WEBVTT 00:00:00.000 --> 00:00:03.220 Imagine trying to, like, actually bottle human 00:00:03.220 --> 00:00:06.879 intuition, just cashing the exact thought process 00:00:06.879 --> 00:00:10.160 of a world -class doctor diagnosing some rare 00:00:10.160 --> 00:00:13.380 blood disease. Or a master chemist, you know, 00:00:13.500 --> 00:00:15.699 deducing the structure of an unknown organic 00:00:15.699 --> 00:00:18.719 molecule from just a massive pile of raw data. 00:00:18.940 --> 00:00:21.320 Yeah, exactly. Taking those rapid subconscious 00:00:21.320 --> 00:00:23.679 leaps of logic that happen inside an expert's 00:00:23.679 --> 00:00:26.879 brain, freezing them into code, and then letting 00:00:26.879 --> 00:00:29.079 a machine just replicate that exact brilliance 00:00:29.079 --> 00:00:31.179 on command. And doing it decades before anyone 00:00:31.179 --> 00:00:33.420 had ever even heard of deep learning or massive 00:00:33.420 --> 00:00:35.320 language models. Right. Well, welcome to the 00:00:35.320 --> 00:00:37.380 deep dive. Today we are exploring the history, 00:00:37.759 --> 00:00:40.039 the mechanics, and the kind of hidden legacy 00:00:40.039 --> 00:00:42.500 of expert systems. Yeah. If you're a learner 00:00:42.500 --> 00:00:44.679 who's obsessing over modern artificial neural 00:00:44.679 --> 00:00:46.700 networks, you really need to hear this because 00:00:46.700 --> 00:00:49.780 we are looking at what is arguably the first 00:00:49.780 --> 00:00:53.259 truly successful form of AI software. It's just 00:00:53.259 --> 00:00:55.039 a fascinating journey. We're going to look at 00:00:55.039 --> 00:00:57.359 how early computer scientists pioneered this 00:00:57.359 --> 00:00:59.719 concept, why it completely revolutionized the 00:00:59.719 --> 00:01:02.359 corporate world in the 1980s, and well, whether 00:01:02.359 --> 00:01:04.780 these systems actually died out during the infamous 00:01:04.780 --> 00:01:07.560 AI winter. Or if they just went undercover. Exactly. 00:01:07.980 --> 00:01:10.239 Okay, let's unpack this. Before we get into the 00:01:10.239 --> 00:01:12.280 corporate espionage and the AI winter and all 00:01:12.280 --> 00:01:14.459 that, we really need to understand the anatomy 00:01:14.459 --> 00:01:17.439 of these mechanical minds. The core concept here 00:01:17.439 --> 00:01:20.650 is emulating human decision -making Using if 00:01:20.650 --> 00:01:23.250 -then rules right instead of conventional procedural 00:01:23.250 --> 00:01:26.310 programming, right? Yeah, so how does that architecture 00:01:26.310 --> 00:01:28.430 actually work under the hood? So fundamentally 00:01:28.430 --> 00:01:31.689 an expert system is divided into two main subsystems 00:01:31.689 --> 00:01:33.430 First you have what's called the knowledge base. 00:01:33.549 --> 00:01:35.870 This is where all the facts and the rules about 00:01:35.870 --> 00:01:38.709 the specific world are stored Okay, and early 00:01:38.709 --> 00:01:42.579 on these were just you know flat standalone assertions. 00:01:42.939 --> 00:01:46.420 But as the systems grew, they evolved into object 00:01:46.420 --> 00:01:49.659 -oriented concepts. So think of classes, subclasses, 00:01:49.859 --> 00:01:51.799 instances. Wait, hold on. Unpack that object 00:01:51.799 --> 00:01:53.959 -oriented shift for me. Why do they need classes 00:01:53.959 --> 00:01:56.239 instead of just like a big list of facts? Well, 00:01:56.280 --> 00:01:58.920 because a flat list takes up massive amounts 00:01:58.920 --> 00:02:01.519 of memory and processing power. Oh, sure. If 00:02:01.519 --> 00:02:03.939 you have to read a specific separate rule stating 00:02:03.939 --> 00:02:06.599 that a sedan has four wheels and then another 00:02:06.599 --> 00:02:09.379 rule that a truck has four wheels and a van has 00:02:09.379 --> 00:02:11.280 four wheels. You're just wasting space. You're 00:02:11.280 --> 00:02:14.379 wasting so much space. By creating a vehicle 00:02:14.379 --> 00:02:18.240 class with the baseline rule of has four wheels, 00:02:18.599 --> 00:02:21.400 you can just create subclasses that inherit those 00:02:21.400 --> 00:02:23.759 traits automatically. Right. It made the knowledge 00:02:23.759 --> 00:02:26.819 base infinitely more efficient. and critically 00:02:26.819 --> 00:02:29.319 scalable. OK, so the knowledge base holds the 00:02:29.319 --> 00:02:31.960 rules and the facts. What's the second subsystem, 00:02:32.120 --> 00:02:34.460 then? That would be the inference engine. This 00:02:34.460 --> 00:02:37.020 is the automated reasoning system. It basically 00:02:37.020 --> 00:02:38.800 looks at the current state of the knowledge base, 00:02:39.319 --> 00:02:41.560 applies the relevant rules to the known facts, 00:02:41.960 --> 00:02:44.580 and deduces entirely new facts. It sounds like, 00:02:44.919 --> 00:02:47.080 well, if I'm picturing this, traditional code 00:02:47.080 --> 00:02:49.020 is like a recipe you have to follow step by step, 00:02:49.159 --> 00:02:51.139 right? Yeah, imperative instructions. Right. 00:02:51.610 --> 00:02:53.849 But an expert system is more like a detective 00:02:53.849 --> 00:02:56.509 walking into a room full of clues and just applying 00:02:56.509 --> 00:02:59.349 logic. Or, actually, maybe it's like a giant 00:02:59.349 --> 00:03:01.909 plinko board. A plinko board? Yeah, like the 00:03:01.909 --> 00:03:03.810 knowledge base forms all the pegs, the facts, 00:03:03.870 --> 00:03:06.490 and the rules. And the inference engine is the 00:03:06.490 --> 00:03:08.710 puck dropping through. Every time it hits a peg, 00:03:09.129 --> 00:03:12.689 it's forced down a specific if -then branch until 00:03:12.689 --> 00:03:14.750 it eventually lands in a conclusion at the bottom. 00:03:14.960 --> 00:03:17.740 That is actually a really brilliant way to visualize 00:03:17.740 --> 00:03:20.419 it. And that puck, the inference engine, can 00:03:20.419 --> 00:03:22.259 actually drop through the board in two different 00:03:22.259 --> 00:03:24.719 ways, forward chaining and backward chaining. 00:03:24.919 --> 00:03:27.300 Okay, break those down for me. Let's use a classic 00:03:27.300 --> 00:03:29.500 logic puzzle from the source material, the Socrates 00:03:29.500 --> 00:03:33.509 example. In forward chaining, the system is totally 00:03:33.509 --> 00:03:36.710 data -driven. You feed it a fact, like Socrates 00:03:36.710 --> 00:03:39.449 is a man, the inference engine hits the rule 00:03:39.449 --> 00:03:42.990 peg that says, if man, then mortal, and it asserts 00:03:42.990 --> 00:03:46.210 a new conclusion. Socrates is mortal. So it moves 00:03:46.210 --> 00:03:48.330 forward from the data to the conclusion. Data 00:03:48.330 --> 00:03:50.949 in, conclusion out. Very straightforward. Exactly. 00:03:51.030 --> 00:03:52.689 But backward chaining sounds like the system 00:03:52.689 --> 00:03:55.889 is working in reverse. Wait, I'm actually a bit 00:03:55.889 --> 00:03:57.870 confused. How does the system even know what 00:03:57.870 --> 00:03:59.750 goal to start with? Does the user have to prompt 00:03:59.750 --> 00:04:03.650 it? Yes, backward chaining is goal -driven. The 00:04:03.650 --> 00:04:05.909 user, or maybe another part of the program, starts 00:04:05.909 --> 00:04:09.449 with a specific question, like, is Socrates mortal? 00:04:09.810 --> 00:04:12.569 The system looks at its rules and says, well, 00:04:12.750 --> 00:04:15.169 to prove he is mortal, I first need to establish 00:04:15.169 --> 00:04:18.110 if he is a man. So it queries its knowledge base. 00:04:18.269 --> 00:04:20.370 And if that fact isn't already recorded, the 00:04:20.370 --> 00:04:22.509 system does something incredibly powerful. Who 00:04:22.509 --> 00:04:25.470 does it do? It basically pauses its processing, 00:04:25.790 --> 00:04:28.529 generates a prompt, and actively interrogates 00:04:28.529 --> 00:04:31.629 the user. It will literally ask, is Socrates 00:04:31.629 --> 00:04:35.269 a man? That is wild. It realizes what it doesn't 00:04:35.269 --> 00:04:38.189 know? and actively questions you to fill in the 00:04:38.189 --> 00:04:40.269 blank so it can solve the puzzle. Right. And 00:04:40.269 --> 00:04:42.870 the source mentions a massive advantage to this 00:04:42.870 --> 00:04:45.069 rigid rule structure, the explanation facility. 00:04:45.389 --> 00:04:47.949 Oh yes, the why factor. This is huge. Because 00:04:47.949 --> 00:04:49.949 the system is bouncing down that plinko board 00:04:49.949 --> 00:04:52.089 of explicit if -then rules, it can trace its 00:04:52.089 --> 00:04:54.470 own steps backward. If it diagnoses a problem 00:04:54.470 --> 00:04:57.689 and you ask it why, it doesn't spit out some 00:04:57.689 --> 00:05:00.370 opaque mathematical probability. It translates 00:05:00.370 --> 00:05:03.449 its logic trail into natural English. It will 00:05:03.449 --> 00:05:06.850 reply. I concluded this because Rule 42 states 00:05:06.850 --> 00:05:10.170 all men are mortal, and Fact B states Socrates 00:05:10.170 --> 00:05:13.379 is a man. I mean, if it can explain itself and 00:05:13.379 --> 00:05:15.740 build that kind of trust with the user, it must 00:05:15.740 --> 00:05:17.800 have jumped out of the lab pretty quickly. Who 00:05:17.800 --> 00:05:20.319 actually trusted this enough to put it into practice 00:05:20.319 --> 00:05:22.740 back then? Well, to put it in context, in the 00:05:22.740 --> 00:05:25.660 late 1950s and 60s, researchers were trying to 00:05:25.660 --> 00:05:28.060 build diagnostic systems, especially in medicine. 00:05:28.139 --> 00:05:30.420 But they were hitting walls using traditional 00:05:30.420 --> 00:05:32.639 flow charts or statistical pattern matching. 00:05:32.980 --> 00:05:35.019 Those early attempts were just simply too rigid. 00:05:35.100 --> 00:05:38.839 Right. Then around 1965, Edward Feigenbaum, he's 00:05:38.839 --> 00:05:40.889 often called the father of experts. systems, 00:05:41.449 --> 00:05:43.569 and his team at Stanford introduced the formal 00:05:43.569 --> 00:05:46.750 concept. And Feigenbaum had this really vital 00:05:46.750 --> 00:05:49.269 insight. Really was. He realized that the true 00:05:49.269 --> 00:05:51.470 power of an intelligent system comes from the 00:05:51.470 --> 00:05:53.730 specific knowledge it possesses, not just from 00:05:53.730 --> 00:05:56.629 general problem -solving math. Knowing a lot 00:05:56.629 --> 00:05:59.189 about a highly specific domain is better than 00:05:59.189 --> 00:06:01.790 being generally smart but totally ignorant of 00:06:01.790 --> 00:06:03.990 the details. Exactly. And they proved it with 00:06:03.990 --> 00:06:07.399 a system called MYCIN. Oh yeah, the sources detail 00:06:07.399 --> 00:06:11.079 how they built MYCIN to diagnose infectious blood 00:06:11.079 --> 00:06:14.160 diseases. But it wasn't just using basic true 00:06:14.160 --> 00:06:17.800 or false logic, was it? Not at all. MYCIN introduced 00:06:17.800 --> 00:06:20.060 certainty factors. Because in medicine, things 00:06:20.060 --> 00:06:22.259 are rarely absolute, right? Yeah, of course. 00:06:22.420 --> 00:06:24.899 A doctor might observe a symptom and be, say, 00:06:25.100 --> 00:06:29.220 70 % sure it indicates a specific bacteria. MYCIN 00:06:29.220 --> 00:06:31.519 allowed the rules to incorporate that fuzzy logic. 00:06:31.660 --> 00:06:33.300 That's fascinating. It would ask the physician 00:06:33.300 --> 00:06:36.459 questions about the patient, calculate the compounding 00:06:36.459 --> 00:06:38.540 probabilities of different infections based on 00:06:38.540 --> 00:06:41.620 those certainty factors, and then recommend an 00:06:41.620 --> 00:06:44.040 antibiotic therapy. Wow. And it was actually 00:06:44.040 --> 00:06:48.339 outperforming at Stanford in the 1970s. The 1970s. 00:06:48.360 --> 00:06:50.839 And it wasn't just medicine. The source lists 00:06:50.839 --> 00:06:53.720 Dendrol identifying organic molecules in chemistry 00:06:53.720 --> 00:06:57.800 and Garvin AS1, which that was the first medical 00:06:57.800 --> 00:07:00.620 expert system used for daily routine diagnosis 00:07:00.620 --> 00:07:02.459 internationally in Australia, right? Yes, running 00:07:02.459 --> 00:07:05.680 on 661 rules, actually. That is incredible. But 00:07:05.680 --> 00:07:07.639 the one that really caught my eye was the APS 00:07:07.639 --> 00:07:09.819 system. They used a programming language called 00:07:09.819 --> 00:07:12.360 Prologue to encode the British Nationality Act 00:07:12.360 --> 00:07:15.240 of 1981. Yeah, that's a famous one. They literally 00:07:15.240 --> 00:07:18.160 turned international law into executable code. 00:07:19.079 --> 00:07:22.439 But why use Prolog? Why not just write it in 00:07:22.439 --> 00:07:25.379 standard languages like C or COBOL that everyone 00:07:25.379 --> 00:07:28.149 in the business world was already using? So standard 00:07:28.149 --> 00:07:31.069 languages like C or COBOL are imperative. You 00:07:31.069 --> 00:07:33.550 have to give the computer step -by -step instructions, 00:07:33.850 --> 00:07:36.490 go to this memory address, add to, move to the 00:07:36.490 --> 00:07:39.389 next step. Right, very procedural. Exactly. Prologue 00:07:39.389 --> 00:07:41.750 is declarative. It is literally a language built 00:07:41.750 --> 00:07:44.610 on logic statements. You declare a truth, like 00:07:44.610 --> 00:07:47.189 a parent of a parent is a grandparent. OK. You 00:07:47.189 --> 00:07:49.209 don't tell the computer how to find the grandparent. 00:07:49.389 --> 00:07:51.730 You just define the logic. And the languages 00:07:51.730 --> 00:07:54.110 built in inference engine figures out the rest. 00:07:54.350 --> 00:07:56.990 It made encoding complex legal statu - it's just 00:07:56.990 --> 00:07:59.350 incredibly natural. Okay, so we have systems 00:07:59.350 --> 00:08:01.350 outperforming doctors and parsing international 00:08:01.350 --> 00:08:04.750 law. By the 1980s, two -thirds of Fortune 500 00:08:04.750 --> 00:08:06.689 companies were applying this technology in their 00:08:06.689 --> 00:08:09.850 daily business. But I have to ask, if these systems 00:08:09.850 --> 00:08:13.029 are essentially just giant complex webs of if 00:08:13.029 --> 00:08:15.470 -then rules, were they actually intelligent or 00:08:15.470 --> 00:08:17.709 were they just incredibly fast filing cabinets 00:08:17.709 --> 00:08:20.259 cross -referencing data? What's fascinating here 00:08:20.259 --> 00:08:22.120 is how the source answers that, specifically 00:08:22.120 --> 00:08:25.199 with a story from 1982 about a software program 00:08:25.199 --> 00:08:28.920 called SID, Synthesis of Integral Design. Right, 00:08:28.920 --> 00:08:32.379 the SID controversy. Yeah. SID was an expert 00:08:32.379 --> 00:08:35.000 system written in Lisp, and it was tasked with 00:08:35.000 --> 00:08:38.759 designing CPU logic gates for the VX9000 computer. 00:08:39.399 --> 00:08:41.460 OK. Human expert logic designers input a set 00:08:41.460 --> 00:08:44.639 of foundational rules. But SID took those rules 00:08:44.639 --> 00:08:47.539 and expanded on them. It generated logic synthesis 00:08:47.539 --> 00:08:49.779 routines that were vastly more complex than the 00:08:49.779 --> 00:08:51.860 original inputs. It beat the masters at their 00:08:51.860 --> 00:08:54.019 own game. Yeah. It ended up designing, what, 00:08:54.220 --> 00:08:57.519 93 % of the CPU's logic gates? Yes. And the design 00:08:57.519 --> 00:08:59.740 actually exceeded the capabilities of the human 00:08:59.740 --> 00:09:01.659 experts who wrote the rules in the first place. 00:09:01.679 --> 00:09:04.620 Yeah. But the ending of that story is crazy. 00:09:05.320 --> 00:09:07.120 The source notes that the program was highly 00:09:07.120 --> 00:09:09.539 controversial. And despite its massive success. 00:09:09.639 --> 00:09:12.600 The human logic designers terminated the AI after 00:09:12.600 --> 00:09:14.259 the project was finished. It pulled the plug. 00:09:14.559 --> 00:09:18.090 They did. It brings up questions of ego, certainly, 00:09:18.429 --> 00:09:22.009 but also trust. When a machine's output is so 00:09:22.009 --> 00:09:24.450 complex that the humans can't fully grasp how 00:09:24.450 --> 00:09:26.870 it optimized everything, even if it technically 00:09:26.870 --> 00:09:29.649 works flawlessly, people get really uncomfortable. 00:09:29.730 --> 00:09:32.690 It felt alien to them. Completely alien. So if 00:09:32.690 --> 00:09:35.330 SID is outperforming the masters and Fortune 00:09:35.330 --> 00:09:38.470 500s are eating this up, why did it hit a wall? 00:09:40.309 --> 00:09:42.360 Because... Spoiler alert to you listening. You 00:09:42.360 --> 00:09:44.620 don't hear the phrase expert system much anymore. 00:09:45.039 --> 00:09:47.440 We don't see massive server farms dedicated to 00:09:47.440 --> 00:09:49.820 prologue logic puzzles. No, we don't. The collapse 00:09:49.820 --> 00:09:52.259 really came down to the fatal flaws of the if 00:09:52.259 --> 00:09:55.779 -then architecture itself. The first major roadblock 00:09:55.779 --> 00:09:58.799 was what researchers called the knowledge acquisition 00:09:58.799 --> 00:10:02.100 problem The human bottleneck exactly. I mean 00:10:02.100 --> 00:10:03.620 think about your own job for a second if you're 00:10:03.620 --> 00:10:05.960 listening to this How hard would it be to sit 00:10:05.960 --> 00:10:07.879 down and write out every single subconscious 00:10:07.879 --> 00:10:10.580 decision you make in a day without ever contradicting 00:10:10.580 --> 00:10:12.659 yourself? It's nearly impossible. You would have 00:10:12.659 --> 00:10:14.320 a nervous breakdown trying to map out your own 00:10:14.320 --> 00:10:17.299 intuition and To build these systems, companies 00:10:17.299 --> 00:10:19.820 had to hire these specialized knowledge engineers 00:10:19.820 --> 00:10:23.480 whose entire job was to extract these intuitive 00:10:23.480 --> 00:10:26.980 rules from domain experts. But these experts 00:10:26.980 --> 00:10:30.240 are, by definition, highly valued and incredibly 00:10:30.240 --> 00:10:33.360 busy people. Getting hours of their time and 00:10:33.360 --> 00:10:35.500 then trying to translate their gut feelings into 00:10:35.500 --> 00:10:39.179 rigid logic trees, it was agonizingly slow and 00:10:39.179 --> 00:10:41.620 incredibly expensive. And even if you could extract 00:10:41.620 --> 00:10:44.559 all those rules, The computers themselves started 00:10:44.559 --> 00:10:46.539 choking on them, didn't they? The math simply 00:10:46.539 --> 00:10:49.159 caught up with the ambition. Yeah, the scaling 00:10:49.159 --> 00:10:51.519 nightmare. Researchers started envisioning the 00:10:51.519 --> 00:10:54.120 ultimate expert system with, like, 100 million 00:10:54.120 --> 00:10:56.600 rules. Wow. But they hit a hard mathematical 00:10:56.600 --> 00:10:59.779 wall. When you have that many rules, verifying 00:10:59.779 --> 00:11:01.799 that they don't contradict each other turns into 00:11:01.799 --> 00:11:05.200 a massive computational issue known as the SAT 00:11:05.200 --> 00:11:08.500 problem, Boolean Satisfiability. OK, let's break 00:11:08.500 --> 00:11:11.330 that down practically. Why is checking for contradictions 00:11:11.330 --> 00:11:13.889 so hard for a computer? I mean, it's just cross 00:11:13.889 --> 00:11:15.710 -referencing text, right? Well, it's not just 00:11:15.710 --> 00:11:18.509 text. It's a combinatorial explosion. Imagine 00:11:18.509 --> 00:11:20.549 you have a system with a hundred thousand rules. 00:11:21.049 --> 00:11:24.610 If you introduce one new rule, you have to ensure 00:11:24.610 --> 00:11:27.070 it doesn't create some hidden paradox with rule 00:11:27.070 --> 00:11:31.009 number 4000. Oh, right. To verify absolute consistency 00:11:31.009 --> 00:11:33.090 across hundreds of interconnected variables, 00:11:33.570 --> 00:11:36.129 the computer has to check every single possible 00:11:36.129 --> 00:11:39.289 combination of true and false states. So the 00:11:39.289 --> 00:11:41.610 search space just grows exponentially. It's like 00:11:41.610 --> 00:11:44.210 a butterfly effect of logic. Precisely. It's 00:11:44.210 --> 00:11:46.870 2 to the power of n. So if you have just 10 variables, 00:11:46.990 --> 00:11:49.350 that's over a thousand combinations. Which is 00:11:49.350 --> 00:11:51.620 fine for a computer. Right. But once you reach 00:11:51.620 --> 00:11:54.179 hundreds of variables, the number of permutations 00:11:54.179 --> 00:11:56.620 quickly exceeds the number of atoms in the known 00:11:56.620 --> 00:12:00.419 universe. Wow. It literally paralyzes the processor. 00:12:01.100 --> 00:12:04.399 Massive systems became fundamentally computationally 00:12:04.399 --> 00:12:06.379 unworkable. And on top of that, they suffered 00:12:06.379 --> 00:12:09.159 from severe overfitting, right? Yes. Because 00:12:09.159 --> 00:12:12.100 the system only knows exactly what is explicitly 00:12:12.100 --> 00:12:14.759 described in its knowledge base, it struggles 00:12:14.759 --> 00:12:17.840 to handle novel situations. If a case comes in 00:12:17.840 --> 00:12:21.080 that is just slightly outside, the pre -programmed 00:12:21.080 --> 00:12:24.139 rules, the system doesn't have like common sense 00:12:24.139 --> 00:12:27.039 to fall back on. It just breaks. And then you 00:12:27.039 --> 00:12:29.960 add all of that to what the source calls integration 00:12:29.960 --> 00:12:33.019 woes. You have these incredible AI systems built 00:12:33.019 --> 00:12:37.159 in exotic languages like Lisp and Prolog running 00:12:37.159 --> 00:12:40.580 on highly specialized. Very expensive Lisp machines. 00:12:41.019 --> 00:12:43.799 Which corporate IT departments absolutely hated? 00:12:44.080 --> 00:12:45.620 I can imagine. They were sitting there running 00:12:45.620 --> 00:12:48.000 their entire massive businesses on imperative 00:12:48.000 --> 00:12:51.340 languages like COBOL on IBM mainframes, and then 00:12:51.340 --> 00:12:55.639 1981 hits, the IBM PC is introduced, and suddenly 00:12:55.639 --> 00:12:58.360 we have this huge paradigm shift to the client 00:12:58.360 --> 00:13:00.399 server model. Corporate IT departments wanted 00:13:00.399 --> 00:13:03.259 AI on their affordable PCs and standard servers. 00:13:03.539 --> 00:13:05.779 They didn't want it on weird, isolated hardware 00:13:05.779 --> 00:13:07.779 that couldn't even talk to their legacy databases. 00:13:08.039 --> 00:13:10.519 It was just totally incompatible with the business 00:13:10.519 --> 00:13:12.860 infrastructure of the time. So between the human 00:13:12.860 --> 00:13:15.179 bottlenecks, the mathematical limits of the SAT 00:13:15.179 --> 00:13:17.720 problem, and the IT departments just rejecting 00:13:17.720 --> 00:13:20.340 the hardware, we entered the AI winter of the 00:13:20.340 --> 00:13:23.519 1990s. The term expert system practically vanished 00:13:23.519 --> 00:13:25.919 from the IT lexicon. It really did. Which brings 00:13:25.919 --> 00:13:27.700 us to the ultimate question of this deep dive. 00:13:28.279 --> 00:13:30.580 Did this technology actually die, or is it just 00:13:30.580 --> 00:13:33.639 in disguise? Well, the source outlines two totally 00:13:33.639 --> 00:13:35.820 opposite interpretations of what happened next. 00:13:36.059 --> 00:13:39.179 Theory one is the pessimistic view. Expert systems 00:13:39.179 --> 00:13:42.259 failed. They simply couldn't deliver on the overhyped 00:13:42.259 --> 00:13:45.259 promises of the 1980s. The scaling problems were 00:13:45.259 --> 00:13:47.600 totally insurmountable and the tech world basically 00:13:47.600 --> 00:13:49.940 just cut its losses and moved on. But theory 00:13:49.940 --> 00:13:52.860 two is the stealth resurrection. Yes. They didn't 00:13:52.860 --> 00:13:55.039 die at all. They were victims of their own massive 00:13:55.039 --> 00:13:58.039 success. The IT world just swallowed them whole. 00:13:58.279 --> 00:14:01.360 Exactly. As IT professionals really grasped the 00:14:01.360 --> 00:14:03.399 concept of separating the rules from the inference 00:14:03.399 --> 00:14:06.299 engine, they stopped treating them as magical 00:14:06.299 --> 00:14:08.500 standalone artificial intelligence. Right. The 00:14:08.500 --> 00:14:10.679 mystique wore off. Yeah. They just became another 00:14:10.679 --> 00:14:13.460 standard tool in the software toolbox. Major 00:14:13.460 --> 00:14:17.259 vendors like SAP, Siebel, and Oracle quietly 00:14:17.259 --> 00:14:19.899 integrated expert system abilities right into 00:14:19.899 --> 00:14:22.019 their standard business. application suites. 00:14:22.220 --> 00:14:24.039 And they became known as business rule engines. 00:14:24.279 --> 00:14:27.080 Exactly. Business rule engines. So they migrated 00:14:27.080 --> 00:14:30.840 from being sci -fi AI in a Stanford lab to becoming 00:14:30.840 --> 00:14:33.960 the invisible plumbing of corporate logic. I 00:14:33.960 --> 00:14:36.879 mean, they are still out there right now, quietly 00:14:36.879 --> 00:14:40.039 automating business processes, approving mortgages, 00:14:40.279 --> 00:14:42.980 and running diagnostics behind the scenes. They 00:14:42.980 --> 00:14:45.600 are. And if we connect this to the bigger picture, 00:14:46.360 --> 00:14:49.519 the definition of AI simply shifted. How so? 00:14:50.190 --> 00:14:52.730 Modern artificial intelligence, like the machine 00:14:52.730 --> 00:14:54.809 learning models everyone is obsessed with today, 00:14:55.210 --> 00:14:57.110 operates very differently from these classic 00:14:57.110 --> 00:15:00.289 expert systems. Modern AI, like recurrent neural 00:15:00.289 --> 00:15:03.389 networks, relies on big data and feedback mechanisms. 00:15:04.129 --> 00:15:06.049 Instead of a human trying to write a million 00:15:06.049 --> 00:15:08.350 rigid rules, you just feed the neural network 00:15:08.350 --> 00:15:10.789 millions of examples and it learns to recognize 00:15:10.789 --> 00:15:13.009 the patterns itself. It generalizes far better 00:15:13.009 --> 00:15:15.210 than if -then rules ever could. Tremendously 00:15:15.210 --> 00:15:17.230 better. There is a great example of this in the 00:15:17.230 --> 00:15:19.110 source material regarding voice recognition. 00:15:19.450 --> 00:15:22.009 Early researchers built an expert system called 00:15:22.009 --> 00:15:24.870 Hearsay. Ah, yes. They tried to recognize speech 00:15:24.870 --> 00:15:27.289 using explicit if -then rules about audio phones. 00:15:27.429 --> 00:15:30.110 And it struggled immensely. Because human speech 00:15:30.110 --> 00:15:33.629 is noisy, it is messy, it relies heavily on context. 00:15:34.049 --> 00:15:36.289 You cannot capture the difference between a Boston 00:15:36.289 --> 00:15:38.769 accent and a Texas accent with a strict set of 00:15:38.769 --> 00:15:41.669 logic rules. She's too fuzzy. Right. Hearsay 00:15:41.669 --> 00:15:43.929 proved that interpretation problems, like messy 00:15:43.929 --> 00:15:46.629 pattern matching, were far better suited for 00:15:46.629 --> 00:15:49.049 neural networks than rule -based expert systems. 00:15:48.970 --> 00:15:52.970 So what does this all mean? If we zoom out, we 00:15:52.970 --> 00:15:56.250 see this incredible historical arc. Expert systems 00:15:56.250 --> 00:15:59.309 may not be the shiny face of AI anymore, but 00:15:59.309 --> 00:16:02.629 they completely pioneered the idea that capturing 00:16:02.629 --> 00:16:05.389 human knowledge was inherently valuable. Yes, 00:16:05.389 --> 00:16:07.429 they absolutely did. They pushed the computing 00:16:07.429 --> 00:16:09.610 world forward, they paved the way for client 00:16:09.610 --> 00:16:12.009 -server environments, and they still quietly 00:16:12.009 --> 00:16:14.330 run the business logic of the world today. But 00:16:14.330 --> 00:16:16.309 they also offer something that we have largely 00:16:16.309 --> 00:16:19.210 lost in the modern AI boom. What's that? Today 00:16:19.210 --> 00:16:22.309 we live in a world of complex black box AI. When 00:16:22.309 --> 00:16:24.990 a neural network denies your loan or misidentifies 00:16:24.990 --> 00:16:27.909 a medical scan and you ask it why, it cannot 00:16:27.909 --> 00:16:29.950 really tell you. No, it just points to billions 00:16:29.950 --> 00:16:32.269 of statistical weights scattered across a massive 00:16:32.269 --> 00:16:35.799 matrix. Exactly. The absolute transparency of 00:16:35.799 --> 00:16:38.440 the old expert system's explanation facility, 00:16:38.679 --> 00:16:41.539 that ability to reply in plain English, I decided 00:16:41.539 --> 00:16:44.240 this because of Rule A and Fact B. That is something 00:16:44.240 --> 00:16:47.580 we actually deeply miss today. We traded transparency 00:16:47.580 --> 00:16:50.220 for scale. We got the answers, but we lost the 00:16:50.220 --> 00:16:52.200 explanation. Which brings up a fascinating thought 00:16:52.200 --> 00:16:54.740 to leave you with. When an old expert system 00:16:54.740 --> 00:16:57.220 made a terrible mistake, you could trace the 00:16:57.220 --> 00:16:59.779 exact rule that caused it, find the human knowledge 00:16:59.779 --> 00:17:01.940 engineer who wrote it, and hold them accountable. 00:17:02.220 --> 00:17:06.460 But today, when a black box neural network hallucinates 00:17:06.460 --> 00:17:08.740 a legal precedent or misdiagnoses a patient, 00:17:09.220 --> 00:17:12.970 the logic is entirely invisible. As these modern 00:17:12.970 --> 00:17:15.190 systems take over our most critical infrastructure, 00:17:15.849 --> 00:17:18.049 next time you make a complex decision at work, 00:17:18.490 --> 00:17:21.089 ask yourself, could your thought process be broken 00:17:21.089 --> 00:17:23.930 down into a massive tree of if -then rules? Probably 00:17:23.930 --> 00:17:26.970 not. And if it can't, what exactly is the secret 00:17:26.970 --> 00:17:28.849 ingredient of human intuition that a machine 00:17:28.849 --> 00:17:31.670 can't capture? And when the machine gets it wrong, 00:17:32.089 --> 00:17:35.130 without being able to explain why, who exactly 00:17:35.130 --> 00:17:36.210 do we hold accountable?