WEBVTT 00:00:00.000 --> 00:00:02.580 If you ask someone on the street for directions 00:00:02.580 --> 00:00:05.419 to the nearest coffee shop, they don't usually 00:00:05.419 --> 00:00:07.559 look at you and say, well, put your left foot 00:00:07.559 --> 00:00:10.699 forward, shift your weight, now flex your knee 00:00:10.699 --> 00:00:12.740 20 degrees. Right, that would be terrifying. 00:00:13.099 --> 00:00:14.900 And, you know, completely exhausting. They just 00:00:14.900 --> 00:00:17.579 say, take a left. It's brick build. Exactly. 00:00:18.140 --> 00:00:19.940 They give you the layout and they just trust 00:00:19.940 --> 00:00:22.210 your brain to figure out the actual. physical 00:00:22.210 --> 00:00:23.829 mechanics of getting there. Yeah, they don't 00:00:23.829 --> 00:00:26.250 micromanage your joints. But for decades, I mean, 00:00:26.309 --> 00:00:29.429 we forced computers to do exactly that. We communicated 00:00:29.429 --> 00:00:32.369 with them using that exhausting foot -by -foot 00:00:32.369 --> 00:00:35.530 method. But today, we are looking at a completely 00:00:35.530 --> 00:00:38.149 different paradigm. So welcome to another deep 00:00:38.149 --> 00:00:40.710 dive specially tailored for you, the learner. 00:00:41.450 --> 00:00:44.009 Today, we're skipping the dense academic textbooks 00:00:44.009 --> 00:00:46.630 and jumping straight into a really fascinating 00:00:46.630 --> 00:00:49.689 source, a comprehensive Wikipedia article on 00:00:49.689 --> 00:00:52.490 logic programming. It's a great piece. It really 00:00:52.490 --> 00:00:55.369 is. And the mission of this deep dive is to extract 00:00:55.369 --> 00:00:58.490 the absolute best nuggets from this text. We're 00:00:58.490 --> 00:01:00.810 going to demystify what logic programming is, 00:01:01.210 --> 00:01:04.590 explore this dramatic history behind it, and 00:01:04.590 --> 00:01:07.409 uncover how it actually mimics human thought 00:01:07.409 --> 00:01:09.870 and human law. Yeah, and the goal today is really 00:01:09.870 --> 00:01:12.250 to look at the why behind the code, not just 00:01:12.250 --> 00:01:15.010 how it works, but why this approach even exists 00:01:15.010 --> 00:01:17.549 and why it matters. OK, let's unpack this. Yeah. 00:01:17.569 --> 00:01:19.230 Because before we can understand why this sparked 00:01:19.230 --> 00:01:21.989 an actual academic war, we need to know how it 00:01:21.989 --> 00:01:24.430 actually works. Right. So the fundamental shift 00:01:24.430 --> 00:01:28.230 here is moving from procedural to declarative 00:01:28.230 --> 00:01:30.390 programming. Declarative, meaning? Meaning you 00:01:30.390 --> 00:01:32.469 don't give the computer step -by -step instructions. 00:01:32.909 --> 00:01:34.329 That foot -by -foot method, that's procedural. 00:01:34.469 --> 00:01:36.989 You're dictating the control flow. But logic 00:01:36.989 --> 00:01:39.200 programming is declarative. You just give the 00:01:39.200 --> 00:01:41.379 computer a set of sentences in logical form, 00:01:41.519 --> 00:01:43.120 facts, and rules, and you just let it figure 00:01:43.120 --> 00:01:45.000 out the solution. So you're just describing the 00:01:45.000 --> 00:01:47.659 problem, basically. Exactly. You describe what 00:01:47.659 --> 00:01:50.400 the problem is, not how to solve it. If you look 00:01:50.400 --> 00:01:52.560 at the structure of a rule in this system, it 00:01:52.560 --> 00:01:55.280 looks sort of backwards. You write the conclusion 00:01:55.280 --> 00:01:58.540 first. So let's say conclusion A. Then you write 00:01:58.540 --> 00:02:01.040 a colon and a dash, followed by the conditions. 00:02:01.099 --> 00:02:04.640 Let's call them B and C. Wait, a colon and a 00:02:04.640 --> 00:02:07.140 dash, like a sideways smiley face. Yeah, a lot 00:02:07.140 --> 00:02:09.680 of people see that. But think of that colon and 00:02:09.680 --> 00:02:11.919 dash as the computer's way of saying the word 00:02:11.919 --> 00:02:16.560 I F. So it translates to A is true, I F, B and 00:02:16.560 --> 00:02:19.000 C are true. Oh, I see. Yeah. And A is the head 00:02:19.000 --> 00:02:21.639 of the rule, the goal. B and C are the body. 00:02:21.840 --> 00:02:25.020 And these are known as horn clauses. Horn clauses. 00:02:25.199 --> 00:02:27.479 Right. Which is just a formal way to describe 00:02:27.479 --> 00:02:30.159 rules that have one clear, definite conclusion. 00:02:30.479 --> 00:02:32.520 And if you just have A with no conditions, that's 00:02:32.520 --> 00:02:34.840 just a fact. Then you just ask the system queries, 00:02:35.060 --> 00:02:36.840 and it does a deduction. OK, let me try to make 00:02:36.840 --> 00:02:39.300 this concrete using the family tree example from 00:02:39.300 --> 00:02:41.039 the source text. Sure, go ahead. So if I tell 00:02:41.039 --> 00:02:44.340 the computer, Charles is William's parent. That's 00:02:44.340 --> 00:02:46.620 a fact, right? Yep, a base fact. And then I give 00:02:46.620 --> 00:02:48.800 it a rule, like a grandparent is a parent of 00:02:48.800 --> 00:02:52.219 a parent. In a normal procedural language, I 00:02:52.219 --> 00:02:54.460 have to write a loop to search the whole database, 00:02:54.560 --> 00:02:57.259 hold it in memory, cross -reference everything. 00:02:57.419 --> 00:02:59.539 Which takes a lot of manual coding, yeah. Right. 00:02:59.900 --> 00:03:01.659 But here, I don't need to write a loop. I just 00:03:01.659 --> 00:03:04.580 ask the query, who is William's grandparent? 00:03:05.180 --> 00:03:07.199 And the computer does the logical heavy lifting. 00:03:07.520 --> 00:03:09.840 It just connects the dots. That is the exact 00:03:09.840 --> 00:03:12.520 elegance of it. It searches its database of horn 00:03:12.520 --> 00:03:16.340 clauses and figures it out. So if it's that elegant... 00:03:16.460 --> 00:03:18.860 I mean, why wasn't this the only way we programmed 00:03:18.860 --> 00:03:21.939 computers from the start? Because it caused a 00:03:21.939 --> 00:03:24.500 massive academic rivalry. Really? Oh, yeah. In 00:03:24.500 --> 00:03:27.479 the late 1960s and early 1970s, the AI community 00:03:27.479 --> 00:03:30.520 was completely split. On one side, you had the 00:03:30.520 --> 00:03:32.819 declarative camp. Which our sources say was mostly 00:03:32.819 --> 00:03:34.460 Stanford and Edinburgh, right? Right. With guys 00:03:34.460 --> 00:03:36.479 like John McCarthy, Cordell Green, Robert Kowalski. 00:03:36.599 --> 00:03:38.800 Exactly. They believed knowledge should be represented 00:03:38.800 --> 00:03:41.039 logically. But on the other side, you had the 00:03:41.039 --> 00:03:45.349 procedural camp, mainly at MIT. Ah, MIT. Marvin 00:03:45.349 --> 00:03:47.889 Minsky and Seymour Paper. Right, and they wanted 00:03:47.889 --> 00:03:50.169 step -by -step control. They looked at pure logic 00:03:50.169 --> 00:03:52.509 and thought it was way too inefficient for real 00:03:52.509 --> 00:03:54.789 -world computing. Wait, so you're telling me 00:03:54.789 --> 00:03:57.169 that one of the most important programming languages 00:03:57.169 --> 00:04:01.689 in AI history, Prologue, basically emerged because 00:04:01.689 --> 00:04:04.789 two rival academic factions couldn't agree on 00:04:04.789 --> 00:04:06.750 whether to give computers rules or instructions. 00:04:06.889 --> 00:04:09.280 That is exactly what happened. The declarative 00:04:09.280 --> 00:04:12.000 camp wanted mathematical purity, and the procedural 00:04:12.000 --> 00:04:14.979 camp wanted control, using things like backtracking, 00:04:15.080 --> 00:04:18.220 where the computer tries a path and rewinds if 00:04:18.220 --> 00:04:20.980 it hits a dead end. Like a maze. Yeah, exactly 00:04:20.980 --> 00:04:24.319 like a maze. But the two sides finally reconciled, 00:04:24.420 --> 00:04:26.899 thanks to Elaine Kolmerauer and Robert Kowalski. 00:04:26.959 --> 00:04:29.339 This was in Marseille, right? Around 1971 or 00:04:29.339 --> 00:04:32.040 72? Yes. Holmerhauer was working on a French 00:04:32.040 --> 00:04:34.439 question answering system. He invited Kowalski 00:04:34.439 --> 00:04:36.959 over, and together they realized you could actually 00:04:36.959 --> 00:04:39.740 execute pure logic as a controlled series of 00:04:39.740 --> 00:04:42.420 SNPs. They combined the two philosophies. Wow. 00:04:42.540 --> 00:04:45.019 And that reconciliation led directly to the creation 00:04:45.019 --> 00:04:48.259 of Prolog in 1972. Which stands for programming 00:04:48.259 --> 00:04:52.220 in logic. You got it. And this whole reconciliation 00:04:52.220 --> 00:04:55.060 led to a really fundamental realization about 00:04:55.060 --> 00:04:57.420 problem solving, which Kowalski actually captured 00:04:57.420 --> 00:05:01.060 in an equation. Algorithm equals logic plus control. 00:05:01.519 --> 00:05:03.860 The equation of the century, some call it. It 00:05:03.860 --> 00:05:06.759 means any algorithm has two parts. The logic 00:05:06.759 --> 00:05:10.139 is the knowledge, the rules and facts, and the 00:05:10.139 --> 00:05:13.319 control is the strategy the machine uses to process 00:05:13.319 --> 00:05:16.779 it. So logic is knowing the recipe for a cake. 00:05:17.019 --> 00:05:19.480 And control is whether you start by gathering 00:05:19.480 --> 00:05:21.839 the ingredients, which the text calls forward 00:05:21.839 --> 00:05:24.720 reasoning, or visualizing the finished cake and 00:05:24.720 --> 00:05:26.879 working backward to find out what you need, which 00:05:26.879 --> 00:05:29.040 is backward reasoning. That's a perfect analogy. 00:05:29.439 --> 00:05:31.720 Backward reasoning is top -down, like Prologue 00:05:31.720 --> 00:05:35.439 uses. It reduces goals down to base facts. Forward 00:05:35.439 --> 00:05:37.879 reasoning is bottom -up, deriving all possible 00:05:37.879 --> 00:05:39.759 facts from the base. Does it really make a big 00:05:39.759 --> 00:05:41.560 difference which one you use? Oh, absolutely. 00:05:41.899 --> 00:05:43.860 The text uses the Fibonacci sequence to show 00:05:43.860 --> 00:05:45.579 this. Oh, right, where each number is the sum 00:05:45.579 --> 00:05:47.680 of the two before it. Yeah. If you use backward 00:05:47.680 --> 00:05:49.740 reasoning to find a large Fibonacci number, the 00:05:49.740 --> 00:05:51.720 computer starts at the top and says, OK, I need 00:05:51.720 --> 00:05:54.920 n minus 1 and n minus 2. And then for n minus 00:05:54.920 --> 00:05:57.319 1, it needs n minus 2 and n minus 3. So it's 00:05:57.319 --> 00:06:00.600 calculating n minus 2 twice. Exactly. It redundantly 00:06:00.600 --> 00:06:03.160 calculates the same numbers over and over. Its 00:06:03.160 --> 00:06:05.319 exponential complexity, the computer will just 00:06:05.319 --> 00:06:08.019 crash. Yikes. But if you just change the control 00:06:08.019 --> 00:06:10.459 to forward reasoning, keeping the exact same 00:06:10.459 --> 00:06:12.839 logic, it starts at 0 and 1 and just adds them 00:06:12.839 --> 00:06:15.620 up linearly. It takes almost no time at all. 00:06:15.740 --> 00:06:18.139 Wow, so no recomputation. It just walks up the 00:06:18.139 --> 00:06:22.180 sequence. But real life isn't a neat math problem 00:06:22.180 --> 00:06:25.060 with perfect facts. How does it handle missing 00:06:25.060 --> 00:06:27.240 information? That is where we get to a concept 00:06:27.240 --> 00:06:30.420 called negation as failure. Negation as failure. 00:06:30.519 --> 00:06:33.060 Right. In this paradigm, if the computer cannot 00:06:33.060 --> 00:06:36.120 prove a positive condition is true, it just assumes 00:06:36.120 --> 00:06:38.800 it's false. Here's where it gets really interesting. 00:06:39.459 --> 00:06:41.199 Doesn't this mean the computer is just jumping 00:06:41.199 --> 00:06:44.000 to conclusions based on ignorance? Is that safe? 00:06:44.170 --> 00:06:47.430 Well, it makes it a non -monotonic logic, meaning 00:06:47.430 --> 00:06:49.930 conclusions can be withdrawn when new information 00:06:49.930 --> 00:06:52.509 appears. It's highly adaptable. Can we use the 00:06:52.509 --> 00:06:54.490 Tom the thief example from the article? Let's 00:06:54.490 --> 00:06:58.009 do it. So say we have a rule. A person is punished 00:06:58.009 --> 00:07:00.709 if they're a thief unless they are a minor. OK. 00:07:01.089 --> 00:07:04.569 We tell the system, Tom is a thief. It checks 00:07:04.569 --> 00:07:06.790 if he's a minor, finds no evidence of it, so 00:07:06.790 --> 00:07:09.930 it assumes he's not. Verdict. Tom is punished. 00:07:10.230 --> 00:07:12.709 So we jump to that conclusion, but then what 00:07:12.709 --> 00:07:15.569 if we learn he's 15? We input that fact. The 00:07:15.569 --> 00:07:17.850 system immediately withdraws the punishment conclusion 00:07:17.850 --> 00:07:20.470 because now the minor condition is met. The new 00:07:20.470 --> 00:07:23.970 verdict is Tom is rehabilitated. And if we learn 00:07:23.970 --> 00:07:27.089 he's a violent offender. We add a rule that violence 00:07:27.089 --> 00:07:30.410 defeats rehabilitation. The system reevaluates, 00:07:30.550 --> 00:07:32.810 the exception triggers, and the punishment is 00:07:32.810 --> 00:07:35.189 reinstated. It's just arguing with itself. It's 00:07:35.189 --> 00:07:37.230 bouncing between rules and exceptions. Exactly. 00:07:37.310 --> 00:07:39.829 And that exact tendency to jump to conclusions 00:07:39.829 --> 00:07:42.870 and revise them later isn't a flaw. It actually 00:07:42.870 --> 00:07:45.009 perfectly aligns with how the human brain works. 00:07:45.209 --> 00:07:47.290 Really? Yeah. Cognitive scientists like Paul 00:07:47.290 --> 00:07:49.149 Thagard and Keith Stenning, who are cited in 00:07:49.149 --> 00:07:51.610 The Source, argue that this is psychologically 00:07:51.610 --> 00:07:54.269 plausible. Because humans use defaults and exceptions 00:07:54.269 --> 00:07:56.860 all the time. Right. If I say bird, you assume 00:07:56.860 --> 00:07:59.199 it flies. If I say penguin, you pull up the exception. 00:07:59.379 --> 00:08:02.019 You don't crash. And Stenning even says this 00:08:02.019 --> 00:08:04.620 kind of backward reasoning has an appealing neural 00:08:04.620 --> 00:08:07.500 implementation. That makes total sense. So instead 00:08:07.500 --> 00:08:10.060 of writing endless if -then loops to cover every 00:08:10.060 --> 00:08:12.620 single edge case, we are basically giving the 00:08:12.620 --> 00:08:15.420 computer a law degree. like feeding it the Civil 00:08:15.420 --> 00:08:17.439 Code and letting it argue with itself to find 00:08:17.439 --> 00:08:19.379 the verdict. That is quite literally what they 00:08:19.379 --> 00:08:21.839 did. In the 1980s, they digitized the British 00:08:21.839 --> 00:08:24.459 Nationality Act using logic programming. Oh, 00:08:24.459 --> 00:08:27.319 wow. And today, Japan has a modern system called 00:08:27.319 --> 00:08:30.740 Proleg. Right, the prologue -based legal reasoning 00:08:30.740 --> 00:08:33.940 system. Yes. And it uses roughly 2 ,500 rules 00:08:33.940 --> 00:08:36.659 from the Japanese Civil Code. It models human 00:08:36.659 --> 00:08:39.299 legal arguments perfectly. But wait, if it's 00:08:39.299 --> 00:08:42.480 this good... Why did it kind of fade out? The 00:08:42.480 --> 00:08:44.360 text talks about the Japanese fifth -generation 00:08:44.360 --> 00:08:47.519 computer systems project, the FGCS, in the 80s. 00:08:47.620 --> 00:08:49.820 Yeah, they poured massive money into trying to 00:08:49.820 --> 00:08:52.200 use concurrent logic programming for parallel 00:08:52.200 --> 00:08:54.700 AI, meaning multiple processors working at the 00:08:54.700 --> 00:08:57.620 same time. But it failed. It did. It turns out, 00:08:58.000 --> 00:09:00.139 forcing logic to commit to choices in parallel 00:09:00.139 --> 00:09:03.159 hardware ruined the mathematical purity. It introduced 00:09:03.159 --> 00:09:06.080 timing bugs. And meanwhile, regular, general 00:09:06.080 --> 00:09:08.419 -purpose computers just got incredibly fast. 00:09:08.779 --> 00:09:10.840 So standard procedural programming just outpaced 00:09:10.840 --> 00:09:13.500 it in raw speed. Exactly. It couldn't compete. 00:09:13.820 --> 00:09:16.299 But the paradigm didn't die, right? Because the 00:09:16.299 --> 00:09:18.220 source text goes into how it splintered into 00:09:18.220 --> 00:09:22.440 these modern flavors. Like, uh... Datalog. Right. 00:09:22.519 --> 00:09:25.139 Datalog is huge for databases. Standard relational 00:09:25.139 --> 00:09:28.139 databases like SQL choke on recursive queries 00:09:28.139 --> 00:09:30.820 like finding an ancestor 15 generations back. 00:09:30.940 --> 00:09:32.639 Because it doesn't know how many times to loop. 00:09:33.159 --> 00:09:36.340 Exactly. But Datalog handles recursive rules 00:09:36.340 --> 00:09:39.580 natively and efficiently. And then there's ASP, 00:09:39.740 --> 00:09:42.320 answer set programming. The text describes it 00:09:42.320 --> 00:09:45.960 as a generate and pest method. Yes. ASP is amazing 00:09:45.960 --> 00:09:48.899 for constraint problems. Take map coloring. You 00:09:48.899 --> 00:09:51.559 need to color a map red, green, or blue, but 00:09:51.559 --> 00:09:53.500 no touching countries can be the same color. 00:09:53.759 --> 00:09:55.500 Procedurally, that's a nightmare of checking 00:09:55.500 --> 00:09:58.279 every single border. But in ASP, you just generate 00:09:58.279 --> 00:10:00.759 every possible color combination for the universe 00:10:00.759 --> 00:10:03.399 and then apply a constraint test to filter out 00:10:03.399 --> 00:10:05.700 any maps where touching countries share a color. 00:10:05.840 --> 00:10:07.899 Generate the universe, then destroy the ones 00:10:07.899 --> 00:10:10.419 that break the rules. That is wild. It's incredible. 00:10:10.379 --> 00:10:13.200 powerful. But the last one mentioned is inductive 00:10:13.200 --> 00:10:16.700 logic programming or ILP, and this one flips 00:10:16.700 --> 00:10:19.960 the script completely. It does. ILP is a form 00:10:19.960 --> 00:10:22.360 of machine learning. Instead of giving it the 00:10:22.360 --> 00:10:24.559 rules, you give it the facts and the answers, 00:10:24.720 --> 00:10:26.860 positive and negative examples, and you make 00:10:26.860 --> 00:10:29.500 the machine learn the rules. If inductive logic 00:10:29.500 --> 00:10:32.580 programming is about inducing the general rules 00:10:32.580 --> 00:10:34.679 from a bunch of positive and negative examples, 00:10:35.340 --> 00:10:38.039 isn't that essentially exactly how you or I learned 00:10:38.039 --> 00:10:40.620 what a dog or a car was? when we were toddlers. 00:10:40.919 --> 00:10:42.980 It's exactly the same mechanism. You show a kid 00:10:42.980 --> 00:10:46.259 a dog, say dog. Show a cat, say not dog. Eventually, 00:10:46.480 --> 00:10:48.500 they induce the rule of what a dog is. That's 00:10:48.500 --> 00:10:51.179 fascinating. And Stuart Russell, a major AI researcher, 00:10:51.299 --> 00:10:53.419 notes in the text that this kind of concept invention 00:10:53.419 --> 00:10:57.539 is key to reaching true human -level AI. Wow. 00:10:58.379 --> 00:11:01.399 We have cut so much ground today. from the simple 00:11:01.399 --> 00:11:03.639 idea of just giving computers declarative rules, 00:11:04.080 --> 00:11:06.820 to the academic turf wars at MIT and Stanford, 00:11:07.259 --> 00:11:09.879 all the way to modeling the Japanese civil code 00:11:09.879 --> 00:11:12.600 and mimicking toddler cognition. So what does 00:11:12.600 --> 00:11:14.019 this all mean? What does this leave us today? 00:11:14.240 --> 00:11:15.919 Well, it leaves us with a really important thought 00:11:15.919 --> 00:11:18.779 to mull over. Right now, our modern AI, like 00:11:18.779 --> 00:11:21.340 massive neural networks, they're incredibly powerful, 00:11:21.639 --> 00:11:23.940 but they're completely unpredictable black boxes. 00:11:24.179 --> 00:11:26.830 Right, they just... guess based on statistics. 00:11:27.370 --> 00:11:30.389 Exactly. So if logic programming so closely mirrors 00:11:30.389 --> 00:11:33.690 human cognitive rules, defaults, and the legal 00:11:33.690 --> 00:11:36.750 system, might the future of AI require us to 00:11:36.750 --> 00:11:38.950 combine our modern neural networks with this 00:11:38.950 --> 00:11:41.509 classic rule -based logic programming? Just to 00:11:41.509 --> 00:11:44.429 keep the AI's reasoning explainable. Explainable 00:11:44.429 --> 00:11:47.019 and actually aligned with human law. blending 00:11:47.019 --> 00:11:49.220 the statistical guessing with an actual logical 00:11:49.220 --> 00:11:51.720 law degree. Love that. Well, be you the learner. 00:11:52.019 --> 00:11:53.559 Thank you so much for joining us on this deep 00:11:53.559 --> 00:11:55.519 dive. Stay curious, keep questioning how things 00:11:55.519 --> 00:11:56.500 work, and we'll see you next time.