 00:12:54.169 rule book away completely. Yeah. He introduced 00:12:54.169 --> 00:12:56.690 something called Situated Robotics, or Nouvelle 00:12:56.690 --> 00:12:59.990 AI. He entirely rejected the use of central symbols 00:12:59.990 --> 00:13:02.370 and logical representations. So he built robots 00:13:02.370 --> 00:13:05.230 without a central brain. Essentially, yes. He 00:13:05.230 --> 00:13:07.669 built robots using what he called a subsumption 00:13:07.669 --> 00:13:11.149 architecture. It relies on layered behaviors. 00:13:11.419 --> 00:13:13.759 Sort of like an insect's nervous system. Okay, 00:13:13.779 --> 00:13:16.440 how does that work? The bottom layer is incredibly 00:13:16.440 --> 00:13:19.139 dumb. It just uses raw sensor data to say if 00:13:19.139 --> 00:13:22.200 you bump into a wall, turn left. The next layer 00:13:22.200 --> 00:13:24.940 up says move forward. The robot doesn't have 00:13:24.940 --> 00:13:27.220 an internal logical map of the room it's in. 00:13:27.700 --> 00:13:29.980 Brooks argued that the real world is its own 00:13:29.980 --> 00:13:32.379 best model. So you just let the robot interact 00:13:32.379 --> 00:13:34.740 with it directly? Yeah, and intelligence emerges 00:13:34.740 --> 00:13:37.330 from those simple layers overlapping. This complete 00:13:37.330 --> 00:13:39.570 rejection of symbols opened the floodgates for 00:13:39.570 --> 00:13:41.830 the anti -symbolic movement. And this is the 00:13:41.830 --> 00:13:44.049 foundation of the AI we see everywhere today. 00:13:44.509 --> 00:13:47.830 Deep learning or connectionist AI? Connectionism 00:13:47.830 --> 00:13:50.190 is based on artificial neural networks, which 00:13:50.190 --> 00:13:52.370 are loosely inspired by the biological neurons 00:13:52.370 --> 00:13:54.909 in our brains. Right. Instead of giving the computer 00:13:54.909 --> 00:13:57.309 rules, you feed it massive amounts of data and 00:13:57.309 --> 00:13:59.230 let it find the statistical patterns on its own. 00:13:59.629 --> 00:14:02.549 It learns intuitively, processing raw sensory 00:14:02.549 --> 00:14:05.190 input much closer to how we actually experience 00:14:05.190 --> 00:14:07.659 the world. This source material highlights just 00:14:07.659 --> 00:14:10.139 how intense the sociological conflict between 00:14:10.139 --> 00:14:12.919 these two camps became. It wasn't just a polite 00:14:12.919 --> 00:14:15.700 academic disagreement over coffee. It was a full 00:14:15.700 --> 00:14:19.860 on ideological feud. Oh, the rhetoric was surprisingly 00:14:19.860 --> 00:14:22.740 aggressive. The deep learning community, led 00:14:22.740 --> 00:14:25.200 by figures like Jeffrey Hinton and Jan LeCun. 00:14:25.580 --> 00:14:28.519 fiercely attacked the very concept of symbolic 00:14:28.519 --> 00:14:31.259 logic. Jeffrey Hinton gave a talk where he compared 00:14:31.259 --> 00:14:35.179 symbolic AI to ether, you know, the invisible 00:14:35.179 --> 00:14:38.279 substance that 19th century scientists mistakenly 00:14:38.279 --> 00:14:40.340 believed filled the universe. Right. A complete 00:14:40.340 --> 00:14:42.879 scientific dead end. He also told European Union 00:14:42.879 --> 00:14:45.500 leaders that investing in symbolic AI was like 00:14:45.500 --> 00:14:47.620 investing in internal combustion engines in the 00:14:47.620 --> 00:14:50.419 era of electric cars. He wanted outright replacement, 00:14:50.659 --> 00:14:53.340 not reconciliation. The deep learning camp firmly 00:14:53.340 --> 00:14:55.200 believed that if you just have enough data and 00:14:55.200 --> 00:14:57.639 enough computing power, intelligent behavior 00:14:57.639 --> 00:15:00.379 will organically emerge. You don't need a human 00:15:00.379 --> 00:15:02.779 to write a single rule. And that philosophy brings 00:15:02.779 --> 00:15:05.519 us to the present day. Deep learning absolutely 00:15:05.519 --> 00:15:07.740 won the battle of perception. It can recognize 00:15:07.740 --> 00:15:10.379 faces in a crowd, translate languages fluidly, 00:15:10.580 --> 00:15:13.240 generate breathtaking art. But it struggles with 00:15:13.240 --> 00:15:16.120 something fundamental. Transparent, step -by 00:15:16.120 --> 00:15:18.659 -step reasoning. It can taste the soup perfectly, 00:15:18.659 --> 00:15:21.379 but it cannot tell you the recipe. Exactly. It's 00:15:21.379 --> 00:15:24.500 an opaque black box. When a large language model 00:15:24.639 --> 00:15:27.960 hallucinates a completely fake legal case. The 00:15:27.960 --> 00:15:30.159 programmers can't just go in and fix line 42 00:15:30.159 --> 00:15:32.659 of the code because the logic isn't written down 00:15:32.659 --> 00:15:35.399 anywhere. It's just a blur of statistical weights. 00:15:35.899 --> 00:15:37.720 Because of that limitation, we're seeing something 00:15:37.720 --> 00:15:39.620 remarkable happen right now in research labs. 00:15:40.340 --> 00:15:44.070 The bitter rivals of AI. the deep learning neural 00:15:44.070 --> 00:15:47.029 networks and the old school symbolic logic systems 00:15:47.029 --> 00:15:49.850 are being forced to work together. Neuro symbolic 00:15:49.850 --> 00:15:51.929 AI. Yeah. And if we connect this to the bigger 00:15:51.929 --> 00:15:54.710 picture, it maps beautifully onto how human psychology 00:15:54.710 --> 00:15:56.909 actually works. Are you familiar with Daniel 00:15:56.909 --> 00:15:59.490 Kahneman's book, Thinking Fast and Slow? Absolutely. 00:15:59.830 --> 00:16:02.190 The idea that the human mind uses system one 00:16:02.190 --> 00:16:04.990 and system two thinking. Right. Kahneman argues 00:16:04.990 --> 00:16:07.629 human thinking has two distinct gears. System 00:16:07.629 --> 00:16:11.500 one is fast, intuitive, automatic, and unconscious. 00:16:12.120 --> 00:16:14.600 It's how you instantly recognize a friend's face 00:16:14.600 --> 00:16:17.039 across a crowded room. And that is perfectly 00:16:17.039 --> 00:16:20.159 modeled by deep learning. Exactly. But System 00:16:20.159 --> 00:16:23.500 2 is slower, deliberative, and step -by -step. 00:16:23.779 --> 00:16:26.600 It's what you use to solve a complex math equation 00:16:26.600 --> 00:16:30.159 or plan a multi -city vacation. Right. That is 00:16:30.159 --> 00:16:33.779 perfectly modeled by symbolic AI. So, neuro -symbolic 00:16:33.779 --> 00:16:36.259 AI is literally trying to build a machine with 00:16:36.259 --> 00:16:39.379 both systems. Gut instinct paired with logical 00:16:39.379 --> 00:16:41.639 math. Yeah. And we've already seen this hybrid 00:16:41.639 --> 00:16:44.120 approach work in the real world, haven't we? 00:16:44.159 --> 00:16:46.299 We have. The most famous example is AlphaGo, 00:16:46.779 --> 00:16:49.259 the AI that defeated the human world champion 00:16:49.259 --> 00:16:52.279 at the incredibly complex board games Go. It 00:16:52.279 --> 00:16:55.139 is a hybrid. It uses deep learning neural networks 00:16:55.139 --> 00:16:57.620 to intuitively evaluate the visual pattern of 00:16:57.620 --> 00:17:00.279 the stones on the board that is its system one 00:17:00.279 --> 00:17:02.639 gut instinct. Okay, what about system two? It 00:17:02.639 --> 00:17:04.480 uses something called a symbolic Monte Carlo 00:17:04.480 --> 00:17:07.019 tree search to explicitly plan out its future 00:17:07.019 --> 00:17:09.259 moves. I've heard that term Monte Carlo tree 00:17:09.259 --> 00:17:11.000 search. How does that actually work in practice? 00:17:11.200 --> 00:17:13.839 Imagine standing at a crossroads in a dense forest. 00:17:14.500 --> 00:17:17.140 Monte Carlo Tree Search is like the AI rapidly 00:17:17.140 --> 00:17:19.400 playing out thousands of random games in its 00:17:19.400 --> 00:17:22.019 head down each path before taking a single physical 00:17:22.019 --> 00:17:25.319 step. Oh, wow. It simulates the future, tracking 00:17:25.319 --> 00:17:27.859 the statistics of wins and losses for each branch, 00:17:28.359 --> 00:17:30.200 ensuring it picks the route with the highest 00:17:30.200 --> 00:17:33.839 mathematical chance of victory. It is pure step 00:17:33.839 --> 00:17:37.039 -by -step system two planning. So it uses the 00:17:37.039 --> 00:17:39.450 neural net to sense the board. And the symbolic 00:17:39.450 --> 00:17:41.869 search to plan the attack. Exactly. And this 00:17:41.869 --> 00:17:44.210 synthesis isn't just a clever engineering trick. 00:17:44.349 --> 00:17:46.349 It might be scientifically necessary for the 00:17:46.349 --> 00:17:49.190 future of the field. The source cites a recent 00:17:49.190 --> 00:17:51.589 theoretical proof by a researcher named Hang 00:17:51.589 --> 00:17:54.410 Zhang and his colleagues. What did they prove? 00:17:54.589 --> 00:17:57.430 They mathematically proved that mainstream knowledge 00:17:57.430 --> 00:18:00.269 representation formalisms are recursively isomorphic. 00:18:00.470 --> 00:18:02.349 OK. You are going to have to unpack recursively 00:18:02.349 --> 00:18:05.269 isomorphic for us. In plain terms, it means that 00:18:05.269 --> 00:18:07.670 from a purely theoretical standpoint, neither 00:18:07.670 --> 00:18:09.970 the symbolic approach nor the connectionist deep 00:18:09.970 --> 00:18:11.910 learning approach is fundamentally superior. 00:18:12.190 --> 00:18:15.029 Wait, really? Yeah. What one system can represent, 00:18:15.210 --> 00:18:17.849 the other can also represent. They are functionally 00:18:17.849 --> 00:18:20.849 equivalent in their expressive power. Therefore, 00:18:21.309 --> 00:18:23.690 the decades -long feud over which is the one 00:18:23.690 --> 00:18:26.569 true path was entirely missing the point. We 00:18:26.569 --> 00:18:29.769 need the strengths of both. We do. Current large 00:18:29.769 --> 00:18:32.130 language models, the transformers everyone is 00:18:32.130 --> 00:18:36.140 using today, are incredibly powerful, but entirely 00:18:36.140 --> 00:18:39.640 opaque. To get robust AI that we can actually 00:18:39.640 --> 00:18:41.839 trust in high -stake situations like medical 00:18:41.839 --> 00:18:44.920 diagnosis or autonomous driving, we might need 00:18:44.920 --> 00:18:47.299 the internal combustion engine of simple manipulation 00:18:47.299 --> 00:18:50.920 to handle abstract logical knowledge and verify 00:18:50.920 --> 00:18:53.759 the neural network's guesses. So what does this 00:18:53.759 --> 00:18:56.000 all mean for you, the learner listening right 00:18:56.000 --> 00:18:58.299 now? I mean, does the ultimate artificial intelligence 00:18:58.299 --> 00:19:00.740 need to have a split brain, just like humans 00:19:00.740 --> 00:19:04.859 do, to actually be smart? The evidence strongly 00:19:04.859 --> 00:19:07.299 points in that direction. Pure logic failed because 00:19:07.299 --> 00:19:09.740 it couldn't handle the messy, unpredictable reality 00:19:09.740 --> 00:19:12.119 of a banana and a tailpipe. Right. And pure deep 00:19:12.119 --> 00:19:14.059 learning is struggling right now because it hallucinates 00:19:14.059 --> 00:19:16.119 facts, it can't explain its own reasoning, and 00:19:16.119 --> 00:19:18.180 it lacks basic common sense. The real takeaway 00:19:18.180 --> 00:19:20.579 from this deep dive is context. The next time 00:19:20.579 --> 00:19:22.819 you see a news headline hyping up a terrifying 00:19:22.819 --> 00:19:25.200 new deep learning model that seems like magic... 00:19:25.200 --> 00:19:27.019 Or an article warning that we are headed for 00:19:27.019 --> 00:19:29.339 another catastrophic AI winter... Exactly. You 00:19:29.339 --> 00:19:32.349 can see through the noise. You understand that 00:19:32.349 --> 00:19:35.130 the brittleness of AI isn't some new phenomenon 00:19:35.130 --> 00:19:38.089 born in the last five years. It's the exact same 00:19:38.089 --> 00:19:40.390 banana in the tailpipe problem that brilliant 00:19:40.390 --> 00:19:42.250 researchers have been wrestling with since the 00:19:42.250 --> 00:19:45.309 1950s. Capturing common sense remains the holy 00:19:45.309 --> 00:19:47.390 grail of computer science. And it leaves you 00:19:47.390 --> 00:19:50.650 with one final thought to mull over as you go 00:19:50.650 --> 00:19:54.069 about your day. If neuro -symbolic AI really 00:19:54.069 --> 00:19:57.079 is the path forward, If it perfectly mirrors 00:19:57.079 --> 00:20:00.319 human cognition, where deep learning is our opaque, 00:20:00.559 --> 00:20:03.519 unconscious intuition. And symbolic AI is our 00:20:03.519 --> 00:20:05.880 readable, conscious reasoning. Yeah. Then think 00:20:05.880 --> 00:20:07.400 about what we have been building over the last 00:20:07.400 --> 00:20:09.599 decade. Are we currently just building a massive 00:20:09.599 --> 00:20:13.240 planetary AI subconscious? Is this colossal network 00:20:13.240 --> 00:20:15.779 of data rapidly expanding, completely in the 00:20:15.779 --> 00:20:17.940 dark, just waiting for us to finally hand it 00:20:17.940 --> 00:20:19.880 the symbolic tools of a conscious mind?