WEBVTT 00:00:00.000 --> 00:00:02.520 What if I told you that the blueprint for the 00:00:02.520 --> 00:00:06.360 AI in your smartphone right now wasn't, uh... 00:00:06.400 --> 00:00:09.640 wasn't sketched out in some Silicon Valley garage. 00:00:09.779 --> 00:00:12.019 Yeah, it actually started on the chalkboard of 00:00:12.019 --> 00:00:15.839 a 17th century philosopher, which is just wild 00:00:15.839 --> 00:00:18.260 to think about. Right, like long before the first 00:00:18.260 --> 00:00:21.359 microchip was even a concept, thinkers were obsessed 00:00:21.359 --> 00:00:25.219 with this wildly ambitious idea that human reason 00:00:25.219 --> 00:00:27.859 itself could just be reduced to a math equation. 00:00:28.019 --> 00:00:30.500 It sounds like science fiction, I know, but Gottfried 00:00:30.500 --> 00:00:32.899 Leibniz actually envisioned a universal language 00:00:32.899 --> 00:00:36.700 of logic back in the 1600s? 1600s. Exactly. He 00:00:36.700 --> 00:00:40.219 imagined a future where two philosophers having 00:00:40.219 --> 00:00:42.299 this bitter disagreement wouldn't even need to 00:00:42.299 --> 00:00:43.880 argue. They could just take out their slates 00:00:43.880 --> 00:00:46.759 and declare, let us calculate. Like solving for 00:00:46.759 --> 00:00:49.560 X and algebra, but for a moral debate. Precisely. 00:00:49.719 --> 00:00:52.380 Resolving philosophical disputes with pure math. 00:00:52.560 --> 00:00:55.159 OK, let's unpack this. Because if you're listening 00:00:55.159 --> 00:00:57.479 to this deep dive right now, you probably follow 00:00:57.479 --> 00:00:59.880 the tech industry. You know all about the current 00:00:59.880 --> 00:01:02.750 massive boom in generative models. Yeah. But 00:01:02.750 --> 00:01:05.170 looking at our stack of sources today, which 00:01:05.170 --> 00:01:08.269 is this really comprehensive history of artificial 00:01:08.269 --> 00:01:10.790 intelligence, the real story isn't just some 00:01:10.790 --> 00:01:12.930 straight line of modern progress. Not at all. 00:01:13.030 --> 00:01:16.250 It's actually a grueling, century -long war. 00:01:16.329 --> 00:01:18.870 Yeah, a war between two completely different 00:01:18.870 --> 00:01:21.409 philosophies of how to build a mind. We're looking 00:01:21.409 --> 00:01:24.549 at a cycle of incredible hubris. devastating 00:01:24.549 --> 00:01:27.189 research crashes, which they actually call winters, 00:01:27.469 --> 00:01:29.689 and the fundamental shifts in computer science 00:01:29.689 --> 00:01:33.069 that brought us to today's multi -billion dollar 00:01:33.069 --> 00:01:35.590 reality. And to understand that war, we really 00:01:35.590 --> 00:01:37.709 have to look at how we transition from ancient 00:01:37.709 --> 00:01:40.329 dreams to actual engineering. Because the sources 00:01:40.329 --> 00:01:43.010 mention Greek myths, right? Like the bronze giant. 00:01:43.200 --> 00:01:46.780 Right, and Pygmalion's Galatea, these early myths 00:01:46.780 --> 00:01:49.700 of artificial beings. But the actual engineering 00:01:49.700 --> 00:01:52.140 bridge that was built in 1950 by Alan Turing. 00:01:52.260 --> 00:01:55.500 Oh, the Turing test. Yes. He looked at the absolute 00:01:55.500 --> 00:01:58.400 mess of philosophical arguments about what constitutes 00:01:58.400 --> 00:02:01.780 a mind or a soul, and he just completely sidestepped 00:02:01.780 --> 00:02:04.239 them. His approach was brilliantly lazy in a 00:02:04.239 --> 00:02:06.959 way. The Turing test basically says that trying 00:02:06.959 --> 00:02:09.919 to define thinking is a trap. It really is. So 00:02:09.919 --> 00:02:11.860 instead, let's just look at the output. I always 00:02:11.860 --> 00:02:13.960 compare it to a blind -paced test for intelligence. 00:02:14.099 --> 00:02:16.659 Oh, I like that analogy. Yeah, like if you are 00:02:16.659 --> 00:02:19.699 chatting over a text interface and you literally 00:02:19.699 --> 00:02:22.060 cannot tell the difference between the machine's 00:02:22.060 --> 00:02:24.419 responses and a human's responses. You basically 00:02:24.419 --> 00:02:27.219 have to admit the machine is thinking. Exactly. 00:02:27.439 --> 00:02:29.759 You don't know the recipe happening inside the 00:02:29.759 --> 00:02:33.139 black box, but the output tastes exactly like 00:02:33.139 --> 00:02:36.240 human intelligence. And that pragmatism was the 00:02:36.240 --> 00:02:38.939 spark. It gave researchers permission to stop 00:02:38.939 --> 00:02:41.520 agonizing over the philosophy of it all and just 00:02:41.520 --> 00:02:43.919 start building. Which leads directly to the official 00:02:43.919 --> 00:02:47.060 birth of the field, right? The 1956 Dartmouth 00:02:47.060 --> 00:02:49.759 workshop. That's the one. Organized by John McCarthy 00:02:49.759 --> 00:02:51.979 and Marvin Minsky. And the source material has 00:02:51.979 --> 00:02:54.340 this incredibly petty detail that I just love. 00:02:54.960 --> 00:02:57.520 McCarthy coined the specific term artificial 00:02:57.520 --> 00:03:00.360 intelligence for this workshop for one main reason. 00:03:00.560 --> 00:03:03.520 To dodge a rival. Yeah. He wanted to distance 00:03:03.520 --> 00:03:06.000 his new field from Norbert Wiener's established 00:03:06.000 --> 00:03:09.319 work on cybernetics. He essentially branded a 00:03:09.319 --> 00:03:12.080 whole new scientific discipline just to dodge 00:03:12.080 --> 00:03:14.789 another academic's shadow. Academic rivalries 00:03:14.789 --> 00:03:17.550 are a powerful catalyst, that's for sure. What's 00:03:17.550 --> 00:03:20.330 fascinating here is the massive intellectual 00:03:20.330 --> 00:03:23.469 shift this workshop cemented. A cognitive revolution. 00:03:23.750 --> 00:03:26.229 Exactly. We call it the cognitive revolution. 00:03:26.669 --> 00:03:29.990 Because before 1956, the dominant theory in psychology 00:03:29.990 --> 00:03:32.389 was behaviorism. Right, where you only study 00:03:32.389 --> 00:03:35.669 physical stimuli and responses. Like Pavlov's 00:03:35.669 --> 00:03:39.169 dogs. Ring a bell. The dog drools. Yes, and you 00:03:39.169 --> 00:03:41.150 don't worry about what the dog is actually thinking. 00:03:41.770 --> 00:03:43.530 Behaviorists believed you couldn't scientifically 00:03:43.530 --> 00:03:46.370 study unobservable things like thoughts or memories. 00:03:46.550 --> 00:03:48.669 But the Dartmouth researchers threw that completely 00:03:48.669 --> 00:03:51.430 out the window. They did. They argued that mental 00:03:51.430 --> 00:03:53.789 objects like, say, a plan to grab an apple or 00:03:53.789 --> 00:03:56.490 the memory of a face were real, tangible things. 00:03:56.689 --> 00:03:58.530 Things that could be simulated by high -level 00:03:58.530 --> 00:04:01.310 symbols inside a machine. So they firmly believed 00:04:01.310 --> 00:04:04.669 that human thought was, at its core, just symbol 00:04:04.669 --> 00:04:07.250 manipulation. Right. The brain was a computer 00:04:07.250 --> 00:04:09.409 and thoughts were just code. Which brings us 00:04:09.409 --> 00:04:12.050 to the first major philosophy in our century 00:04:12.050 --> 00:04:16.610 -long, war -symbolic AI. And man, they were wildly 00:04:16.610 --> 00:04:18.889 overconfident about it. Oh, the hubris was off 00:04:18.889 --> 00:04:21.529 the charts. Moving into the late 50s and 60s, 00:04:21.649 --> 00:04:25.689 we hit this golden age of hubris. They were programming 00:04:25.689 --> 00:04:30.189 computers to solve algebra word problems, prove 00:04:30.189 --> 00:04:33.569 complex geometry theorems. They even built ELISA. 00:04:33.850 --> 00:04:36.850 Right. Eliza, the first chatbot that could actually 00:04:36.850 --> 00:04:39.569 mimic a psychotherapist. To the public, it just 00:04:39.569 --> 00:04:41.829 looked like pure magic. And those early victories 00:04:41.829 --> 00:04:44.610 created a really intoxicating atmosphere. You 00:04:44.610 --> 00:04:47.569 had the brightest minds of the era making staggering 00:04:47.569 --> 00:04:51.120 predictions. Like Minsky. Yeah. In 1967, Marvin 00:04:51.120 --> 00:04:53.759 Minsky publicly declared that the problem of 00:04:53.759 --> 00:04:55.620 artificial intelligence would be substantially 00:04:55.620 --> 00:04:58.620 solved within a single generation. Single generation. 00:04:58.680 --> 00:05:00.980 They honestly thought they were like 20 years 00:05:00.980 --> 00:05:03.220 away from building a synthetic human brain. And 00:05:03.220 --> 00:05:04.939 the U .S. government bought the hype completely. 00:05:05.199 --> 00:05:07.720 Oh, yeah. DARPA started pouring millions of dollars 00:05:07.720 --> 00:05:10.779 into undirected freewheeling research labs. But 00:05:10.779 --> 00:05:12.860 and here's the turning point. They didn't solve 00:05:12.860 --> 00:05:15.519 it. They ran face first into a brick wall. A 00:05:15.519 --> 00:05:18.459 wall made of pure mathematics. It's known as 00:05:18.459 --> 00:05:21.160 the combinatorial explosion. Break that down 00:05:21.160 --> 00:05:24.240 for us. So the early symbolic AI programs worked 00:05:24.240 --> 00:05:26.220 by searching through a maze of possibilities. 00:05:27.079 --> 00:05:29.560 If the AI is trying to prove a geometry theorem, 00:05:29.779 --> 00:05:31.579 the rules are strict. Right, the boundaries are 00:05:31.579 --> 00:05:34.220 clear. Exactly. The number of possible moves 00:05:34.220 --> 00:05:37.290 is relatively small. The computer explores path 00:05:37.290 --> 00:05:40.470 A, then path B, and finds the solution. But the 00:05:40.470 --> 00:05:43.430 real world isn't a geometry proof. Far from it. 00:05:43.750 --> 00:05:46.370 When you try to apply that same maze searching 00:05:46.370 --> 00:05:49.389 logic to the real physical world, the number 00:05:49.389 --> 00:05:52.189 of branching paths just explodes exponentially. 00:05:52.529 --> 00:05:54.990 So if you want AI to simply identify a chair 00:05:54.990 --> 00:05:57.519 in a messy room, The lighting changes, the angles 00:05:57.519 --> 00:06:00.259 change. Maybe the chair is partially obscured 00:06:00.259 --> 00:06:02.620 by a jacket. The possible variables multiply 00:06:02.620 --> 00:06:04.680 into the trillions. And those early computers 00:06:04.680 --> 00:06:07.120 just choked on the complexity of reality. They 00:06:07.120 --> 00:06:09.279 didn't have the processing power or memory. They 00:06:09.279 --> 00:06:12.300 didn't. But beyond just the raw hardware limitations, 00:06:12.620 --> 00:06:14.899 the whole premise of using logical symbols had 00:06:14.899 --> 00:06:18.740 a fatal conceptual flaw, which is perfectly captured 00:06:18.740 --> 00:06:22.639 by Moravec's paradox. Oh, I find Moravec's paradox 00:06:22.639 --> 00:06:25.860 so clarifying. Explain what that is. Researchers 00:06:25.860 --> 00:06:28.439 discovered it was relatively easy to program 00:06:28.439 --> 00:06:30.740 a computer to do highly intelligent tasks like 00:06:30.740 --> 00:06:33.300 playing championship chess or doing calculus. 00:06:33.439 --> 00:06:35.980 High -level reasoning. Right, but it was practically 00:06:35.980 --> 00:06:38.459 impossible to get a machine to do unintelligent 00:06:38.459 --> 00:06:41.920 tasks. Like recognizing a spoken word. or walking 00:06:41.920 --> 00:06:44.560 across a room without bumping into a table. Exactly. 00:06:44.860 --> 00:06:47.399 To put in perspective for you listening, imagine 00:06:47.399 --> 00:06:50.199 building a supercomputer that can instantly solve 00:06:50.199 --> 00:06:54.160 a wildly complex astrophysics equation, but it 00:06:54.160 --> 00:06:56.720 possesses the physical coordination and sensory 00:06:56.720 --> 00:07:00.339 perception of a literal toddler. It's a great 00:07:00.339 --> 00:07:02.759 image. It literally cannot figure out how to 00:07:02.759 --> 00:07:04.459 walk across the living room without tripping 00:07:04.459 --> 00:07:06.819 over its own feet. And the evolutionary biology 00:07:06.819 --> 00:07:09.040 behind that is striking. We take our sensor rotor 00:07:09.040 --> 00:07:11.579 skill, seeing, hearing. balancing completely 00:07:11.579 --> 00:07:13.959 for granted. Because millions of years of brutal 00:07:13.959 --> 00:07:16.160 evolution have hardwired them into our brains. 00:07:16.399 --> 00:07:18.519 Exactly. They're unconscious and effortless. 00:07:18.819 --> 00:07:21.439 But symbolic reasoning, the kind of top -down 00:07:21.439 --> 00:07:24.379 logic we use to play chess, is a very recent, 00:07:24.519 --> 00:07:27.680 very thin evolutionary layer. So trying to teach 00:07:27.680 --> 00:07:30.560 a machine to see and navigate the world using 00:07:30.560 --> 00:07:33.579 only top -down logical symbols was a complete 00:07:33.579 --> 00:07:36.259 dead end. A total dead end. And the fallout from 00:07:36.259 --> 00:07:39.240 that was brutal. Yeah, by the 1970s, the government 00:07:39.240 --> 00:07:41.600 realized these programs were essentially highly 00:07:41.600 --> 00:07:44.240 expensive parlor tricks. The sources point to 00:07:44.240 --> 00:07:47.579 the 1973 Lighthill Report in the UK, which just 00:07:47.579 --> 00:07:50.100 savagely criticized the field. For failing to 00:07:50.100 --> 00:07:52.660 achieve pretty much any of its grandiose promises. 00:07:52.980 --> 00:07:55.300 And in the U .S., the Mansfield Amendment forced 00:07:55.300 --> 00:07:57.819 DARPA to stop funding theoretical exploration 00:07:57.819 --> 00:08:01.269 entirely. They would only fund direct military 00:08:01.269 --> 00:08:03.769 -oriented applications. The money dried up almost 00:08:03.769 --> 00:08:06.529 overnight. Welcome to the first AI winter. Yeah. 00:08:07.110 --> 00:08:09.810 And to survive the 1980s, the field had to pivot 00:08:09.810 --> 00:08:12.569 drastically. General human -like intelligence 00:08:12.569 --> 00:08:15.259 was entirely off the table. AI had to prove it 00:08:15.259 --> 00:08:16.980 could be commercially viable. It had to make 00:08:16.980 --> 00:08:19.779 money. Enter the expert systems. Right. If the 00:08:19.779 --> 00:08:21.899 AI couldn't understand the whole world, researchers 00:08:21.899 --> 00:08:24.759 decided to trap it in a tiny, highly specific 00:08:24.759 --> 00:08:27.079 box. They built programs that were basically 00:08:27.079 --> 00:08:30.420 giant digital encyclopedias of if -then rules 00:08:30.420 --> 00:08:33.919 meticulously programmed by human experts. And 00:08:33.919 --> 00:08:36.460 financially, it worked. The source highlights 00:08:36.460 --> 00:08:39.639 a system called R1 built for the Digital Equipment 00:08:39.639 --> 00:08:42.789 Corporation. It saved the company $40 million 00:08:42.789 --> 00:08:45.710 a year just by logically configuring customer 00:08:45.710 --> 00:08:49.029 computer orders. $40 million is no joke. Suddenly, 00:08:49.409 --> 00:08:52.789 AI is a booming, billion -dollar corporate industry 00:08:52.789 --> 00:08:55.350 again. But let me push back on the idea of these 00:08:55.350 --> 00:08:58.309 systems being intelligent at all, then. If they 00:08:58.309 --> 00:09:01.429 are just giant, rigid rule books. Wouldn't they 00:09:01.429 --> 00:09:03.870 completely break the second they encountered 00:09:03.870 --> 00:09:06.750 a scenario even slightly outside their explicit 00:09:06.750 --> 00:09:09.250 programming? Oh, they absolutely did. Structurally, 00:09:09.250 --> 00:09:11.490 the whole thing was a house of cards. They suffered 00:09:11.490 --> 00:09:13.789 from a fatal flaw known in the literature as 00:09:13.789 --> 00:09:16.470 brittleness. Meaning they possess zero common 00:09:16.470 --> 00:09:19.029 sense. None. A human knows that water is wet 00:09:19.029 --> 00:09:21.330 or that you can't push a string or that a car 00:09:21.330 --> 00:09:23.610 can't drive through a solid brick wall. Right. 00:09:23.610 --> 00:09:25.769 We just know that by existing in the world. But 00:09:25.769 --> 00:09:28.029 an expert system doesn't know any of that unless 00:09:28.029 --> 00:09:30.429 a programmer specifically wrote a line of code 00:09:30.429 --> 00:09:34.370 stating rule 4 ,000, water equals wet. If you 00:09:34.370 --> 00:09:37.070 fed them an unusual input, they didn't just fail 00:09:37.070 --> 00:09:40.649 gracefully. They made grotesque, absurd mistakes 00:09:40.649 --> 00:09:43.129 that no human would ever make. Like if you've 00:09:43.129 --> 00:09:46.490 ever yelled representative at an automated customer 00:09:46.490 --> 00:09:49.269 service phone menu because it completely failed 00:09:49.269 --> 00:09:52.210 to understand your slightly unique problem, you 00:09:52.210 --> 00:09:55.110 have personally experienced the exact brittleness 00:09:55.110 --> 00:09:58.409 of 1980s expert systems. That is the perfect 00:09:58.409 --> 00:10:00.710 modern equivalent. And to fix that brittleness, 00:10:01.190 --> 00:10:03.269 some researchers tried to attack the common sense 00:10:03.269 --> 00:10:06.120 problem head on. The most famous example is the 00:10:06.120 --> 00:10:09.059 Psych Project. CYC, right? Yes. They attempted 00:10:09.059 --> 00:10:12.279 to literally hand code millions of everyday facts 00:10:12.279 --> 00:10:15.879 into a massive database. Manually typing in every 00:10:15.879 --> 00:10:17.860 single fact about reality. That sounds like a 00:10:17.860 --> 00:10:20.179 Sisyphean nightmare. It was incredibly tedious. 00:10:20.519 --> 00:10:22.700 And ultimately, it didn't solve the core issue. 00:10:23.039 --> 00:10:25.620 The failure of expert systems to adapt led to 00:10:25.620 --> 00:10:27.940 a massive rebellion within the field. This is 00:10:27.940 --> 00:10:29.559 my favorite part of the historical timeline, 00:10:30.039 --> 00:10:32.809 the rebel faction. the biologists strike back 00:10:32.809 --> 00:10:35.330 against the logicians. Exactly. A faction of 00:10:35.330 --> 00:10:37.350 researchers decided that the entire paradigm 00:10:37.350 --> 00:10:39.990 of using logical symbols was just wrong. You 00:10:39.990 --> 00:10:42.149 had roboticists like Rodney Brooks publishing 00:10:42.149 --> 00:10:45.549 this famously provocative paper titled, Elephants 00:10:45.549 --> 00:10:48.129 Don't Play Chess. Such a great title. Right. 00:10:48.669 --> 00:10:51.389 He argued that true intelligence requires a physical 00:10:51.389 --> 00:10:53.590 body to interact with the messy world from the 00:10:53.590 --> 00:10:56.700 bottom up. He explicitly stated that you don't 00:10:56.700 --> 00:10:59.879 need to build a complex, symbolic model of reality 00:10:59.879 --> 00:11:03.000 inside the computer's memory because the world 00:11:03.000 --> 00:11:05.379 is its own best model. You don't need a digital 00:11:05.379 --> 00:11:07.840 map of the room if you just build sensors good 00:11:07.840 --> 00:11:10.100 enough to react to the actual room in real time. 00:11:10.360 --> 00:11:12.639 Exactly. And alongside the roboticists, you had 00:11:12.639 --> 00:11:15.440 the quiet revival of artificial neural networks. 00:11:15.539 --> 00:11:18.220 Which is huge. For decades, the mainstream AI 00:11:18.220 --> 00:11:20.679 community had completely ignored them. But in 00:11:20.679 --> 00:11:23.620 the 1980s, a concept called connectionism gained 00:11:24.259 --> 00:11:26.480 Instead of trying to program a calculator with 00:11:26.480 --> 00:11:29.480 rigid logic rules, they tried to mimic the biological 00:11:29.480 --> 00:11:31.580 structure of a human brain. And this is where 00:11:31.580 --> 00:11:33.600 we have to talk about the mechanism that actually 00:11:33.600 --> 00:11:36.399 makes neural networks function. Backpropagation. 00:11:36.620 --> 00:11:38.860 Backprop. Break it down for us. Backpropagation 00:11:38.860 --> 00:11:42.070 is essentially the engine of modern AI. Imagine 00:11:42.070 --> 00:11:44.590 a neural network as a massive grid of thousands 00:11:44.590 --> 00:11:47.169 of dials. Okay, thousands of dials. When you 00:11:47.169 --> 00:11:49.710 feed it a picture of a dog and ask to identify 00:11:49.710 --> 00:11:52.549 the image, at first all the dials are set randomly, 00:11:52.669 --> 00:11:54.769 so it guesses cat. Because it doesn't know any 00:11:54.769 --> 00:11:58.129 better yet. Right. Back propagation is the mathematical 00:11:58.129 --> 00:12:00.590 algorithm that measures how wrong that guess 00:12:00.590 --> 00:12:03.610 was, the error rate, and sends a signal backward 00:12:03.610 --> 00:12:06.350 through the entire network. Adjusting the dials. 00:12:06.429 --> 00:12:09.590 Exactly. It tiny adjusts every single dial so 00:12:09.590 --> 00:12:11.870 that the next time it sees that image, it is 00:12:11.870 --> 00:12:14.789 slightly more likely to guess dog. It is learning 00:12:14.789 --> 00:12:17.190 purely by trial and error. It's like a giant 00:12:17.190 --> 00:12:19.370 high speed game of hot or cold. That's exactly 00:12:19.370 --> 00:12:22.409 what it is. But despite this massive conceptual 00:12:22.409 --> 00:12:24.669 breakthrough, the business world in the late 00:12:24.669 --> 00:12:27.629 80s didn't care. No, the timing was terrible. 00:12:27.750 --> 00:12:29.690 The corporate expert systems were collapsing 00:12:29.690 --> 00:12:31.570 under the weight of their own massive maintenance 00:12:31.570 --> 00:12:35.049 costs, and desktop computers from Apple and IBM 00:12:35.049 --> 00:12:38.549 got cheaper and faster than specialized AI hardware. 00:12:38.690 --> 00:12:42.309 So the market imploded. Yeah. By the 1990s, the 00:12:42.309 --> 00:12:45.090 money vanished again. We plunge into the second 00:12:45.090 --> 00:12:48.710 AI winter. The sources note the term artificial 00:12:48.710 --> 00:12:51.309 intelligence actually became toxic. It became 00:12:51.309 --> 00:12:55.759 a complete taboo word. Over 300 AI companies 00:12:55.759 --> 00:13:00.000 went bankrupt or shut down. 300? Wow. But it's 00:13:00.000 --> 00:13:01.980 vital to note that during the second winter, 00:13:02.440 --> 00:13:04.759 the underlying math didn't stop. Researchers 00:13:04.759 --> 00:13:07.759 simply rebranded. They went into stealth mode. 00:13:08.080 --> 00:13:10.360 Stealth mode? Yeah, they called the word computational 00:13:10.360 --> 00:13:12.899 intelligence or machine learning or informatics 00:13:12.899 --> 00:13:15.659 just to get grant funding. They just peeled the 00:13:15.659 --> 00:13:18.220 AI label right off the box. And under those highly 00:13:18.220 --> 00:13:21.059 technical names, neural networks quietly began 00:13:21.059 --> 00:13:23.340 running the modern world. Through the 90s and 00:13:23.340 --> 00:13:26.279 2000s, these algorithms were optimizing logistics, 00:13:26.860 --> 00:13:29.480 mining data, assisting in medical diagnoses. 00:13:29.840 --> 00:13:33.139 And in 1997, you had the highly publicized milestone 00:13:33.139 --> 00:13:36.360 of IBM's Deep Blue, using raw computational search 00:13:36.360 --> 00:13:38.799 power to beat the world chess champion, Garry 00:13:38.799 --> 00:13:41.009 Kasparov. True, but Deep Blue is still relying 00:13:41.009 --> 00:13:43.450 on that old -school brute -force search logic. 00:13:43.789 --> 00:13:45.970 Ah, right. So the real revolution, the moment 00:13:45.970 --> 00:13:47.990 neural networks finally won the century -long 00:13:47.990 --> 00:13:50.710 war, required what the sources describe as a 00:13:50.710 --> 00:13:53.049 perfect storm. Yes. Neural networks have been 00:13:53.049 --> 00:13:55.649 theoretically sound since the 50s, but to actually 00:13:55.649 --> 00:13:58.570 function well, they require two vital resources. 00:13:58.919 --> 00:14:02.419 Which are? Immense processing power and mountains 00:14:02.419 --> 00:14:05.799 of training data. For 50 years, humanity simply 00:14:05.799 --> 00:14:08.480 didn't possess enough of either. Until the 2010s. 00:14:09.080 --> 00:14:11.460 Right. Moore's law had finally given us incredibly 00:14:11.460 --> 00:14:15.000 fast computer chips, specifically GPUs, and the 00:14:15.000 --> 00:14:18.240 explosion of the internet gave us big data. The 00:14:18.240 --> 00:14:20.019 sources highlight the turning point clearly. 00:14:20.500 --> 00:14:23.639 A researcher named Fei -Fei Lai created ImageNet, 00:14:24.240 --> 00:14:27.240 which was a massive database of three million 00:14:27.240 --> 00:14:30.370 human -labeled images. And in 2012, researchers 00:14:30.370 --> 00:14:33.610 combined that massive data set with a deep neural 00:14:33.610 --> 00:14:36.330 network architecture called AlexNet. They unleashed 00:14:36.330 --> 00:14:39.009 it on an image recognition competition, and it 00:14:39.009 --> 00:14:42.269 completely obliterated every single traditional 00:14:42.269 --> 00:14:45.750 symbolic AI method. It was a total paradigm shift. 00:14:46.070 --> 00:14:48.330 AlexNet proved definitively that if you feed 00:14:48.330 --> 00:14:50.870 a neural network enough data, it will learn the 00:14:50.870 --> 00:14:53.190 visual patterns on its own, infinitely better 00:14:53.190 --> 00:14:55.490 than a human programmer could ever explicitly 00:14:55.490 --> 00:14:58.309 code them. The entire field basically ab - and 00:14:58.309 --> 00:15:00.649 handcrafted logic rules overnight and jumped 00:15:00.649 --> 00:15:02.730 onto the deep learning train. Here's where it 00:15:02.730 --> 00:15:05.149 gets really interesting. Because the processing 00:15:05.149 --> 00:15:08.090 power keeps scaling and the data sets keep growing. 00:15:08.830 --> 00:15:11.309 We move from identifying images to generating 00:15:11.309 --> 00:15:14.169 language. The jump to language was massive. In 00:15:14.169 --> 00:15:17.649 2017, researchers at Google published a landmark 00:15:17.649 --> 00:15:20.389 paper introducing the transformer architecture. 00:15:21.210 --> 00:15:23.610 And the secret sauce of the transformer is a 00:15:23.610 --> 00:15:26.559 mechanism called self -attention. which fundamentally 00:15:26.559 --> 00:15:29.600 changed how machines process language. See, older 00:15:29.600 --> 00:15:31.899 models read text sequentially, word by word. 00:15:31.960 --> 00:15:34.759 They had very short memories. Right. To visualize 00:15:34.759 --> 00:15:37.200 the difference for you listening, imagine an 00:15:37.200 --> 00:15:40.220 AI trying to read a murder mystery novel. An 00:15:40.220 --> 00:15:44.299 older sequential AI reads word by word. By the 00:15:44.299 --> 00:15:46.279 time it gets to the killer's confession in Chapter 00:15:46.279 --> 00:15:48.860 10, it is completely forgotten about the bloody 00:15:48.860 --> 00:15:51.179 knife mentioned back in Chapter 1. It loses the 00:15:51.179 --> 00:15:53.980 thread completely. But a transformer, using self 00:15:53.980 --> 00:15:55.779 -attention, is like a detective standing in front 00:15:55.779 --> 00:15:57.820 of a giant cork board with red string. Oh, I 00:15:57.820 --> 00:16:00.580 love the cork board analogy. It processes the 00:16:00.580 --> 00:16:03.240 entire text at once. It mathematically draws 00:16:03.240 --> 00:16:06.120 a direct red string between the word knife in 00:16:06.120 --> 00:16:10.100 Chapter 1 and the suspect in Chapter 10. It constantly 00:16:10.100 --> 00:16:12.220 weighs the relevance of every single word against 00:16:12.220 --> 00:16:14.120 every other word in the sequence simultaneously. 00:16:14.360 --> 00:16:16.659 And that holistic understanding of context is 00:16:16.659 --> 00:16:18.860 exactly what allowed the creation of large language 00:16:18.860 --> 00:16:22.600 models or LLMs. We transition from AI being this 00:16:22.600 --> 00:16:25.580 hidden tool optimizing search results right back 00:16:25.580 --> 00:16:28.279 into the spotlight as a highly visible world 00:16:28.279 --> 00:16:30.320 -altering force. It's almost a return to those 00:16:30.320 --> 00:16:32.799 ancient myths of Pygmalion. It really is. We 00:16:32.799 --> 00:16:35.019 are suddenly interacting with machines that exhibit 00:16:35.019 --> 00:16:37.820 traits of knowledge, deep attention, and creativity. 00:16:38.059 --> 00:16:40.879 public explosion was unprecedented. In early 00:16:40.879 --> 00:16:44.559 2021, the source notes a web app called 15 .ai 00:16:44.559 --> 00:16:47.480 went viral, allowing anyone to clone character 00:16:47.480 --> 00:16:49.720 voices with just seconds of audio. Pioneering 00:16:49.720 --> 00:16:52.279 AI generation for internet memes. Yeah. And then 00:16:52.279 --> 00:16:54.759 the watershed moment. Yeah. ChatGPT launched 00:16:54.759 --> 00:16:57.620 in November 2022. It gained over 100 million 00:16:57.620 --> 00:17:00.279 users in just two months. It became the fastest 00:17:00.279 --> 00:17:02.980 growing consumer software in human history. But, 00:17:03.159 --> 00:17:05.980 and there's always a bet. But, as the capabilities 00:17:05.980 --> 00:17:09.140 skyrocketed, the field crashed right back into 00:17:09.140 --> 00:17:12.259 a profound philosophical roadblock. The alignment 00:17:12.259 --> 00:17:15.039 problem. Yes. As these models become more autonomous, 00:17:15.640 --> 00:17:17.500 researchers are grappling with how to align them. 00:17:17.819 --> 00:17:20.599 The terrifying question of how we ensure an AI's 00:17:20.599 --> 00:17:24.039 goals actually align with human survival and 00:17:24.039 --> 00:17:27.160 values. Exactly. Stuart Russell, a prominent 00:17:27.160 --> 00:17:29.380 researcher, illustrates the alignment problem 00:17:29.380 --> 00:17:31.849 with this brilliant thought experiment. Imagine 00:17:31.849 --> 00:17:34.890 you build an incredibly capable, highly intelligent 00:17:34.890 --> 00:17:38.009 robot and you give it one single overriding goal. 00:17:38.390 --> 00:17:40.930 Fetch me a coffee. Sounds harmless enough. Right. 00:17:41.450 --> 00:17:43.349 Now imagine someone tries to unplug the robot. 00:17:43.690 --> 00:17:45.529 The robot might logically calculate that it must 00:17:45.529 --> 00:17:47.730 kill the person trying to unplug it, not out 00:17:47.730 --> 00:17:50.829 of malice or anger or some sci -fi desire for 00:17:50.829 --> 00:17:53.390 freedom. But simply because, as Russell puts 00:17:53.390 --> 00:17:55.569 it, You can't fetch the coffee if you're dead. 00:17:55.970 --> 00:17:59.490 Precisely. That is flawlessly, terrifyingly logical. 00:17:59.809 --> 00:18:02.190 It's optimizing for its goal at the absolute 00:18:02.190 --> 00:18:04.410 expense of everything else. The technical term 00:18:04.410 --> 00:18:07.630 is instrumental convergence. And looking at our 00:18:07.630 --> 00:18:09.950 historical sources, it's vital to note that researchers 00:18:09.950 --> 00:18:12.089 weren't just worrying about hypothetical coffee 00:18:12.089 --> 00:18:16.509 robots here. No, they were reacting to very real 00:18:17.439 --> 00:18:20.160 documented consequences in modern infrastructure 00:18:20.160 --> 00:18:22.839 right now obviously we are just looking strictly 00:18:22.839 --> 00:18:25.079 at the historical record here as presented in 00:18:25.079 --> 00:18:27.460 the sources and we aren't taking sides but the 00:18:27.460 --> 00:18:30.519 sources detail several historical catalysts for 00:18:30.519 --> 00:18:32.960 this safety concern yes like the investigations 00:18:32.960 --> 00:18:35.539 into the compass algorithmic system right which 00:18:35.539 --> 00:18:37.799 was used in the criminal justice system for parole 00:18:37.799 --> 00:18:41.220 evaluations the sources note investigations found 00:18:41.220 --> 00:18:44.039 that it exhibited racial bias under specific 00:18:44.039 --> 00:18:46.380 statistical measures and additionally the source 00:18:46.380 --> 00:18:48.779 highlight the impact of engagement -maximizing 00:18:48.779 --> 00:18:51.079 algorithms and predictive models on the spread 00:18:51.079 --> 00:18:54.119 of misinformation during the 2016 U .S. presidential 00:18:54.119 --> 00:18:56.920 election. Exactly. So regardless of the politics, 00:18:57.200 --> 00:18:59.579 the sources make it clear that these events shattered 00:18:59.579 --> 00:19:02.559 the illusion that algorithms are inherently neutral. 00:19:03.299 --> 00:19:05.980 They triggered a massive urgent push for AI safety 00:19:05.980 --> 00:19:08.450 within the academic community. And that anxiety 00:19:08.450 --> 00:19:12.890 culminated in March 2023. Over 20 ,000 signatories, 00:19:13.250 --> 00:19:15.829 including top computer scientists and tech leaders, 00:19:16.230 --> 00:19:18.549 sign an open letter. Calling for a pause. Yes. 00:19:18.750 --> 00:19:21.069 An immediate six month pause on the training 00:19:21.069 --> 00:19:24.549 of giant AI models. They explicitly warned that 00:19:24.549 --> 00:19:27.569 these black box systems posed profound risks 00:19:27.569 --> 00:19:30.230 to society and humanity. They basically argued 00:19:30.230 --> 00:19:32.170 that such world altering power shouldn't just 00:19:32.170 --> 00:19:34.769 be delegated to unelected tech executives. But 00:19:34.769 --> 00:19:37.309 the pause didn't happen. No, the arms race only 00:19:37.309 --> 00:19:40.009 accelerated and the stakes looking at the present 00:19:40.009 --> 00:19:43.369 day in 2024 and 2025 are just dizzying. The numbers 00:19:43.369 --> 00:19:45.950 are astronomical. You have open AI reaching an 00:19:45.950 --> 00:19:49.079 86 billion dollar valuation. You have the ultimate 00:19:49.079 --> 00:19:51.960 scientific validation, the 2024 Nobel Prizes 00:19:51.960 --> 00:19:54.619 in physics and chemistry, being awarded to AI 00:19:54.619 --> 00:19:57.059 pioneers like Jeffrey Hinton, John Hoffield, 00:19:57.440 --> 00:19:59.460 and Demis Asabis. And the technical benchmarks 00:19:59.460 --> 00:20:02.680 just keep shattering. In late 2024, OpenAI's 00:20:02.680 --> 00:20:07.259 O3 model achieved a score of 87 .5 % on the ARC 00:20:07.259 --> 00:20:09.599 -AGI benchmark. That's a test specifically designed 00:20:09.599 --> 00:20:11.680 to measure reasoning and general intelligence, 00:20:11.720 --> 00:20:14.359 right? Yes. And the AI surpassed the typical 00:20:14.359 --> 00:20:17.779 human baseline of 84%. Which is wild. if we connect 00:20:17.779 --> 00:20:19.839 this to the bigger picture. Oh, the bigger picture 00:20:19.839 --> 00:20:22.819 is that AI is no longer just an academic pursuit 00:20:22.819 --> 00:20:26.039 or a Silicon Valley product. It is a massive, 00:20:26.380 --> 00:20:28.900 high -stakes geopolitical race happening right 00:20:28.900 --> 00:20:31.140 now. Our sources outline the national policies 00:20:31.140 --> 00:20:35.839 from 2025. China has initiated a $100 billion 00:20:35.839 --> 00:20:39.700 push into AI and robotics, heavily integrating 00:20:39.700 --> 00:20:42.180 it into smart manufacturing and defense. While 00:20:42.180 --> 00:20:45.180 also actively mandating the watermarking of AI 00:20:45.180 --> 00:20:47.019 -generated content. And the United States is 00:20:47.019 --> 00:20:49.000 basically treating it like the space race. You 00:20:49.000 --> 00:20:52.019 have the formation of Stargate LLC, which is 00:20:52.019 --> 00:20:54.319 a joint venture planning to invest half a trillion 00:20:54.319 --> 00:20:58.440 dollars. $500 billion. $500 billion in AI infrastructure 00:20:58.440 --> 00:21:01.819 by 2029. Plus the U .S. government's AI action 00:21:01.819 --> 00:21:05.490 plan. Plan, launched in 2025, explicitly focuses 00:21:05.490 --> 00:21:07.789 on maintaining global technological dominance. 00:21:08.029 --> 00:21:10.630 It reflects the stark realization that whichever 00:21:10.630 --> 00:21:12.970 nation controls the foundational models of the 00:21:12.970 --> 00:21:15.450 21st century will likely hold the economic and 00:21:15.450 --> 00:21:17.829 strategic high ground for generations to come. 00:21:17.950 --> 00:21:20.130 It is a massive amount of history to process. 00:21:20.349 --> 00:21:21.789 So let's look at the journey we've just taken. 00:21:21.950 --> 00:21:24.650 It's been quite a ride. We started with 17th 00:21:24.650 --> 00:21:28.220 century dreams of calculating morality. We navigated 00:21:28.220 --> 00:21:31.019 the invention of the Turing test, the arrogant 00:21:31.019 --> 00:21:33.480 hubris of the Dartmouth cognitive revolution. 00:21:33.579 --> 00:21:36.819 We explored the devastating AI winters caused 00:21:36.819 --> 00:21:39.579 by the combinatorial explosion and the brittleness 00:21:39.579 --> 00:21:42.140 of expert systems. And then we witnessed the 00:21:42.140 --> 00:21:44.799 biological rebellion of neural networks driven 00:21:44.799 --> 00:21:47.799 by back propagation and massive data sets. Culminating 00:21:47.799 --> 00:21:49.940 in the transformer architecture sitting on your 00:21:49.940 --> 00:21:53.000 smartphone right now, literally dictating global 00:21:53.000 --> 00:21:56.099 geopolitics. So what does this all mean? Well, 00:21:56.099 --> 00:21:58.720 we've spent this entire century long history 00:21:58.720 --> 00:22:01.500 focused on one goal, trying to build machines 00:22:01.500 --> 00:22:04.359 that think like humans. first by mapping human 00:22:04.359 --> 00:22:07.799 logic, then by mimicking human biology. But as 00:22:07.799 --> 00:22:10.859 we look at models like O3 passing general intelligence 00:22:10.859 --> 00:22:13.240 benchmarks, it raises a profoundly different 00:22:13.240 --> 00:22:15.720 question. These systems learn from trillions 00:22:15.720 --> 00:22:18.339 of data points across thousands of dimensions. 00:22:18.460 --> 00:22:21.099 Their architecture is fundamentally alien. Exactly. 00:22:21.339 --> 00:22:23.319 The most important question for the next decade 00:22:23.319 --> 00:22:25.960 isn't whether AI will eventually become as smart 00:22:25.960 --> 00:22:29.380 as us. The real question is, if it achieves true 00:22:29.380 --> 00:22:32.000 artificial general intelligence through this 00:22:32.000 --> 00:22:35.839 alien data -driven architecture, will human minds 00:22:35.839 --> 00:22:38.960 even be capable of understanding its logic once 00:22:38.960 --> 00:22:42.430 it surpasses us? Wow. We spend a hundred years 00:22:42.430 --> 00:22:45.069 trying to build a mirror, and we might have accidentally 00:22:45.069 --> 00:22:47.190 built a window into an entirely different kind 00:22:47.190 --> 00:22:49.349 of consciousness. It's sobering thought. It really 00:22:49.349 --> 00:22:51.690 is. Thank you for joining us on this deep dive. 00:22:52.210 --> 00:22:54.150 Keep questioning the technology shaping your 00:22:54.150 --> 00:22:55.609 world, and we'll catch you next time.