WEBVTT 00:00:00.000 --> 00:00:02.240 You know, when you snap a picture of your dog 00:00:02.240 --> 00:00:05.599 and save it to your phone or when you type just 00:00:05.599 --> 00:00:08.119 a quick grocery list into a notes app, there 00:00:08.119 --> 00:00:11.300 is this fundamental illusion at play. Oh, absolutely. 00:00:11.439 --> 00:00:13.320 We tend to think the computer actually knows 00:00:13.320 --> 00:00:15.980 what a golden retriever is or what a gallon of 00:00:15.980 --> 00:00:19.079 milk is, but it doesn't. No, not at all. To the 00:00:19.079 --> 00:00:21.519 machine, it's literally just storing ones and 00:00:21.519 --> 00:00:24.699 zeros. It's essentially just a highly organized 00:00:24.699 --> 00:00:27.379 filing cabinet. Yeah, exactly. Which brings up 00:00:27.379 --> 00:00:29.949 a really massive question for you. to think about 00:00:29.949 --> 00:00:33.490 today as we start this deep dive, how do we actually 00:00:33.490 --> 00:00:36.710 teach computers to understand the world rather 00:00:36.710 --> 00:00:40.549 than just passively store data about it? It is 00:00:40.549 --> 00:00:42.729 arguably the defining question of artificial 00:00:42.729 --> 00:00:45.990 intelligence. It really is. And to answer it, 00:00:45.990 --> 00:00:48.229 we are going to look at a highly detailed source 00:00:48.229 --> 00:00:51.829 today. It's this comprehensive breakdown on knowledge 00:00:51.829 --> 00:00:54.270 representation and reasoning. Right, which you'll 00:00:54.270 --> 00:00:57.929 often see abbreviated as KRR or KR squared. Exactly, 00:00:57.969 --> 00:01:01.009 KR squared. So our mission for this deep dive 00:01:01.009 --> 00:01:04.890 is to completely demystify how AI uses structured 00:01:04.890 --> 00:01:07.969 information to actually solve complex tasks. 00:01:08.549 --> 00:01:10.349 We're going to explore the history of this field, 00:01:10.849 --> 00:01:13.989 the massive, honestly, almost comical hurdles 00:01:13.989 --> 00:01:16.609 of trying to program common sense. Oh, the common 00:01:16.609 --> 00:01:18.829 sense stuff is wild. It's so crazy. And we'll 00:01:18.829 --> 00:01:21.010 get into how all of this is secretly shaping 00:01:21.010 --> 00:01:23.069 the architecture of the Internet you use every 00:01:23.069 --> 00:01:26.170 single day. Yeah. And what becomes clear very 00:01:26.170 --> 00:01:28.109 quickly when you read through this material is 00:01:28.109 --> 00:01:31.750 that Knowledge representation isn't simply about 00:01:31.750 --> 00:01:34.769 writing lines of code. It really merges psychology. 00:01:35.010 --> 00:01:37.250 Like how we as humans naturally solve problems. 00:01:37.390 --> 00:01:39.989 Exactly, how we categorize our reality. It merges 00:01:39.989 --> 00:01:42.030 that with mathematical logic, which is how we 00:01:42.030 --> 00:01:44.629 automate reasoning. So by the time we finish 00:01:44.629 --> 00:01:46.730 today, I really think this will fundamentally 00:01:46.730 --> 00:01:48.609 change how you view every interaction you have 00:01:48.609 --> 00:01:51.870 with a search engine or an AI assistant. Okay, 00:01:51.930 --> 00:01:54.260 let's unpack this. To understand how a computer 00:01:54.260 --> 00:01:56.599 thinks, we have to define what knowledge representation 00:01:56.599 --> 00:01:59.640 actually is. Our source brings up a foundational 00:01:59.640 --> 00:02:03.200 paper from 1993 by an MIT researcher named Randall 00:02:03.200 --> 00:02:06.560 Davis, and he outlined five distinct roles that 00:02:06.560 --> 00:02:09.120 define a knowledge representation framework. 00:02:09.840 --> 00:02:12.379 But the very first role he mentions is the one 00:02:12.379 --> 00:02:14.479 that really anchors the whole concept. Yeah, 00:02:14.639 --> 00:02:16.780 it's the core of it all. He says that a knowledge 00:02:16.780 --> 00:02:19.300 representation is fundamentally a surrogate. 00:02:19.460 --> 00:02:22.300 It is a substitute for the thing itself. And 00:02:22.300 --> 00:02:24.740 the core justification for doing this is that 00:02:24.740 --> 00:02:27.740 you want an entity, in this case a computer system, 00:02:28.180 --> 00:02:31.620 to be able to determine consequences by thinking. 00:02:31.849 --> 00:02:33.949 rather than by acting. Right. You want it to 00:02:33.949 --> 00:02:36.569 reason about the world rather than, you know, 00:02:36.689 --> 00:02:38.710 taking physical action within it to see what 00:02:38.710 --> 00:02:40.870 happens. So is that basically like running a 00:02:40.870 --> 00:02:43.090 physics simulation in a video game before doing 00:02:43.090 --> 00:02:45.129 a stunt in real life? Like the computer gets 00:02:45.129 --> 00:02:48.349 to think about the crash and calculate the trajectory 00:02:48.349 --> 00:02:50.990 and the force without actually crashing a real 00:02:50.990 --> 00:02:53.610 million dollar car. Close. But there's a crucial 00:02:53.610 --> 00:02:55.990 distinction to make there. A video game is mimicking 00:02:55.990 --> 00:02:59.069 physics. Gnarly representation is actually about 00:02:59.069 --> 00:03:01.780 mimicking reasoning itself. Oh. It's less about 00:03:01.780 --> 00:03:04.879 calculating gravity or wind resistance and more 00:03:04.879 --> 00:03:07.400 about the computer deducing why the car crashed 00:03:07.400 --> 00:03:10.370 based on facts it holds. The computer builds 00:03:10.370 --> 00:03:13.629 a surrogate reality of logic, not just a surrogate 00:03:13.629 --> 00:03:17.750 reality of physics, but building that surrogate 00:03:17.750 --> 00:03:20.689 introduces what early AI researchers realized 00:03:20.689 --> 00:03:23.289 was the ultimate balancing act in this field. 00:03:23.750 --> 00:03:26.469 It's this massive trade -off between expressivity 00:03:26.469 --> 00:03:29.610 and tractability. Expressivity and tractability. 00:03:29.710 --> 00:03:31.990 Let me try another analogy here to see if I grasp 00:03:31.990 --> 00:03:34.909 that trade -off. It feels a lot like designing 00:03:34.909 --> 00:03:37.409 a map. Okay. If you use something like First 00:03:37.409 --> 00:03:39.250 Order Logic, which our source mentions is the 00:03:39.250 --> 00:03:41.229 gold standard for high expressivity because it 00:03:41.229 --> 00:03:43.750 can formalize basically all of mathematics. It's 00:03:43.750 --> 00:03:45.889 incredibly powerful. It's like having a one -to 00:03:45.889 --> 00:03:48.789 -one scale map of a city. It includes every single 00:03:48.789 --> 00:03:50.650 blade of grass, every pebble on the street. It 00:03:50.650 --> 00:03:53.150 is perfectly accurate. It is perfectly expressive. 00:03:53.770 --> 00:03:56.449 But it is entirely useless to actually carry 00:03:56.449 --> 00:03:58.889 around and read. It's intractable. That captures 00:03:58.889 --> 00:04:02.050 the problem perfectly. First Order Logic is amazing. 00:04:02.280 --> 00:04:05.219 But because of its high expressive power, it 00:04:05.219 --> 00:04:07.379 allows for countless different ways of expressing 00:04:07.379 --> 00:04:10.379 the exact same piece of information. Wow. So 00:04:10.379 --> 00:04:12.539 when you feed that into a computer, the machine 00:04:12.539 --> 00:04:14.860 spends a massive amount of computational power 00:04:14.860 --> 00:04:17.540 just translating and managing all those different 00:04:17.540 --> 00:04:19.519 expressions. Instead of actually solving the 00:04:19.519 --> 00:04:21.860 problem. Exactly. It becomes intimidating for 00:04:21.860 --> 00:04:24.399 human software developers to write and incredibly 00:04:24.399 --> 00:04:26.939 inefficient for computers to process. You simply 00:04:26.939 --> 00:04:28.980 don't need the full power of first -order logic 00:04:28.980 --> 00:04:31.540 to get most jobs done. Right. So many of the 00:04:31.540 --> 00:04:34.160 early AI knowledge representation formalisms 00:04:34.160 --> 00:04:37.089 were basically just design decisions. They were 00:04:37.089 --> 00:04:39.009 trying to figure out how to balance that expressive 00:04:39.009 --> 00:04:43.269 power with actual workable efficiency. So if 00:04:43.269 --> 00:04:45.889 first -order logic was too heavy and intractable 00:04:45.889 --> 00:04:48.689 for computers to actually use back then, what 00:04:48.689 --> 00:04:52.350 did early AI pioneers do? I mean, they couldn't 00:04:52.350 --> 00:04:54.550 just abandon the idea of a digital brain, so 00:04:54.550 --> 00:04:56.230 they must have started out with just massive 00:04:56.230 --> 00:04:58.870 sweeping ambitions to try and solve this. Oh, 00:04:58.910 --> 00:05:02.040 absolutely. If we go back to 1959, You see the 00:05:02.040 --> 00:05:04.660 earliest work focusing on exactly that, general 00:05:04.660 --> 00:05:06.899 problem solvers. Like universal brains. Yeah. 00:05:07.360 --> 00:05:09.600 You had Alan Newell and Herbert A. Simon developing 00:05:09.600 --> 00:05:12.019 a system literally called the general problem 00:05:12.019 --> 00:05:15.560 solver, or GPS. VPS, right. It featured data 00:05:15.560 --> 00:05:18.279 structures designed entirely for planning. It 00:05:18.279 --> 00:05:20.939 would take a large goal, decompose it into smaller 00:05:20.939 --> 00:05:23.300 sub -goals, and then construct strategies to 00:05:23.300 --> 00:05:26.279 accomplish each one. And in that exact same year, 00:05:26.779 --> 00:05:29.439 John McCarthy proposed the Advice Taker, which 00:05:29.439 --> 00:05:31.800 attempted to implement common sense reasoning 00:05:31.800 --> 00:05:35.139 using formal logic. But reading through the history 00:05:35.139 --> 00:05:37.879 here, the field was incredibly fractured on how 00:05:37.879 --> 00:05:40.319 to actually achieve this. Like in the 1970s, 00:05:40.319 --> 00:05:43.439 there was a massive turf war. Huge. On one side, 00:05:43.500 --> 00:05:46.019 you had purists who wanted AI to solve problems 00:05:46.019 --> 00:05:48.920 using pure mathematical logic. But on the other 00:05:48.920 --> 00:05:51.100 side, researchers at MIT argued that knowledge 00:05:51.100 --> 00:05:53.360 isn't just math. It's procedures. It's a set 00:05:53.360 --> 00:05:55.899 of instructions. Right. The conflict between 00:05:55.899 --> 00:05:58.800 logical representations and procedural representations. 00:05:59.800 --> 00:06:02.860 The MIT researchers completely rejected the idea. 00:06:02.959 --> 00:06:05.879 that a computer should just look at a giant database 00:06:05.879 --> 00:06:08.980 of facts and mathematically deduce an answer. 00:06:09.160 --> 00:06:11.319 They thought that was the wrong approach. Totally. 00:06:11.600 --> 00:06:14.180 They advocated for the procedural embedding of 00:06:14.180 --> 00:06:17.240 knowledge. To put it simply, they believed knowledge 00:06:17.240 --> 00:06:19.740 should be encoded as a set of rules or instructions. 00:06:19.779 --> 00:06:22.500 Right. Think of it like baking a cake. A logical 00:06:22.500 --> 00:06:24.519 representation would just give the computer the 00:06:24.519 --> 00:06:27.180 chemical properties of flour, sugar, and heat 00:06:27.180 --> 00:06:30.290 and expect it to deduce a cake. Which is insane 00:06:30.290 --> 00:06:32.870 for a computer to do. Right, whereas a procedural 00:06:32.870 --> 00:06:35.829 representation says, step one, mix flour and 00:06:35.829 --> 00:06:39.110 sugar. Step two, apply heat. That makes so much 00:06:39.110 --> 00:06:41.089 more sense. And this debate was eventually settled 00:06:41.089 --> 00:06:43.310 by the development of logic programming languages 00:06:43.310 --> 00:06:46.029 like Prolog, which managed to kind of blend these 00:06:46.029 --> 00:06:48.149 two approaches. But wait, I have a pointed question 00:06:48.149 --> 00:06:51.370 here. If early AI pioneers were trying to build 00:06:51.370 --> 00:06:54.769 a literal general problem solver, why did they 00:06:54.769 --> 00:06:57.709 suddenly pivot in the 1970s and 80s to hyper 00:06:57.709 --> 00:07:00.910 -specific expert systems. Wasn't that a massive 00:07:00.910 --> 00:07:03.670 step backward in ambition, like abandoning the 00:07:03.670 --> 00:07:06.129 dream of a general AI to build something that 00:07:06.129 --> 00:07:08.870 only does one specific task? What's fascinating 00:07:08.870 --> 00:07:11.329 here is that it wasn't a retreat at all, though 00:07:11.329 --> 00:07:13.029 I know it might look like one on the surface. 00:07:13.089 --> 00:07:15.689 It definitely looks like one. Sure, but it was 00:07:15.689 --> 00:07:18.569 a highly practical realization driven by the 00:07:18.569 --> 00:07:22.420 cognitive revolution in psychology. Researchers 00:07:22.420 --> 00:07:25.560 like Ed Feigenbaum in North America realized 00:07:25.560 --> 00:07:27.980 that if you actually wanted an AI to match human 00:07:27.980 --> 00:07:31.100 competence on a specific task, like diagnosing 00:07:31.100 --> 00:07:33.019 a medical condition, general purpose reasoning 00:07:33.019 --> 00:07:35.779 just wasn't enough. Exactly. Humans don't solve 00:07:35.779 --> 00:07:38.579 complex medical problems just by being generally 00:07:38.579 --> 00:07:40.980 smart. They solve them because they went to medical 00:07:40.980 --> 00:07:43.920 school and acquired thousands of highly specific 00:07:43.920 --> 00:07:47.100 facts. Right. So to make an AI useful, broad 00:07:47.100 --> 00:07:50.339 and shallow rules had to lose out to narrow and 00:07:50.339 --> 00:07:53.279 deep knowledge. So the AI field shifted its focus 00:07:53.279 --> 00:07:55.699 from building a universal brain to building highly 00:07:55.699 --> 00:07:58.860 educated specialists. Yes. And this pivot to 00:07:58.860 --> 00:08:01.459 expert systems gave us the architectural terminology 00:08:01.459 --> 00:08:04.300 that AI still uses today. They divided these 00:08:04.300 --> 00:08:07.220 systems into two distinct parts. First, the knowledge 00:08:07.220 --> 00:08:10.000 base. which holds the raw facts and rules about 00:08:10.000 --> 00:08:13.040 a specific problem domain. And second, the inference 00:08:13.040 --> 00:08:15.579 engine, which is the mechanism that takes the 00:08:15.579 --> 00:08:18.379 knowledge in that base and applies it to answer 00:08:18.379 --> 00:08:20.699 questions or solve problems. So they literally 00:08:20.699 --> 00:08:23.579 split the storage of facts from the logic of 00:08:23.579 --> 00:08:25.980 how to use those facts. Precisely. That makes 00:08:25.980 --> 00:08:28.980 complete sense for something bounded like diagnosing 00:08:28.980 --> 00:08:31.839 a blood disease. You give the knowledge base 00:08:31.839 --> 00:08:34.240 all the symptoms of blood diseases and the inference 00:08:34.240 --> 00:08:36.580 engine matches a patient's symptoms to the right 00:08:36.580 --> 00:08:39.139 disease. But the source points out that while 00:08:39.139 --> 00:08:41.740 expert systems were amazing in a laboratory for 00:08:41.740 --> 00:08:44.740 those specific tasks, they failed completely 00:08:44.740 --> 00:08:47.840 at everyday human life. Well, yeah. Medical diagnosis 00:08:47.840 --> 00:08:50.240 is complex, but it's bounded by the walls of 00:08:50.240 --> 00:08:52.519 a hospital and the biology of the human body. 00:08:52.750 --> 00:08:55.409 Everyday life is completely unbounded. Right. 00:08:55.710 --> 00:08:57.490 Which brings us to the challenge of encoding 00:08:57.490 --> 00:09:00.250 common sense. Here's where it gets really interesting. 00:09:00.769 --> 00:09:03.649 In the mid -1970s, Marvin Minsky developed a 00:09:03.649 --> 00:09:06.409 concept called frames to try and tackle this 00:09:06.409 --> 00:09:09.090 exact problem. Yes, frames are so important here. 00:09:09.330 --> 00:09:11.690 A frame is described as an abstract description 00:09:11.690 --> 00:09:15.149 of a category. It has these slots for data and 00:09:15.149 --> 00:09:17.090 constraints on what can actually fit in those 00:09:17.090 --> 00:09:19.610 slots. And the example in the source that really 00:09:19.610 --> 00:09:22.509 helped me visualize this was ordering food in 00:09:22.509 --> 00:09:24.909 a restaurant. The restaurant frame is just a 00:09:24.909 --> 00:09:27.870 fantastic way to understand how we try to take 00:09:27.870 --> 00:09:30.529 computers context. So a frame is basically like 00:09:30.529 --> 00:09:33.470 a social Mad Libs template. If I walk into a 00:09:33.470 --> 00:09:36.509 restaurant, the restaurant frame activates in 00:09:36.509 --> 00:09:39.850 the AI. It automatically knows to expect a menu, 00:09:40.110 --> 00:09:43.129 a waiter, and food. Exactly. It instantly narrows 00:09:43.129 --> 00:09:45.450 the search space so the computer isn't suddenly 00:09:45.450 --> 00:09:47.870 preparing for a medical exam or trying to solve 00:09:47.870 --> 00:09:50.370 a calculus problem. But what happens if something 00:09:50.370 --> 00:09:53.370 unexpected occurs? What if the waiter drops a 00:09:53.370 --> 00:09:56.509 plate? That is exactly why frames are so revolutionary. 00:09:56.720 --> 00:09:59.200 When the waiter drops the plate, the AI doesn't 00:09:59.200 --> 00:10:01.840 have to relearn what gravity is or figure out 00:10:01.840 --> 00:10:04.500 if the waiter spontaneously exploded. Because 00:10:04.500 --> 00:10:07.139 the restaurant frame is loaded with default expectations 00:10:07.139 --> 00:10:10.440 about physics and human behavior, it can constrain 00:10:10.440 --> 00:10:13.330 the possibilities. The AI understands that the 00:10:13.330 --> 00:10:15.950 waiter slot is still filled by a human, and the 00:10:15.950 --> 00:10:18.429 plate slot just changed from intact to broken 00:10:18.429 --> 00:10:21.610 due to gravity. Wow. And by the early 1980s, 00:10:21.870 --> 00:10:24.590 communities realized they could combine the real 00:10:24.590 --> 00:10:27.710 -world representation of frames with the complex 00:10:27.710 --> 00:10:31.129 logic of rule -based systems. Merging them together. 00:10:31.330 --> 00:10:33.710 Yeah. In 1983, a system called the Knowledge 00:10:33.710 --> 00:10:36.889 Engineering Environment, or KEY, did exactly 00:10:36.889 --> 00:10:39.759 that. It had a complete rule engine to handle 00:10:39.759 --> 00:10:42.519 logic, but also a frame -based knowledge base 00:10:42.519 --> 00:10:44.980 to handle the messy reality of categories and 00:10:44.980 --> 00:10:47.639 objects. But even with all that power, they kept 00:10:47.639 --> 00:10:49.860 slamming into the common sense reasoning problem. 00:10:50.090 --> 00:10:52.350 The source mentions the frame problem, which 00:10:52.350 --> 00:10:54.049 is just mind -boggling when you think about it. 00:10:54.129 --> 00:10:56.149 It really is. How do you teach a computer that 00:10:56.149 --> 00:10:58.610 an object maintains its exact position from one 00:10:58.610 --> 00:11:01.149 moment to the next unless it is moved by an external 00:11:01.149 --> 00:11:03.190 force? Right, this raises an important question. 00:11:03.450 --> 00:11:06.289 How do you explicitly code the infinite unwritten 00:11:06.289 --> 00:11:09.009 rules of human existence? It seems impossible. 00:11:09.389 --> 00:11:12.730 As humans, we regularly draw on an extensive, 00:11:13.090 --> 00:11:15.649 invisible foundation of knowledge about the real 00:11:15.649 --> 00:11:19.129 world. We understand basic physics. We know that 00:11:19.129 --> 00:11:21.210 if you put a cup on a table, it stays there. 00:11:21.690 --> 00:11:24.909 We know that water flows downhill. We understand 00:11:24.909 --> 00:11:28.149 causality and human intentions. We take it entirely 00:11:28.149 --> 00:11:30.250 for granted because we've been interacting with 00:11:30.250 --> 00:11:32.690 the physical world since we were babies. But 00:11:32.690 --> 00:11:35.649 to a computer... To an artificial agent, a digital 00:11:35.649 --> 00:11:38.190 surrogate that has never felt gravity or dropped 00:11:38.190 --> 00:11:40.629 a cup... None of that is obvious. It all has 00:11:40.629 --> 00:11:42.889 to be explicitly stated. And someone actually 00:11:42.889 --> 00:11:45.250 tried to write it all down. Doug Lean it with 00:11:45.250 --> 00:11:47.769 The Psych Project. The Psych Project spelled 00:11:47.769 --> 00:11:50.769 C -Y -C. It was one of the most ambitious programs 00:11:50.769 --> 00:11:53.210 in AI history. They created their own language. 00:11:53.399 --> 00:11:56.700 cycle and hired armies of analysts to manually 00:11:56.700 --> 00:11:59.019 document various areas of common sense reasoning. 00:11:59.340 --> 00:12:01.320 Just taping out common sense. Literally. Yeah. 00:12:01.419 --> 00:12:03.600 They tried to manually type out models of time, 00:12:03.840 --> 00:12:06.220 causality, physics, intentions. They were writing 00:12:06.220 --> 00:12:08.360 rules like you cannot be in two places at the 00:12:08.360 --> 00:12:11.200 same time or once you consume a liquid, it is 00:12:11.200 --> 00:12:13.919 no longer in the glass. That is why. It was a 00:12:13.919 --> 00:12:17.799 monumental, seemingly endless task to type out 00:12:17.799 --> 00:12:20.940 reality. But here's my logical next step on this. 00:12:21.019 --> 00:12:23.759 If one guy and his team spent decades just trying 00:12:23.759 --> 00:12:25.860 to type out the basic physical rules of the world, 00:12:26.320 --> 00:12:29.299 how do different systems or even different human 00:12:29.299 --> 00:12:33.000 researchers ever agree on what the rules actually 00:12:33.000 --> 00:12:36.080 are? Like if I build a database for my company 00:12:36.080 --> 00:12:38.940 and you build one for yours, how do we ever share 00:12:38.940 --> 00:12:41.419 this knowledge? That exact problem gave rise 00:12:41.419 --> 00:12:44.559 to the discipline of ontology engineering. As 00:12:44.559 --> 00:12:47.159 knowledge bases scaled up, researchers needed 00:12:47.159 --> 00:12:49.730 them to be modular. They needed different systems 00:12:49.730 --> 00:12:52.370 to communicate with one another. But defining 00:12:52.370 --> 00:12:54.870 the boundaries of concepts is incredibly subjective. 00:12:55.070 --> 00:12:57.590 The source includes a quote from Tom Gruber that 00:12:57.590 --> 00:13:00.330 summarizes this beautifully. He says, every ontology 00:13:00.330 --> 00:13:02.710 is a treaty, a social agreement among people 00:13:02.710 --> 00:13:05.700 with common motive and sharing. He argues that 00:13:05.700 --> 00:13:08.580 a truly general purpose universal ontology is 00:13:08.580 --> 00:13:10.899 impossible. So what does this all mean? I have 00:13:10.899 --> 00:13:12.519 a hard time wrapping my head around that. If 00:13:12.519 --> 00:13:14.879 every ontology is just a treaty or a social agreement, 00:13:15.080 --> 00:13:17.059 does that mean there is no objective truth in 00:13:17.059 --> 00:13:19.799 computer logic? Let me give you an example. Is 00:13:19.799 --> 00:13:23.500 a tomato a fruit or a vegetable? If I'm building 00:13:23.500 --> 00:13:26.360 a system for a botanist, the tomato is objectively 00:13:26.360 --> 00:13:28.980 a fruit. Right. If I'm building it for a chef, 00:13:29.059 --> 00:13:32.980 it's a vegetable. Are we literally just teaching 00:13:32.980 --> 00:13:35.720 computers our subjective human perspectives? 00:13:36.139 --> 00:13:39.080 If we connect this to the bigger picture, mathematical 00:13:39.080 --> 00:13:41.539 logic, the ones and zeros, the inference engine 00:13:41.539 --> 00:13:46.220 has objective truth. But categorizing the world, 00:13:46.500 --> 00:13:48.740 deciding how to represent reality within the 00:13:48.740 --> 00:13:51.600 knowledge base, that requires perspective and 00:13:51.600 --> 00:13:54.399 bias. Interesting. The source gives two stark 00:13:54.399 --> 00:13:57.720 examples of how the exact same objective reality 00:13:57.720 --> 00:14:00.039 can be represented in completely different ways, 00:14:00.480 --> 00:14:02.399 depending on what the user needs the system to 00:14:02.399 --> 00:14:04.539 do. Oh, the electronic circuits example really 00:14:04.539 --> 00:14:06.879 highlighted this for me. Yeah. If you were representing 00:14:06.879 --> 00:14:09.059 an electronic circuit, you might use what's called 00:14:09.059 --> 00:14:11.570 the lumped element model. In this view, you see 00:14:11.570 --> 00:14:13.750 the circuit purely in terms of components and 00:14:13.750 --> 00:14:16.049 connections, and you assume electrical signals 00:14:16.049 --> 00:14:18.669 flow instantaneously between them. It's a simple, 00:14:18.710 --> 00:14:21.370 highly useful treaty for most basic electronics. 00:14:21.690 --> 00:14:23.549 But what if you need more detail? Right. If you 00:14:23.549 --> 00:14:26.070 need to pay attention to electrodynamics, maybe 00:14:26.070 --> 00:14:27.970 you're designing a high -speed computer processor, 00:14:28.409 --> 00:14:31.870 that ontology is totally useless. In the electrodynamic 00:14:31.870 --> 00:14:35.669 view, signals propagate at a finite speed. A 00:14:35.669 --> 00:14:37.769 simple resistor is no longer just a component. 00:14:38.289 --> 00:14:40.649 It's an extended medium that an electromagnetic 00:14:40.649 --> 00:14:43.370 wave flows through. But the physical reality 00:14:43.370 --> 00:14:45.210 of the circuit on the table hasn't changed at 00:14:45.210 --> 00:14:48.110 all. Exactly. Copper wire is still copper wire. 00:14:48.860 --> 00:14:51.639 But the representation, the treaty we agreed 00:14:51.639 --> 00:14:54.100 upon to make the math work has completely shifted. 00:14:54.379 --> 00:14:56.559 And the exact same dynamic played out in early 00:14:56.559 --> 00:14:58.779 medicine. We talked about expert systems earlier. 00:14:59.159 --> 00:15:02.820 A famous early system called amycin viewed medicine 00:15:02.820 --> 00:15:06.200 as a set of empirical rules. OK. It created simple 00:15:06.200 --> 00:15:08.799 associations connecting a specific symptom to 00:15:08.799 --> 00:15:12.139 a specific disease. But another system called 00:15:12.139 --> 00:15:15.240 internist view the exact same medical field as 00:15:15.240 --> 00:15:17.639 a set of prototypes. Wait, how is that different? 00:15:17.820 --> 00:15:20.179 It tried to match the patient's entire case against 00:15:20.179 --> 00:15:22.620 prototypical diseases rather than just symptom 00:15:22.620 --> 00:15:25.059 to disease rules. There are two totally different 00:15:25.059 --> 00:15:27.259 ways to encode the same objective reality of 00:15:27.259 --> 00:15:29.759 human biology. And this isn't just ancient history 00:15:29.759 --> 00:15:33.179 for massive room size mainframes. This concept 00:15:33.179 --> 00:15:36.539 of ontologies and treaties is literally the architecture 00:15:36.539 --> 00:15:39.179 of the modern Internet. Yes. The source talks 00:15:39.179 --> 00:15:41.519 about the semantic web, and this is where it 00:15:41.519 --> 00:15:44.419 really connects to your daily life as a listener. 00:15:44.590 --> 00:15:47.870 The semantic web is the ongoing attempt to add 00:15:47.870 --> 00:15:50.870 a layer of actual meaning semantics on top of 00:15:50.870 --> 00:15:52.549 the current internet. Because right now it's 00:15:52.549 --> 00:15:55.039 mostly just text matching, right? Exactly. Traditional 00:15:55.039 --> 00:15:58.019 search engines primarily index websites by matching 00:15:58.019 --> 00:16:01.360 text keywords. If you type the word Apple, the 00:16:01.360 --> 00:16:03.539 traditional search engine looks for web pages 00:16:03.539 --> 00:16:06.220 that contain the letters A -P -P -L -E. Just 00:16:06.220 --> 00:16:08.879 the string of characters. Right. But the semantic 00:16:08.879 --> 00:16:12.240 web uses large ontologies of concepts. It relies 00:16:12.240 --> 00:16:14.779 on markup languages like RDF, which stands for 00:16:14.779 --> 00:16:16.639 Resource Description Framework. And what does 00:16:16.639 --> 00:16:19.799 RDF actually do? RDF creates a structure of subject, 00:16:20.019 --> 00:16:22.240 predicate, and object. It tells the computer, 00:16:22.320 --> 00:16:26.419 Apple is a fruit. or Apple is a technology company. 00:16:26.659 --> 00:16:29.320 Oh, wow. So when I search Apple, the computer 00:16:29.320 --> 00:16:31.059 isn't just looking for letters anymore. It looks 00:16:31.059 --> 00:16:34.360 at the ontology and effectively asks, wait, is 00:16:34.360 --> 00:16:36.679 the user's current frame technology or fruit? 00:16:37.100 --> 00:16:40.440 It uses logic to deduce the context. The semantic 00:16:40.440 --> 00:16:44.070 web... also uses OWL, the Web Ontology Language, 00:16:44.529 --> 00:16:46.889 to add complex logic to those relationships. 00:16:47.549 --> 00:16:50.129 And it uses automated reasoning engines, called 00:16:50.129 --> 00:16:53.149 classifiers, to constantly update the Internet's 00:16:53.149 --> 00:16:55.750 ontology on the fly as new information gets added. 00:16:55.980 --> 00:16:58.740 So the internet is slowly moving from a giant 00:16:58.740 --> 00:17:01.240 filing cabinet of text strings into a structured 00:17:01.240 --> 00:17:03.480 web that actually understands the relationships 00:17:03.480 --> 00:17:06.579 between concepts. Exactly. But that entire web 00:17:06.579 --> 00:17:09.519 of meaning, all those classifications, are governed 00:17:09.519 --> 00:17:11.779 by those human treaties we talked about. The 00:17:11.779 --> 00:17:14.099 computer only knows a tomato is a fruit if the 00:17:14.099 --> 00:17:16.380 treaty we wrote says so. It's all built on the 00:17:16.380 --> 00:17:18.200 perspectives we choose to code into it. That 00:17:18.200 --> 00:17:20.809 is just incredible. To summarize our journey 00:17:20.809 --> 00:17:22.950 today for you listening, we started with the 00:17:22.950 --> 00:17:25.750 massive dreams of the 1950s, trying to build 00:17:25.750 --> 00:17:27.910 general problem solvers that could figure out 00:17:27.910 --> 00:17:30.670 anything. Then we saw the practical computational 00:17:30.670 --> 00:17:32.950 trade -offs that forced the field to pivot to 00:17:32.950 --> 00:17:35.970 expert systems, trading broad logic for deep 00:17:35.970 --> 00:17:39.490 specialized domain knowledge. We explored the 00:17:39.490 --> 00:17:42.250 quirky, incredibly complex world of teaching 00:17:42.250 --> 00:17:45.230 computers common sense using frames so they know 00:17:45.230 --> 00:17:47.250 how to navigate something as simple as a restaurant 00:17:47.250 --> 00:17:51.420 without breaking down. Exactly. And finally, 00:17:51.779 --> 00:17:54.539 we arrived at ontology engineering, realizing 00:17:54.539 --> 00:17:57.740 that the very fabric of the semantic web is bound 00:17:57.740 --> 00:18:00.960 by digital treaties, social agreements on how 00:18:00.960 --> 00:18:03.099 we choose to represent our world to the machines 00:18:03.099 --> 00:18:05.400 we build. Which leaves us with a really profound 00:18:05.400 --> 00:18:08.859 implication to consider. If every knowledge representation 00:18:08.859 --> 00:18:12.230 is fundamentally a treaty, a specific lens chosen 00:18:12.230 --> 00:18:14.789 to view the world, like viewing a circuit as 00:18:14.789 --> 00:18:17.210 simple connections rather than complex waves, 00:18:17.990 --> 00:18:20.029 what happens to the human knowledge that doesn't 00:18:20.029 --> 00:18:22.569 fit neatly into these digital treaties? Oh, that's 00:18:22.569 --> 00:18:24.950 a heavy thought. As our machines become the primary 00:18:24.950 --> 00:18:27.470 way we store, retrieve, and interact with information, 00:18:27.869 --> 00:18:30.329 Are we slowly reshaping human thought to only 00:18:30.329 --> 00:18:32.250 fit the structures that computers can easily 00:18:32.250 --> 00:18:35.549 process? That is a brilliant and slightly terrifying 00:18:35.549 --> 00:18:37.490 question for you to mull over as you go about 00:18:37.490 --> 00:18:40.269 your day. Thank you so much for joining us on 00:18:40.269 --> 00:18:43.089 this deep dive into the architecture of artificial 00:18:43.089 --> 00:18:46.170 thought. Keep questioning the digital world around 00:18:46.170 --> 00:18:46.410 you.