WEBVTT 00:00:00.000 --> 00:00:02.759 Imagine yelling at your smart speaker to turn 00:00:02.759 --> 00:00:04.480 off the living room lights, but leave the porch 00:00:04.480 --> 00:00:07.360 on. And instead, it cheerfully sets an alarm 00:00:07.360 --> 00:00:10.500 for 3 a .m. Oh, yeah. We've all been there. Right. 00:00:11.060 --> 00:00:13.759 Like we have computers that can instantly calculate 00:00:13.759 --> 00:00:16.719 the precise orbital trajectory to land a rover 00:00:16.719 --> 00:00:19.839 on Mars. Yet they completely fail to understand 00:00:19.839 --> 00:00:23.039 a basic sentence spoken by a human in a kitchen. 00:00:23.219 --> 00:00:25.820 It is surprisingly common. It really is. Yeah. 00:00:25.940 --> 00:00:28.920 So why is that? Today, we are diving into that 00:00:28.920 --> 00:00:31.719 exact frustration. Welcome to today's deep dive. 00:00:31.820 --> 00:00:34.100 Glad to be here. Our mission today is to explore 00:00:34.100 --> 00:00:37.340 a comprehensive Wikipedia article on natural 00:00:37.340 --> 00:00:41.500 language understanding, or NLU. We're going to 00:00:41.500 --> 00:00:43.679 figure out how scientists and engineers have 00:00:43.679 --> 00:00:46.859 been trying, honestly, for decades, to teach 00:00:46.859 --> 00:00:49.640 computers to genuinely comprehend our messy, 00:00:49.659 --> 00:00:51.939 complicated human language. And it is a massive 00:00:51.939 --> 00:00:54.020 challenge. You probably interact with some form 00:00:54.020 --> 00:00:56.060 of NLU every single day, whether you realize 00:00:56.060 --> 00:00:57.759 it or not. Oh, for sure. But, you know, there 00:00:57.759 --> 00:00:59.500 is a fundamental difference between a computer 00:00:59.500 --> 00:01:02.100 simply recording the audio of your voice and 00:01:02.100 --> 00:01:04.319 a computer actually comprehending your intent. 00:01:04.620 --> 00:01:07.290 OK, let's unpack this. Because... Natural language 00:01:07.290 --> 00:01:09.629 understanding is categorized in the sources as 00:01:09.629 --> 00:01:12.890 a very specific subset of a broader field, right? 00:01:12.950 --> 00:01:15.409 Yeah, natural language processing. Right, NLP 00:01:15.409 --> 00:01:18.790 within artificial intelligence. But NLU deals 00:01:18.790 --> 00:01:20.829 strictly with something called machine reading 00:01:20.829 --> 00:01:23.370 comprehension. Exactly. That distinction is vital. 00:01:23.769 --> 00:01:25.450 I mean, in the field of artificial intelligence, 00:01:25.650 --> 00:01:29.280 NLU is considered what we call an AI hard. problem. 00:01:29.400 --> 00:01:32.180 AI hard. I mean, that sounds deliberately intimidating. 00:01:32.379 --> 00:01:34.739 Well, it's meant to convey the sheer scale of 00:01:34.739 --> 00:01:37.540 the obstacle. It means that to truly solve this 00:01:37.540 --> 00:01:39.840 problem, you essentially have to make a computer 00:01:39.840 --> 00:01:42.420 as intelligent as a human being. Wait, really? 00:01:42.680 --> 00:01:45.219 Just to understand language? Yeah. I mean, we've 00:01:45.219 --> 00:01:48.079 actually gotten incredibly good at turning spoken 00:01:48.079 --> 00:01:51.120 audio into written text, but that's just transcription. 00:01:51.299 --> 00:01:53.939 Right, just copying it down. Exactly. NLU is 00:01:53.939 --> 00:01:55.939 about automated reasoning. It's about taking 00:01:55.939 --> 00:01:58.659 those transcribed words, extracting the actual 00:01:58.659 --> 00:02:01.239 intent, and, you know, using it for large -scale 00:02:01.239 --> 00:02:03.680 content analysis or complex question answering. 00:02:04.019 --> 00:02:06.620 It is deriving actual meaning, not just copying 00:02:06.620 --> 00:02:08.879 letters. So let me push back on this a little 00:02:08.879 --> 00:02:12.099 bit. Is this just a machine doing advanced keyword 00:02:12.099 --> 00:02:15.199 matching? How do you mean? Well, think of a student 00:02:15.199 --> 00:02:17.639 hunting for vocabulary words in a textbook without 00:02:17.639 --> 00:02:20.400 actually reading the chapter. They scan the page, 00:02:20.780 --> 00:02:22.680 find the bolded word, write down the sentence, 00:02:22.939 --> 00:02:25.580 but they have absolutely no idea what the underlying 00:02:25.580 --> 00:02:28.360 concept means. Is that what these computers are 00:02:28.360 --> 00:02:30.620 doing? Historically, and even in many common 00:02:30.620 --> 00:02:33.120 systems today, yes, that is exactly what they're 00:02:33.120 --> 00:02:35.439 doing. Just advanced keyword matching. Right. 00:02:36.120 --> 00:02:38.639 But true natural language understanding requires 00:02:38.639 --> 00:02:41.439 something much deeper than that student hunting 00:02:41.439 --> 00:02:44.039 for a bolded word. It requires the system to 00:02:44.039 --> 00:02:46.599 build an internal representation of the semantics 00:02:46.599 --> 00:02:48.580 of a sentence. Let's define that for a second, 00:02:48.939 --> 00:02:51.159 because we hear syntax and semantics thrown around 00:02:51.159 --> 00:02:54.110 a lot. Syntax is just the grammar. Right? The 00:02:54.110 --> 00:02:57.090 rules of how words are ordered. Spot on. Syntax 00:02:57.090 --> 00:02:59.689 is relatively easy for a machine. Noun, verb, 00:02:59.930 --> 00:03:02.969 adjective. Sure. But semantic is the actual logic 00:03:02.969 --> 00:03:06.270 and meaning behind those words. True NLU relies 00:03:06.270 --> 00:03:08.289 on mapping that meaning into something called 00:03:08.289 --> 00:03:11.110 first -order logic. And what does first -order 00:03:11.110 --> 00:03:14.030 logic look like in plain English? Well, it's 00:03:14.030 --> 00:03:16.830 a way of translating our messy language into 00:03:16.830 --> 00:03:19.990 strict mathematical relationships. OK. So instead 00:03:19.990 --> 00:03:22.650 of just seeing the words a dog barks, first order 00:03:22.650 --> 00:03:25.110 logic forces the computer to establish a rule. 00:03:25.189 --> 00:03:28.289 It says, for all entities, if an entity is a 00:03:28.289 --> 00:03:31.289 dog, then that entity barks. Oh, wow. It breaks 00:03:31.289 --> 00:03:33.539 it down that strictly. It has to. The machine 00:03:33.539 --> 00:03:36.180 has to explicitly understand who is doing what 00:03:36.180 --> 00:03:38.599 to whom, the relationship between the objects 00:03:38.599 --> 00:03:41.400 and the context. It cannot just rely on the fact 00:03:41.400 --> 00:03:44.020 that the sentence is grammatically correct. And 00:03:44.020 --> 00:03:46.740 because achieving that level of true reading 00:03:46.740 --> 00:03:50.479 comprehension, getting a machine to map out relationships 00:03:50.479 --> 00:03:53.750 like a human is so absurdly difficult. Early 00:03:53.750 --> 00:03:55.990 computer scientists had to rely on some pretty 00:03:55.990 --> 00:03:57.990 brilliant illusions to make computers appear 00:03:57.990 --> 00:03:59.729 as though they understood us. Oh, absolutely. 00:03:59.789 --> 00:04:02.189 They were forced to work within very strict artificial 00:04:02.189 --> 00:04:04.030 constraints. If you can't teach a machine the 00:04:04.030 --> 00:04:06.490 whole world, you teach it a tiny fake world. 00:04:06.729 --> 00:04:09.330 Which brings us to the 1960s. The early days. 00:04:09.530 --> 00:04:11.889 Yeah, the source material highlights this pioneering 00:04:11.889 --> 00:04:15.490 program called Student, written in 1964 by Daniel 00:04:15.490 --> 00:04:18.810 Bobrow at MIT. This was just eight years after 00:04:18.810 --> 00:04:21.370 the term artificial intelligence was even coined. 00:04:21.550 --> 00:04:24.160 It's wild to think of. about. And student could 00:04:24.160 --> 00:04:27.019 actually take simple natural language input and 00:04:27.019 --> 00:04:29.980 solve algebra word problems. Which was hailed 00:04:29.980 --> 00:04:32.980 as a massive breakthrough at the time. But it 00:04:32.980 --> 00:04:36.120 worked specifically because the world it was 00:04:36.120 --> 00:04:38.160 operating in algebra was extremely restricted. 00:04:38.459 --> 00:04:40.399 Right. Highly logical and mathematically predictable. 00:04:40.579 --> 00:04:45.000 Exactly. So then a year later, in 1965, another 00:04:45.000 --> 00:04:48.180 MIT researcher, Joseph Weissenbaum, writes a 00:04:48.180 --> 00:04:50.560 program called ELIZA. And this one is famous. 00:04:50.779 --> 00:04:53.500 Oh, very famous. Eliza was an interactive program 00:04:53.500 --> 00:04:55.540 that carried on a dialogue in English, acting 00:04:55.540 --> 00:04:57.839 as a psychotherapist. You would type in your 00:04:57.839 --> 00:05:00.399 problems, and Eliza would respond. And people 00:05:00.399 --> 00:05:02.660 genuinely poured their hearts out to this machine. 00:05:02.819 --> 00:05:04.699 They really did. They felt deeply understood 00:05:04.699 --> 00:05:07.079 by it. But the source notes that Weissenbaum 00:05:07.079 --> 00:05:09.660 completely sidestepped the problem of giving 00:05:09.660 --> 00:05:12.899 the program a real world database. Eliza didn't 00:05:12.899 --> 00:05:15.290 have a rich vocabulary. Not at all. So if it 00:05:15.290 --> 00:05:18.129 just used simple parsing and substituted keywords 00:05:18.129 --> 00:05:21.350 into canned phrases, wasn't Eliza basically just 00:05:21.350 --> 00:05:23.910 an interactive game of Mad Libs? Like, it was 00:05:23.910 --> 00:05:26.110 just an illusion of empathy. What's fascinating 00:05:26.110 --> 00:05:28.430 here is it was entirely an illusion. A brilliant 00:05:28.430 --> 00:05:30.850 trick. Just a trick. Yeah, if you typed, I am 00:05:30.850 --> 00:05:33.370 unhappy, Eliza would just stand for the syntax 00:05:33.370 --> 00:05:36.230 I am, flip the pronoun, grab the next word, and 00:05:36.230 --> 00:05:38.050 plug it into a pre -written template. So it would 00:05:38.050 --> 00:05:40.430 just spit out, why are you unhappy? Exactly. 00:05:40.670 --> 00:05:43.069 It had zero understanding of what happiness or 00:05:43.069 --> 00:05:45.889 unhappiness actually - meant. Yet, despite being 00:05:45.889 --> 00:05:48.430 a toy project based on keyword substitution, 00:05:49.029 --> 00:05:51.589 it became wildly popular. People loved it. They 00:05:51.589 --> 00:05:54.470 did. In many ways, Eliza served as the blueprint 00:05:54.470 --> 00:05:57.370 for early commercial customer service bots like 00:05:57.370 --> 00:06:00.670 ask .com. The illusion of understanding was enough 00:06:00.670 --> 00:06:03.509 to fool us. And the history of these restricted 00:06:03.509 --> 00:06:06.209 microworlds just gets more interesting. Like 00:06:06.209 --> 00:06:10.009 in 1971, Terry Winograd writes a program called 00:06:10.009 --> 00:06:13.889 SHRDLU. Right, that's S -H -R -D -L -U. Yes, 00:06:14.149 --> 00:06:17.129 SHRDLU. It could understand English sentences 00:06:17.129 --> 00:06:19.889 well enough to direct a robotic arm to move children's 00:06:19.889 --> 00:06:22.149 blocks around on a computer screen. It was incredibly 00:06:22.149 --> 00:06:25.389 visual. Yeah. You could literally tell it. Find 00:06:25.389 --> 00:06:27.529 a block. which is taller than the one you are 00:06:27.529 --> 00:06:30.470 holding, and put it into the box. And the machine 00:06:30.470 --> 00:06:32.970 would execute the physical action. But notice 00:06:32.970 --> 00:06:35.990 the constraint again, though. It proved a machine 00:06:35.990 --> 00:06:38.269 could parse language and map it to a physical 00:06:38.269 --> 00:06:41.310 action, but only within a restricted world of 00:06:41.310 --> 00:06:43.569 children's blocks. You didn't need to understand 00:06:43.569 --> 00:06:46.870 poetry or politics. Or sarcasm, right? It only 00:06:46.870 --> 00:06:49.129 needed to understand geometric shapes, colors, 00:06:49.290 --> 00:06:52.449 and spatial relationships, like on top of or 00:06:52.449 --> 00:06:55.230 inside. There's also this incredible tridia in 00:06:55.230 --> 00:06:58.490 the sources about the lineage of this tech. A 00:06:58.490 --> 00:07:01.529 developer named Wayne Ratliff developed a program 00:07:01.529 --> 00:07:04.899 in the late 70s called Vulcan. Oh, the Star Trek 00:07:04.899 --> 00:07:07.480 connection. Yes. He built it with an English 00:07:07.480 --> 00:07:09.720 -like syntax, specifically because he wanted 00:07:09.720 --> 00:07:12.779 to mimic the voice -activated computer from Star 00:07:12.779 --> 00:07:14.660 Trek. He wanted to talk to his computer like 00:07:14.660 --> 00:07:16.540 they did on the Enterprise. The science fiction 00:07:16.540 --> 00:07:19.399 inspiration is always there in early AI. And 00:07:19.399 --> 00:07:22.199 that program, Vulkan, eventually evolved into 00:07:22.199 --> 00:07:25.040 a system called D -Base, which practically launched 00:07:25.040 --> 00:07:27.220 the entire personal computer database industry. 00:07:27.360 --> 00:07:30.410 It's a huge legacy. Even the massive software 00:07:30.410 --> 00:07:33.490 company Symantec originally started out in 1982 00:07:33.490 --> 00:07:35.550 as a natural language understanding company. 00:07:35.889 --> 00:07:37.490 They were trying to build a natural language 00:07:37.490 --> 00:07:41.009 interface for database queries. But they pivoted 00:07:41.009 --> 00:07:43.439 away from natural language fairly quickly. Right, 00:07:43.639 --> 00:07:45.759 because the computer mouse was invented. Graphical 00:07:45.759 --> 00:07:48.079 user interfaces took over. It was just easier. 00:07:48.160 --> 00:07:50.040 Yeah, it was suddenly so much easier to just 00:07:50.040 --> 00:07:52.740 point and click a folder icon than to struggle 00:07:52.740 --> 00:07:55.879 with the friction of typing a perfectly grammatically 00:07:55.879 --> 00:07:58.779 correct English sentence to ask the computer 00:07:58.779 --> 00:08:01.139 to open a file. That makes perfect sense. I mean, 00:08:01.139 --> 00:08:03.959 why fight with early clunky NLP when a mouse 00:08:03.959 --> 00:08:05.759 click guarantees the machine understands your 00:08:05.759 --> 00:08:08.660 intent? Exactly. But moving past those early 00:08:08.660 --> 00:08:11.339 decades, researchers realized that tricks like 00:08:11.339 --> 00:08:15.620 ELISA or tiny sandboxes like SHRDLU were dead 00:08:15.620 --> 00:08:18.240 ends for real -world application. They couldn't 00:08:18.240 --> 00:08:22.160 scale it. Right. To handle actual, everyday human 00:08:22.160 --> 00:08:25.120 language, they had to build an incredibly complex 00:08:25.120 --> 00:08:27.399 architecture under the hood. Here's where it 00:08:27.399 --> 00:08:30.259 gets really interesting. Because the amount of 00:08:30.259 --> 00:08:32.960 grueling manual labor required to build these 00:08:32.960 --> 00:08:36.240 internal systems is just staggering. To even 00:08:36.240 --> 00:08:39.600 begin to understand a sentence, a modern NLU 00:08:39.600 --> 00:08:42.980 system needs a lexicon, a massive dictionary 00:08:42.980 --> 00:08:45.299 of the language. But it's not just a standard 00:08:45.299 --> 00:08:47.899 dictionary like you'd find on a bookshelf. The 00:08:47.899 --> 00:08:50.820 source specifies it requires a rich lexicon with 00:08:50.820 --> 00:08:53.639 a suitable ontology. Okay, let's pause there. 00:08:54.000 --> 00:08:56.220 What exactly is an ontology in this context? 00:08:56.429 --> 00:08:59.649 Think of an ontology as a massive, invisible 00:08:59.649 --> 00:09:02.830 web of relationships. Like a mind map. Sort of. 00:09:03.149 --> 00:09:05.009 A regular dictionary just tells you the definition 00:09:05.009 --> 00:09:07.409 of an apple. An ontology tells the computer that 00:09:07.409 --> 00:09:09.750 an apple is a type of fruit, which grows on a 00:09:09.750 --> 00:09:12.269 tree, which is a plant, and an apple is related 00:09:12.269 --> 00:09:15.269 to the action of eating. Ah, I see. It maps out 00:09:15.269 --> 00:09:17.509 how every concept in the universe connects to 00:09:17.509 --> 00:09:19.649 every other concept. And building that web is 00:09:19.649 --> 00:09:22.509 no joke. The source explicitly mentions the word 00:09:22.509 --> 00:09:26.320 net lexicon. It took many person years of pure 00:09:26.320 --> 00:09:28.879 human effort to construct. Just unbelievable 00:09:28.879 --> 00:09:31.139 amounts of labor. Human beings had to sit there, 00:09:31.240 --> 00:09:33.419 meticulously mapping out the relationships between 00:09:33.419 --> 00:09:35.799 tens of thousands of words by hand. Because that 00:09:35.799 --> 00:09:37.740 lexicon is just layer one of the architecture. 00:09:38.059 --> 00:09:40.259 Once you have the words and their relationships, 00:09:40.700 --> 00:09:43.340 you need a parser and grammar rules to break 00:09:43.340 --> 00:09:47.720 the sentences down. I picture a parser like a 00:09:47.720 --> 00:09:49.700 diagramming a sentence back in middle school 00:09:49.700 --> 00:09:53.009 English class. Remember drawing those lines to 00:09:53.009 --> 00:09:55.870 separate the subject, the verb, and the prepositional 00:09:55.870 --> 00:09:58.590 phrase. That's a great analogy. But the computer 00:09:58.590 --> 00:10:01.129 has to do that instantly, almost in 3D, keeping 00:10:01.129 --> 00:10:03.830 track of dozens of overlapping grammatical rules. 00:10:04.049 --> 00:10:07.070 That sounds exhausting. It is. And even after 00:10:07.070 --> 00:10:09.850 the parser diagrams the sentence, you arrive 00:10:09.850 --> 00:10:12.129 at the most difficult layer, the semantic theory. 00:10:12.529 --> 00:10:14.549 This is what guides the actual comprehension. 00:10:15.210 --> 00:10:17.009 Because without a semantic theory, you just have 00:10:17.009 --> 00:10:20.100 a beautifully diagrammed list of words. It's 00:10:20.100 --> 00:10:22.399 like handing someone a perfectly translated dictionary 00:10:22.399 --> 00:10:24.259 of foreign language. They have the words, but 00:10:24.259 --> 00:10:26.659 no meaning. Exactly. They might know what the 00:10:26.659 --> 00:10:28.860 words mean individually, but they don't know 00:10:28.860 --> 00:10:31.799 the culture, the idioms, or the context. They 00:10:31.799 --> 00:10:34.039 wouldn't know if a phrase was meant to be a literal 00:10:34.039 --> 00:10:37.879 instruction or a sarcastic joke. Precisely. And 00:10:37.879 --> 00:10:40.399 the source outlines that these semantic theories 00:10:40.399 --> 00:10:44.019 operate on a spectrum. On one end, you have naive 00:10:44.019 --> 00:10:46.639 semantics. Naive semantics. Yeah, this is a very 00:10:46.639 --> 00:10:48.899 literal one -to -one interpretation. If you say, 00:10:49.259 --> 00:10:52.659 my car died, naive semantics assumes your vehicle 00:10:52.659 --> 00:10:55.240 was a living biological organism that has just 00:10:55.240 --> 00:10:57.460 passed away. Which obviously causes major errors 00:10:57.460 --> 00:10:59.659 in understanding. Right. So researchers try to 00:10:59.659 --> 00:11:02.200 push toward the other end of the spectrum, utilizing 00:11:02.200 --> 00:11:05.279 something called pragmatics. Pragmatics. Pragmatics 00:11:05.279 --> 00:11:08.460 is the ability to look beyond the literal dictionary 00:11:08.460 --> 00:11:11.460 definitions and derive meaning from context and 00:11:11.460 --> 00:11:13.340 speech. speaker intent. Can you give an example? 00:11:13.580 --> 00:11:16.039 Sure. If you are sitting at a dinner table and 00:11:16.039 --> 00:11:18.980 ask, can you pass the salt? Pragmatics tells 00:11:18.980 --> 00:11:22.480 the system you are making a request, not asking 00:11:22.480 --> 00:11:24.820 a literal question about the person's physical 00:11:24.820 --> 00:11:27.320 capability to lift a salt shaker. Right. Figuring 00:11:27.320 --> 00:11:29.679 out pragmatics is hard enough for humans sometimes. 00:11:30.019 --> 00:11:32.379 So we have this massive web of relationships, 00:11:32.559 --> 00:11:35.379 the parser diagramming the grammar and the semantic 00:11:35.379 --> 00:11:38.320 theory trying to guess the context. Then what? 00:11:38.460 --> 00:11:41.399 Then advanced NLU systems attempt to incorporate 00:11:41.399 --> 00:11:44.700 logic inference. They map this derived meaning 00:11:44.700 --> 00:11:47.460 into a set of assertions using predicate logic. 00:11:47.580 --> 00:11:50.700 Okay another big term, predicate logic. It connects 00:11:50.700 --> 00:11:52.740 back to that first -order logic we discussed 00:11:52.740 --> 00:11:56.000 earlier. Predicate logic is basically the mathematical 00:11:56.000 --> 00:11:58.399 machinery used to evaluate if a statement is 00:11:58.399 --> 00:12:01.100 true or false based on the variables. So back 00:12:01.100 --> 00:12:04.039 to the barking dogs. Exactly. If the machine 00:12:04.039 --> 00:12:06.700 establishes the rule all dogs bark and you tell 00:12:06.700 --> 00:12:09.700 it Fido is a dog, the predicate logic engine 00:12:09.700 --> 00:12:13.590 deduces the new truth. phytobarx. So it's learning. 00:12:13.870 --> 00:12:16.809 It is actively deducing conclusions that weren't 00:12:16.809 --> 00:12:19.470 explicitly stated. And this is where the underlying 00:12:19.470 --> 00:12:21.610 programming language becomes a massive hurdle. 00:12:21.889 --> 00:12:23.590 Yeah, the text gets into some heavy specifics 00:12:23.590 --> 00:12:25.909 here about functional languages versus logic 00:12:25.909 --> 00:12:28.970 -oriented languages. Why does the actual coding 00:12:28.970 --> 00:12:30.950 language matter so much if they are all just 00:12:30.950 --> 00:12:33.149 running on computers? Well, it changes how the 00:12:33.149 --> 00:12:35.649 machine thinks at a foundational level. Take 00:12:35.649 --> 00:12:37.490 a functional programming language like Lisp, 00:12:37.590 --> 00:12:40.590 which was very popular in early AI. Functional 00:12:40.590 --> 00:12:42.909 languages are designed to evaluate mathematical 00:12:42.909 --> 00:12:45.649 expressions. They are not naturally built to 00:12:45.649 --> 00:12:48.419 handle logical deductions like like Fido is a 00:12:48.419 --> 00:12:54.299 dog. So how do you make it understand that? massive 00:12:54.299 --> 00:12:57.460 subsystem on top of Lisp just to represent and 00:12:57.460 --> 00:12:59.600 process those logical assertions. That sounds 00:12:59.600 --> 00:13:02.139 clunky. It is. Think of it like buying a standard 00:13:02.139 --> 00:13:04.379 car and trying to bolt airplane wings onto the 00:13:04.379 --> 00:13:07.200 roof so it can fly. Right. It's heavy, complex, 00:13:07.320 --> 00:13:09.360 and wasn't built for that purpose. But the source 00:13:09.360 --> 00:13:11.919 contrasts that with logic -oriented languages 00:13:11.919 --> 00:13:15.240 like Prologue. Right. Prologue is like buying 00:13:15.240 --> 00:13:17.879 an actual airplane. Its core foundation is already 00:13:17.879 --> 00:13:20.240 built on logical representation. Oh, I see. Yeah. 00:13:20.240 --> 00:13:22.559 In Prologue, you don't write complex functions 00:13:22.570 --> 00:13:25.169 to evaluate a phyto barks. You just declare the 00:13:25.169 --> 00:13:28.549 fact. You literally write dog phyto. The logic 00:13:28.549 --> 00:13:30.669 is baked into the foundation of the language 00:13:30.669 --> 00:13:33.610 itself. OK. So if we have all of this today, 00:13:33.690 --> 00:13:36.549 these massive word net lexicons that took years 00:13:36.549 --> 00:13:39.429 to build, 3D parsers, pragmatic semantic theories, 00:13:39.809 --> 00:13:42.129 and predicate logic mapped out in languages like 00:13:42.129 --> 00:13:45.889 prologue, why do our smart speakers still completely 00:13:45.889 --> 00:13:48.820 fail to understand us? Because of the ultimate 00:13:48.820 --> 00:13:51.740 trade -off in this entire field, every system 00:13:51.740 --> 00:13:54.159 built has to make a sacrifice on the matrix of 00:13:54.159 --> 00:13:56.720 natural language understanding. The matrix of 00:13:56.720 --> 00:14:00.179 breadth versus depth. Yes. This is the core bottleneck 00:14:00.179 --> 00:14:03.480 of the technology. When we evaluate an NLU system, 00:14:03.759 --> 00:14:07.120 we plot it on a chart with two axes. Okay, let's 00:14:07.120 --> 00:14:10.909 visualize this. is the sheer size of the vocabulary 00:14:10.909 --> 00:14:14.049 and the grammar it can handle. How wide is its 00:14:14.049 --> 00:14:17.029 knowledge of the world? And depth. Depth is the 00:14:17.029 --> 00:14:19.610 degree to which its understanding actually approximates 00:14:19.610 --> 00:14:22.769 that of a fluent native human speaker. How well 00:14:22.769 --> 00:14:25.389 does it comprehend the nuances, the pragmatics, 00:14:25.710 --> 00:14:28.009 and the hidden logic of what it knows? So if 00:14:28.009 --> 00:14:30.289 we picture this matrix, where do the things we 00:14:30.289 --> 00:14:33.250 use every day sit? Like my earlier example of 00:14:33.250 --> 00:14:34.889 telling a smart speaker to turn on the living 00:14:34.889 --> 00:14:37.230 room lights. That sits squarely in the narrow 00:14:37.230 --> 00:14:39.490 and shallow quadrant. Narrow and shallow. Yeah, 00:14:39.549 --> 00:14:41.490 it's narrow because it only knows commands related 00:14:41.490 --> 00:14:44.629 to timers, weather, and smart home devices. It's 00:14:44.629 --> 00:14:46.549 shallow because it doesn't understand the nuance 00:14:46.549 --> 00:14:49.750 of your request. It's just hunting for the keywords 00:14:49.750 --> 00:14:53.990 turn on and lights. Minimal complexity. But what 00:14:53.990 --> 00:14:56.909 about those old MIT programs? The ones stacking 00:14:56.909 --> 00:15:00.879 blocks like SHRDLU? Those occupy the narrow and 00:15:00.879 --> 00:15:03.759 deep quadrant. They were built primarily by researchers 00:15:03.759 --> 00:15:06.799 to explore the actual mechanisms of human understanding. 00:15:07.360 --> 00:15:11.480 They are incredibly deep. SHRDLU deeply understood 00:15:11.480 --> 00:15:13.940 the spatial relationships and semantics of its 00:15:13.940 --> 00:15:16.830 environment. but their application is impossibly 00:15:16.830 --> 00:15:20.649 narrow. It only knows about blocks. So if SHRDLU 00:15:20.649 --> 00:15:23.649 was so good at deeply understanding blocks, why 00:15:23.649 --> 00:15:25.830 couldn't researchers just add more words to its 00:15:25.830 --> 00:15:28.529 dictionary? Why not just expand its ontology 00:15:28.529 --> 00:15:31.070 to include cars and restaurants and emotions 00:15:31.070 --> 00:15:33.210 and make it understand the whole world? Because 00:15:33.210 --> 00:15:35.429 the moment you expand the vocabulary, the amount 00:15:35.429 --> 00:15:37.769 of ambiguity explodes exponentially. Oh, I didn't 00:15:37.769 --> 00:15:40.190 think about that. Yeah. Block has one meaning 00:15:40.190 --> 00:15:42.950 in a sandbox. In the real world, block could 00:15:42.950 --> 00:15:45.929 mean a toy, a city street, a mental block, or 00:15:45.929 --> 00:15:48.549 to obstruct someone's path. The deep systems 00:15:48.549 --> 00:15:50.909 collapse under the weight of real -world ambiguity. 00:15:51.029 --> 00:15:53.129 Which forces companies to build systems in the 00:15:53.129 --> 00:15:56.169 broad and shallow quadrant. Exactly. This is 00:15:56.169 --> 00:15:58.669 where massive commercial enterprise systems live. 00:15:59.149 --> 00:16:01.730 Imagine a program that automatically reads and 00:16:01.730 --> 00:16:04.110 routes millions of corporate customer service 00:16:04.110 --> 00:16:06.590 emails to the right department. Like reaching 00:16:06.590 --> 00:16:09.120 billing instead of tech support. Right. It has 00:16:09.120 --> 00:16:11.519 to deal with a massive real -world vocabulary. 00:16:11.840 --> 00:16:15.960 It is very broad, but it lacks true comprehension. 00:16:16.600 --> 00:16:19.539 It's not deeply analyzing the customer's childhood 00:16:19.539 --> 00:16:23.039 trauma or sarcasm. It's just classifying text 00:16:23.039 --> 00:16:25.419 based on statistical patterns. It is shallow. 00:16:25.620 --> 00:16:28.919 Which leaves the final Holy Grail quadrant broad 00:16:28.919 --> 00:16:31.600 and deep. A system that knows every word and 00:16:31.600 --> 00:16:33.740 topic in the human experience and understands 00:16:33.740 --> 00:16:36.200 them with the pragmatic depth of a fluent native 00:16:36.200 --> 00:16:39.070 speaker. The dream of AI. But the source material 00:16:39.070 --> 00:16:41.269 is very clear on this. Systems that are both 00:16:41.269 --> 00:16:43.929 very broad and very deep are currently beyond 00:16:43.929 --> 00:16:45.970 the state of the art. Completely beyond it right 00:16:45.970 --> 00:16:47.950 now. So what does this all mean? Because I have 00:16:47.950 --> 00:16:49.830 to push back here on the idea that broad and 00:16:49.830 --> 00:16:52.110 deep doesn't exist. What about a system like 00:16:52.110 --> 00:16:55.289 IBM Watson? Ah. The source brings up Watson using 00:16:55.289 --> 00:16:59.240 machine learning for text classification. When 00:16:59.240 --> 00:17:01.700 Watson went on Jeopardy and beat the human champions, 00:17:02.299 --> 00:17:04.940 didn't that prove machines can understand broad 00:17:04.940 --> 00:17:08.279 and deep topics? It was answering incredibly 00:17:08.279 --> 00:17:10.700 complex, pun -filled questions about history, 00:17:11.259 --> 00:17:13.579 science, and pop culture. If we connect this 00:17:13.579 --> 00:17:15.819 to the bigger picture, that specific question 00:17:15.819 --> 00:17:18.299 is the philosophical debate tearing the AI field 00:17:18.299 --> 00:17:21.559 apart right now. Really? Yes. Did Watson actually 00:17:21.559 --> 00:17:24.059 understand those Jeopardy clues, or is it just 00:17:24.059 --> 00:17:27.069 incredibly good at statistical prediction? The 00:17:27.069 --> 00:17:29.230 source cites the famous philosopher John Searle, 00:17:29.450 --> 00:17:32.009 who argued explicitly that Watson did not understand 00:17:32.009 --> 00:17:34.230 the questions it was answering. How so? It gave 00:17:34.230 --> 00:17:36.230 the right answers. It gave the right answers 00:17:36.230 --> 00:17:39.009 by computing probabilities based on vast amounts 00:17:39.009 --> 00:17:41.549 of data, not by understanding meaning. I'm not 00:17:41.549 --> 00:17:44.339 sure I follow. Think about the word bank. When 00:17:44.339 --> 00:17:46.980 Watson reads a clue with the word bank, it doesn't 00:17:46.980 --> 00:17:49.160 vividly imagine a brick building with a vault 00:17:49.160 --> 00:17:52.140 and security guards. It simply calculates that 00:17:52.140 --> 00:17:55.579 in its massive database of text, the word bank 00:17:55.579 --> 00:17:59.480 appears next to the word money, 85 % of the time, 00:17:59.700 --> 00:18:02.759 and next to the word river, 15 % of the time. 00:18:02.900 --> 00:18:05.720 Oh, wow. Yeah, it uses statistics to guess the 00:18:05.720 --> 00:18:08.559 most probable missing piece of the puzzle. So 00:18:08.559 --> 00:18:11.119 we are right back to the student hunting for 00:18:11.119 --> 00:18:13.680 vocabulary words. Watson is just doing it at 00:18:13.680 --> 00:18:16.619 lightning speed across a billion textbooks simultaneously. 00:18:17.400 --> 00:18:19.500 Exactly. And the source includes the perspective 00:18:19.500 --> 00:18:21.980 of cognitive scientist John Ball, who invented 00:18:21.980 --> 00:18:25.759 Patom Theory. He strongly supports Searle's assessment. 00:18:26.099 --> 00:18:28.720 What does he argue? Ball argues that NLP has 00:18:28.720 --> 00:18:31.200 largely succeeded today by secretly narrowing 00:18:31.200 --> 00:18:33.519 the scope of the application or by relying on 00:18:33.519 --> 00:18:36.440 those statistical parlor tricks, not by actually 00:18:36.440 --> 00:18:39.720 achieving broad and deep comprehension. Right. 00:18:39.859 --> 00:18:42.259 Because there are still thousands of ways a human 00:18:42.259 --> 00:18:44.579 can request something that completely defies 00:18:44.579 --> 00:18:47.859 conventional statistical NLP. Because human language 00:18:47.859 --> 00:18:49.940 isn't just a math equation of probabilities. 00:18:50.480 --> 00:18:53.720 It relies on shared physical context, body language, 00:18:54.220 --> 00:18:56.960 cultural shifts, and shifting definitions. The 00:18:56.960 --> 00:18:59.680 article quotes an executive named Wide Wagemans 00:18:59.680 --> 00:19:02.380 on this exact point. He says that to have a truly 00:19:02.380 --> 00:19:05.279 meaningful conversation with a machine, the computer 00:19:05.279 --> 00:19:08.160 must match every single word to the correct meaning, 00:19:08.680 --> 00:19:10.700 based entirely on the meanings of the other words 00:19:10.700 --> 00:19:13.380 in the sentence without relying on statistical 00:19:13.380 --> 00:19:16.240 guesswork. Which, based on everything we've discussed 00:19:16.240 --> 00:19:19.039 about parsers and ontologies and breadth versus 00:19:19.039 --> 00:19:22.240 depth, sounds technologically impossible. But 00:19:22.240 --> 00:19:24.819 as Wagemans points out, that impossible standard 00:19:24.819 --> 00:19:26.900 is exactly what a three -year -old human does 00:19:26.900 --> 00:19:29.539 naturally. A three -year -old? Yeah. A toddler 00:19:29.539 --> 00:19:32.579 can understand shifting context, tone, and intent 00:19:32.579 --> 00:19:35.279 without running massive statistical probability 00:19:35.279 --> 00:19:38.380 calculations on a server farm. They just understand. 00:19:38.680 --> 00:19:41.400 A machine currently cannot. That is incredibly 00:19:41.400 --> 00:19:43.579 humbling. We have these massive supercomputers, 00:19:43.640 --> 00:19:46.099 and they still can't match the true semantic 00:19:46.099 --> 00:19:48.849 comprehension of a toddler. This debate about 00:19:48.849 --> 00:19:51.210 whether machines truly understand our logic or 00:19:51.210 --> 00:19:53.730 just predict our syntax really highlights where 00:19:53.730 --> 00:19:55.849 we are in the history of computing. It is a profound 00:19:55.849 --> 00:19:57.849 reality check on our technological progress. 00:19:58.210 --> 00:20:00.670 So let's recap this journey we've been on. We 00:20:00.670 --> 00:20:03.890 started in the 1960s with Eliza, proving that 00:20:03.890 --> 00:20:06.990 simple keyword substitution could create a powerful, 00:20:07.210 --> 00:20:10.049 albeit fake, illusion of empathy. A brilliant 00:20:10.049 --> 00:20:11.910 trick. We moved through the decades, discovering 00:20:11.910 --> 00:20:14.490 the grueling, multi -year manual labor required 00:20:14.490 --> 00:20:17.329 to build vast ontological lexicons like WordNet, 00:20:17.789 --> 00:20:20.349 and the necessity of establishing pragmatic semantic 00:20:20.349 --> 00:20:23.130 theories and predicate logic systems. We witnessed 00:20:23.130 --> 00:20:27.250 the evolution from clever parlor tricks to incredibly 00:20:27.250 --> 00:20:30.019 dense structural architecture. And yet to Today, 00:20:30.380 --> 00:20:32.900 we are still caught in that ultimate matrix of 00:20:32.900 --> 00:20:35.740 breadth versus depth. We have systems that are 00:20:35.740 --> 00:20:38.380 broad enough to scan millions of emails and systems 00:20:38.380 --> 00:20:40.140 that are deep enough to perfectly understand 00:20:40.140 --> 00:20:43.019 a tiny sandbox, but combining them into a machine 00:20:43.019 --> 00:20:45.099 that can truly converse about the whole world 00:20:45.099 --> 00:20:48.160 remains the holy grail currently beyond our reach. 00:20:48.779 --> 00:20:50.740 The three -year -old standard remains unbroken. 00:20:50.839 --> 00:20:53.900 It certainly does. So, for you listening, the 00:20:53.900 --> 00:20:56.079 next time an automated system magically routes 00:20:56.079 --> 00:20:58.599 your customer service email to the perfect apartment, 00:20:58.859 --> 00:21:02.359 or completely fails to grasp a simple spoken 00:21:02.359 --> 00:21:05.299 command to set a kitchen timer, you now know 00:21:05.299 --> 00:21:07.660 exactly what is happening under the hood. You've 00:21:07.660 --> 00:21:09.880 seen the matrix. Yeah, you know exactly where 00:21:09.880 --> 00:21:12.200 that specific program sits on the breadth versus 00:21:12.200 --> 00:21:15.299 depth chart. You know it's missing the complex 00:21:15.299 --> 00:21:18.220 semantic modeling of a fluent speaker and why 00:21:18.220 --> 00:21:21.319 it is still struggling to do what a toddler does 00:21:21.319 --> 00:21:24.119 effortlessly. This raises an important question 00:21:24.119 --> 00:21:25.880 and it's something to think about the next time 00:21:25.880 --> 00:21:28.200 you interact with one of these systems. What's 00:21:28.200 --> 00:21:31.289 that? Well, if achieving that true broad and 00:21:31.289 --> 00:21:33.730 deep machine understanding is currently beyond 00:21:33.730 --> 00:21:36.450 our engineering capabilities and computers constantly 00:21:36.450 --> 00:21:38.910 struggle with our messy context -heavy language, 00:21:39.289 --> 00:21:41.089 will we eventually stop waiting for the machines 00:21:41.089 --> 00:21:43.130 to learn our language? That's interesting. Right. 00:21:43.269 --> 00:21:45.950 Or will we subconsciously standardize and simplify 00:21:45.950 --> 00:21:48.950 the way we speak, smoothing out our own messy 00:21:48.950 --> 00:21:52.170 human quirks just to ensure the algorithms finally 00:21:52.170 --> 00:21:55.539 understand us? Wow. We started by saying we like 00:21:55.539 --> 00:21:57.779 computers because they are exact and predictable, 00:21:58.079 --> 00:22:01.380 like a rigid calculator. But maybe, in our desperation 00:22:01.380 --> 00:22:03.539 to be understood by them, we are the ones who 00:22:03.539 --> 00:22:05.700 will start talking like the calculators. Thank 00:22:05.700 --> 00:22:07.819 you for joining us on this deep dive. We will 00:22:07.819 --> 00:22:08.480 see you next time.