WEBVTT 00:00:00.000 --> 00:00:01.860 Have you ever been on vacation maybe sitting 00:00:01.860 --> 00:00:05.620 in a little cafe somewhere and you're staring 00:00:05.620 --> 00:00:08.240 at a menu that might as well be written in hieroglyphics. 00:00:08.519 --> 00:00:10.820 Oh, absolutely. Right. So you pull out your phone, 00:00:11.199 --> 00:00:14.199 you open up a free online translation tool, and 00:00:14.199 --> 00:00:16.839 you just point your camera at it. And for a second, 00:00:17.039 --> 00:00:19.719 it feels like absolute magic. Like you have a 00:00:19.719 --> 00:00:22.179 superpower or something. Exactly. You can suddenly 00:00:22.179 --> 00:00:24.719 read the words. But then, you know, you look 00:00:24.719 --> 00:00:27.800 a little closer, and the app is cheerfully offering 00:00:27.800 --> 00:00:31.640 you a plate of stir -fried Wikipedia with pimientos. 00:00:31.820 --> 00:00:33.920 Yeah, the illusion of magic shatters instantly, 00:00:34.100 --> 00:00:36.759 right? there, or you're trying to read an international 00:00:36.759 --> 00:00:39.420 news site and halfway through this really serious 00:00:39.420 --> 00:00:42.619 political article, the text suddenly spouts total 00:00:42.619 --> 00:00:45.439 surreal nonsense. Just complete gibberish. Right. 00:00:45.659 --> 00:00:48.600 And you realize you aren't talking to some bilingual 00:00:48.600 --> 00:00:51.240 genius. You're basically talking to a calculator 00:00:51.240 --> 00:00:54.100 that just divided by zero. That broken magic 00:00:54.100 --> 00:00:56.759 is exactly what we are getting into today. So 00:00:56.759 --> 00:00:59.200 welcome to this deep dive into the Wikipedia 00:00:59.200 --> 00:01:01.979 article on machine translation. It's a huge topic. 00:01:02.189 --> 00:01:04.890 It really is. Our mission today is to explore 00:01:04.890 --> 00:01:07.469 the incredibly fascinating and frankly often 00:01:07.469 --> 00:01:10.170 frustrating journey of teaching computers to 00:01:10.170 --> 00:01:12.170 understand human language. Yeah, and we're going 00:01:12.170 --> 00:01:15.709 to discover why true human parity in translation 00:01:15.709 --> 00:01:18.409 is, well, it's still an illusion. And we'll unpack 00:01:18.409 --> 00:01:21.010 the high stakes real world consequences of relying 00:01:21.010 --> 00:01:23.049 on machines to speak for us. Because when we 00:01:23.049 --> 00:01:25.109 discuss machine translation, you know, we aren't 00:01:25.109 --> 00:01:27.430 just looking at lines of code. No, not at all. 00:01:27.569 --> 00:01:30.670 We're looking at how machines try to mathematically 00:01:30.670 --> 00:01:34.980 quantify the cultural, emotional, and even contextual 00:01:34.980 --> 00:01:37.480 depths of how you and I actually communicate. 00:01:37.680 --> 00:01:41.120 Which is wild. Now I know the early tech pioneers 00:01:41.120 --> 00:01:43.340 in the 1950s tried to tackle this problem with 00:01:43.340 --> 00:01:46.439 those first room -sized computers, but didn't 00:01:46.439 --> 00:01:48.579 philosophers dream about universal languages 00:01:48.579 --> 00:01:50.819 way before we even had the hardware? Oh, long 00:01:50.819 --> 00:01:53.480 before. I mean, back in 1629, Rene Descartes 00:01:53.480 --> 00:01:55.500 proposed this really beautiful philosophical 00:01:55.500 --> 00:01:58.260 concept of a universal language. Okay, 1629, 00:01:58.459 --> 00:02:00.780 wow. Yeah, you imagine a system where different 00:02:00.780 --> 00:02:03.140 tongues would share one single symbol for equivalent 00:02:03.140 --> 00:02:07.420 ideas. But if you want to find the actual mathematical 00:02:07.420 --> 00:02:10.259 roots of how today's apps work, we have to look 00:02:10.259 --> 00:02:12.819 past the philosophers and find the code breakers. 00:02:12.860 --> 00:02:15.460 Code breakers. Specifically, a ninth century 00:02:15.460 --> 00:02:19.030 Arabic cryptographer named Al -Kindi. He developed 00:02:19.030 --> 00:02:22.150 techniques for breaking secret codes using frequency 00:02:22.150 --> 00:02:24.830 analysis and probability. Wait, breaking secret 00:02:24.830 --> 00:02:27.729 codes? So the foundation of translating a French 00:02:27.729 --> 00:02:31.270 poem is the exact same math used to intercept 00:02:31.270 --> 00:02:33.610 enemy spy communications? It is the exact same 00:02:33.610 --> 00:02:36.169 math. Right. And that mindset carried over directly 00:02:36.169 --> 00:02:38.710 into the mid -20th century. Really? Yeah. So 00:02:38.710 --> 00:02:42.370 1947 and 1949, researchers like A .D. Booth in 00:02:42.370 --> 00:02:44.689 England and Warren Weaver at the Rockefeller 00:02:44.689 --> 00:02:47.490 Foundation, they proposed using digital computers. 00:02:47.340 --> 00:02:50.340 is to translate natural human languages. OK. 00:02:50.599 --> 00:02:53.099 Weaver actually wrote a highly influential memorandum 00:02:53.099 --> 00:02:56.219 in 1949 that essentially framed foreign languages 00:02:56.219 --> 00:02:59.060 not as distinct cultures, but basically as encrypted 00:02:59.060 --> 00:03:01.639 English. OK, let's unpack this. The early pioneers 00:03:01.639 --> 00:03:03.900 basically treated foreign languages like secret 00:03:03.900 --> 00:03:06.319 enemy codes to be cracked. They thought if you 00:03:06.319 --> 00:03:08.560 just found the right decoder ring, the language 00:03:08.560 --> 00:03:11.680 would be solved. But language isn't a static 00:03:11.680 --> 00:03:14.689 code, right? It's a living, breathing thing. 00:03:14.789 --> 00:03:16.590 Right. And they suffered from an incredible amount 00:03:16.590 --> 00:03:20.069 of hubris about it. In the 1950s, there was this 00:03:20.069 --> 00:03:22.650 massive optimism. I can imagine. There was a 00:03:22.650 --> 00:03:25.909 famous 1954 public demonstration by a Georgetown 00:03:25.909 --> 00:03:30.030 University research team and IBM. They fed Russian 00:03:30.030 --> 00:03:33.169 sentences into a massive computer, and it spit 00:03:33.169 --> 00:03:35.569 out English translations. And people bought it. 00:03:35.810 --> 00:03:38.669 Oh, the press went wild. Funding just poured 00:03:38.669 --> 00:03:41.960 in from all over the US, Japan, and Russia. The 00:03:41.960 --> 00:03:44.180 general consensus was that the entire language 00:03:44.180 --> 00:03:46.719 barrier problem would be permanently solved in, 00:03:46.719 --> 00:03:49.900 like, three to five years. Wow. Spoiler alert 00:03:49.900 --> 00:03:51.960 for anyone who has ever used an online translator. 00:03:52.180 --> 00:03:54.120 It wasn't solved in three years. Not even close. 00:03:54.479 --> 00:03:57.439 By 1966, the National Academy of Sciences formed 00:03:57.439 --> 00:03:59.740 a committee called ELAPAC to review the progress. 00:03:59.819 --> 00:04:01.840 And how does that go? They released a report 00:04:01.840 --> 00:04:04.139 that basically brought the entire industry crashing 00:04:04.139 --> 00:04:07.490 down. It concluded that a solid decade of incredibly 00:04:07.490 --> 00:04:10.289 expensive research had completely failed. Yeah. 00:04:10.289 --> 00:04:12.889 Yeah, the translations were terrible. They required 00:04:12.889 --> 00:04:15.870 massive human editing to even be legible. So 00:04:15.870 --> 00:04:17.850 the government drastically reduced all funding 00:04:17.850 --> 00:04:20.470 and the field went into a sort of dark age. So 00:04:20.470 --> 00:04:23.449 it forced a total reinvention. Once computer 00:04:23.449 --> 00:04:26.870 scientists recovered from that 1966 funding crash, 00:04:27.389 --> 00:04:29.730 they realized that treating language like a static 00:04:29.730 --> 00:04:32.069 cryptography puzzle was a complete dead end. 00:04:32.189 --> 00:04:35.069 Right, they needed entirely new frameworks, which 00:04:35.069 --> 00:04:37.649 led to three distinct eras of technology. Okay, 00:04:37.689 --> 00:04:39.449 let's go through them. What was the first era? 00:04:39.870 --> 00:04:42.810 The first era, which dominated for decades after 00:04:42.810 --> 00:04:46.639 that crash, was the rule -based approach. Programmers 00:04:46.639 --> 00:04:49.180 essentially relied on building massive electronic 00:04:49.180 --> 00:04:52.019 dictionaries and hard -coding every single grammar 00:04:52.019 --> 00:04:54.120 rule. Every single one? That sounds impossible. 00:04:54.199 --> 00:04:56.920 It was incredibly tedious. If there was an orthographical 00:04:56.920 --> 00:04:58.860 variation, meaning just a different spelling 00:04:58.860 --> 00:05:01.680 of the same word, like color with it, U in British 00:05:01.680 --> 00:05:04.439 English versus American English, a human programmer 00:05:04.439 --> 00:05:07.060 had to write a specific lexical selection rule. 00:05:07.319 --> 00:05:10.180 Oh my gosh. Yeah, that rule explicitly told the 00:05:10.180 --> 00:05:12.199 computer exactly which dictionary definition 00:05:12.199 --> 00:05:15.009 to pick based on the word surrounding it. Piping 00:05:15.009 --> 00:05:17.490 out millions of individual grammar rules sounds 00:05:17.490 --> 00:05:21.029 like a total nightmare. Did it actually work? 00:05:21.370 --> 00:05:23.750 In highly controlled, very narrow environments, 00:05:23.910 --> 00:05:26.879 yes it did. There is a system called CANT, designed 00:05:26.879 --> 00:05:30.079 in the early 90s, specifically to translate something 00:05:30.079 --> 00:05:33.199 called Caterpillar Technical English. Caterpillar 00:05:33.199 --> 00:05:35.959 like the tractor. Exactly. If you are translating 00:05:35.959 --> 00:05:38.720 tractor repair manuals where the vocabulary is 00:05:38.720 --> 00:05:40.480 restricted and the sentence structures are super 00:05:40.480 --> 00:05:44.139 simple and rigid, rule -based systems yield very 00:05:44.139 --> 00:05:46.199 stable results. Sure, because a tractor part 00:05:46.199 --> 00:05:48.699 is always a tractor part. Right. but you cannot 00:05:48.699 --> 00:05:51.879 program a rule for every single weird exception, 00:05:52.180 --> 00:05:55.240 slang word, or idiom in everyday human speech. 00:05:55.819 --> 00:05:58.500 The system just breaks under the weight of human 00:05:58.500 --> 00:06:01.420 unpredictability. So if rule based is like giving 00:06:01.420 --> 00:06:03.959 the computer a massive grammar textbook and telling 00:06:03.959 --> 00:06:06.540 it to study, how did they pivot for the second 00:06:06.540 --> 00:06:08.660 era? Did they just, I don't know, build a bigger 00:06:08.660 --> 00:06:10.959 textbook? What's fascinating here is the complete 00:06:10.959 --> 00:06:13.399 shift in philosophy. They stopped telling the 00:06:13.399 --> 00:06:15.699 computer how to translate and instead let the 00:06:15.699 --> 00:06:18.579 computer guess how to translate based on massive 00:06:18.579 --> 00:06:21.300 amounts of data. Oh, I see. This was the statistical 00:06:21.300 --> 00:06:24.240 machine translation era, which really took off 00:06:24.240 --> 00:06:26.579 as computing power got cheaper in the late 80s 00:06:26.579 --> 00:06:28.860 and 90s. So it's like dropping the computer. 00:06:28.839 --> 00:06:31.480 in the middle of Paris with a million translated 00:06:31.480 --> 00:06:33.680 documents and saying, figure out the patterns 00:06:33.680 --> 00:06:36.480 yourself. Precisely. They used bilingual text 00:06:36.480 --> 00:06:39.300 corpora, which are massive collections of documents 00:06:39.300 --> 00:06:41.660 already perfectly translated by humans. Like 00:06:41.660 --> 00:06:44.019 what kind of documents? A famous early example 00:06:44.019 --> 00:06:46.759 is the Canadian Hansard corpus. It's the official 00:06:46.759 --> 00:06:49.040 record of the Canadian Parliament, transcribed 00:06:49.040 --> 00:06:52.180 in both English and French. The computer analyzes 00:06:52.180 --> 00:06:55.000 these parallel texts and calculates statistical 00:06:55.000 --> 00:06:57.920 probability. Okay, so just running the numbers. 00:06:58.019 --> 00:07:01.040 Right. It simply notices that when the word taxes 00:07:01.040 --> 00:07:04.120 appears in English, a specific French word appears 00:07:04.120 --> 00:07:06.939 in the corresponding sentence, like 98 % of the 00:07:06.939 --> 00:07:09.139 time. And that's where Google completely changed 00:07:09.139 --> 00:07:12.339 the paradigm, right? In 2005, Google fed its 00:07:12.339 --> 00:07:15.680 internal system approximately 200 billion words 00:07:15.680 --> 00:07:18.199 from United Nations materials. Yep, 200 billion. 00:07:18.519 --> 00:07:22.509 200 billion? But wait, I have to ask. If Google 00:07:22.509 --> 00:07:25.430 fed its system 200 billion words from the UN 00:07:25.430 --> 00:07:28.769 back in 2005, shouldn't that sheer volume of 00:07:28.769 --> 00:07:31.350 data have permanently solved translation? Why 00:07:31.350 --> 00:07:34.230 wasn't 200 billion words enough to capture every 00:07:34.230 --> 00:07:36.689 nuance? Well, because UN documents are highly 00:07:36.689 --> 00:07:39.889 formal. If you train a machine purely on diplomatic 00:07:39.889 --> 00:07:42.470 treaties, it learns to sound exactly like a diplomat. 00:07:42.610 --> 00:07:44.490 Oh, right, so it doesn't know how normal people 00:07:44.490 --> 00:07:47.290 talk. Exactly. It will completely fail to translate 00:07:47.290 --> 00:07:50.050 a teenager texting their friend. Furthermore, 00:07:50.189 --> 00:07:52.009 statistical models fundamentally struggle with 00:07:52.009 --> 00:07:54.810 morphology -rich languages. Which are what? Languages 00:07:54.810 --> 00:07:56.769 where words change their entire spelling based 00:07:56.769 --> 00:08:00.490 on gender, tense, or case. The math just couldn't 00:08:00.490 --> 00:08:02.829 stretch far enough, leading to the third and 00:08:02.829 --> 00:08:05.550 current era neural machine translation. Right, 00:08:05.550 --> 00:08:08.370 the 2020s deep learning era. This is the architecture 00:08:08.370 --> 00:08:11.930 behind large language models or LLMs like ChatGPT 00:08:11.930 --> 00:08:14.189 and specialized translation tools like DeepL. 00:08:14.370 --> 00:08:16.990 Exactly. But how is a neural network actually 00:08:16.990 --> 00:08:18.850 different from just running better statistics? 00:08:19.120 --> 00:08:22.879 Imagine a giant multi -dimensional map. A neural 00:08:22.879 --> 00:08:25.779 network plots every single word as a spatial 00:08:25.779 --> 00:08:28.540 coordinate on that map. It understands words 00:08:28.540 --> 00:08:31.199 as vectors of meaning. Spatial coordinates. Yeah. 00:08:31.620 --> 00:08:35.860 So the word king and the word queen end up geographically 00:08:35.860 --> 00:08:38.000 close to each other on this map because they 00:08:38.000 --> 00:08:40.879 share royal contexts. Oh, I get it. The network 00:08:40.879 --> 00:08:43.120 doesn't just match words anymore. It predicts 00:08:43.120 --> 00:08:45.360 the likelihood of an entire sequence of words 00:08:45.360 --> 00:08:47.899 based on these spatial relationships. Which leads 00:08:47.899 --> 00:08:49.919 to all those marketing claims we see today about 00:08:49.919 --> 00:08:53.200 AI achieving human parity in translation. Claims 00:08:53.200 --> 00:08:55.940 that researchers overwhelmingly agree are a complete 00:08:55.940 --> 00:08:59.879 illusion. Really? Even now? Even now. The sources 00:08:59.879 --> 00:09:01.980 point out that human parity claims are based 00:09:01.980 --> 00:09:05.019 entirely on limited domains, specific language 00:09:05.019 --> 00:09:08.210 pairs, and very narrow test benchmarks. they 00:09:08.210 --> 00:09:10.889 lack true statistical significance power. So 00:09:10.889 --> 00:09:12.690 they're kind of cherry -picking the data to look 00:09:12.690 --> 00:09:15.070 good. Pretty much. Even with an advanced tool 00:09:15.070 --> 00:09:17.789 like DeepL, the outputs almost always require 00:09:17.789 --> 00:09:20.850 post -editing by a human to fix glaring contextual 00:09:20.850 --> 00:09:23.850 errors. Because data volume cannot easily solve 00:09:23.850 --> 00:09:26.750 contextual ambiguity. The machine might have 00:09:26.750 --> 00:09:29.529 a perfect map of the vocabulary, but it has no 00:09:29.529 --> 00:09:32.460 map of the real world. Exactly. This ambiguity 00:09:32.460 --> 00:09:34.559 wall is actually my favorite part of the deep 00:09:34.559 --> 00:09:37.120 dive because it proves how incredibly complex 00:09:37.120 --> 00:09:40.179 human brains really are. The issue of disambiguation 00:09:40.179 --> 00:09:43.519 was actually raised way back in the 1950s by 00:09:43.519 --> 00:09:46.360 a researcher named Yehoshua Bar -Hilal. What 00:09:46.360 --> 00:09:50.179 did he say? He pointed out that without a universal 00:09:50.179 --> 00:09:53.259 encyclopedia, of common sense embedded in its 00:09:53.259 --> 00:09:55.500 programming. A machine will never be able to 00:09:55.500 --> 00:09:57.179 distinguish between two completely different 00:09:57.179 --> 00:09:59.980 meanings of the exact same word. Claude Piron, 00:10:00.299 --> 00:10:02.320 who was a longtime translator for the United 00:10:02.320 --> 00:10:04.960 Nations and the World Health Organization, he 00:10:04.960 --> 00:10:07.139 had a brilliant example of this. He pointed out 00:10:07.139 --> 00:10:09.639 the phrase Japanese prisoners of war camp. Right. 00:10:09.820 --> 00:10:12.700 A perfect example is the text referring to an 00:10:12.700 --> 00:10:14.759 American camp that is holding Japanese prisoners. 00:10:14.820 --> 00:10:16.740 Or is it a Japanese camp that is holding American 00:10:16.740 --> 00:10:19.279 prisoners? Because both interpretations are grammatically 00:10:19.279 --> 00:10:22.340 identical in English. Right. A machine just looks 00:10:22.340 --> 00:10:25.159 at its spatial map, checks statistical probabilities, 00:10:25.600 --> 00:10:28.080 and takes a blind guess based on what it saw 00:10:28.080 --> 00:10:30.919 most often in its training data. But Purin pointed 00:10:30.919 --> 00:10:33.539 out that a human translator encounters that phrase, 00:10:34.379 --> 00:10:37.740 realizes the historical context is missing, and 00:10:37.740 --> 00:10:40.019 literally picks up the phone to call an expert 00:10:40.019 --> 00:10:42.980 in Australia to research the specific World War 00:10:42.980 --> 00:10:45.600 II epidemic being referenced. And a machine cannot 00:10:45.600 --> 00:10:47.940 make that phone call. No, it can't. That lack 00:10:47.940 --> 00:10:51.039 of real world context heavily impacts how machines 00:10:51.039 --> 00:10:54.580 handle named entities too, like people's names, 00:10:54.779 --> 00:10:57.480 organizations, or locations. They really struggle 00:10:57.480 --> 00:10:59.899 with those, don't they? Constantly. The system 00:10:59.899 --> 00:11:01.960 never quite knows when to transliterate versus 00:11:01.960 --> 00:11:04.259 when to translate. Transliteration is simply 00:11:04.259 --> 00:11:06.340 finding the corresponding phonetic letters in 00:11:06.340 --> 00:11:09.440 the target alphabet, whereas translation is converting 00:11:09.440 --> 00:11:12.169 the actual meaning of the word. The classic example 00:11:12.169 --> 00:11:15.149 in the sources is Southern California. The machine 00:11:15.149 --> 00:11:17.730 needs to translate Southern because that's a 00:11:17.730 --> 00:11:20.429 directional descriptor, but it needs to transliterate 00:11:20.429 --> 00:11:23.210 California because that's a proper noun. Right. 00:11:23.570 --> 00:11:25.970 And machines frequently get confused and treat 00:11:25.970 --> 00:11:28.649 them as one single block, either translating 00:11:28.649 --> 00:11:31.509 both or transliterating both, yielding total 00:11:31.509 --> 00:11:34.289 gibberish in the target language. Even more fascinating 00:11:34.289 --> 00:11:37.470 is the statistical quirk within those named entities. 00:11:38.230 --> 00:11:40.570 A Stanford study found that if you ask a machine 00:11:40.519 --> 00:11:42.980 to translate the sentence, Ted is going for a 00:11:42.980 --> 00:11:45.259 walk, it will assign a different mathematical 00:11:45.259 --> 00:11:47.759 probability score and potentially a different 00:11:47.759 --> 00:11:50.559 translation structure than if you ask it to translate, 00:11:51.259 --> 00:11:53.840 Erica is going for a walk. Purely based on the 00:11:53.840 --> 00:11:55.940 name. Here's where it gets really interesting. 00:11:56.179 --> 00:11:58.700 It's like a bouncer at a club who lets Ted walk 00:11:58.700 --> 00:12:01.379 right through the door, but stops Erica to check 00:12:01.379 --> 00:12:04.059 her ID purely because he's seen more guys named 00:12:04.059 --> 00:12:05.960 Ted that week. That's a great way to put it. 00:12:06.100 --> 00:12:08.659 The machine's output is shaped entirely by its 00:12:08.659 --> 00:12:12.129 training diet. If the name Ted appeared 10 ,000 00:12:12.129 --> 00:12:14.570 times in the data and Erica only appeared 50 00:12:14.570 --> 00:12:17.389 times, the neural network weighs those completely 00:12:17.389 --> 00:12:20.129 equal sentences differently. The system also 00:12:20.129 --> 00:12:22.769 stumbles massively outside of standard formal 00:12:22.769 --> 00:12:25.879 language. Because these tools are trained overwhelmingly 00:12:25.879 --> 00:12:28.200 on government records, published books, and news 00:12:28.200 --> 00:12:31.259 articles, they fail aggressively when faced with 00:12:31.259 --> 00:12:34.639 vernacular, slang, or just the casual way people 00:12:34.639 --> 00:12:37.440 type on their mobile phones. Right. And the Ted 00:12:37.440 --> 00:12:41.279 versus Erica math quirk, or a failure to grasp 00:12:41.279 --> 00:12:43.740 local slang, might just be a funny anecdote if 00:12:43.740 --> 00:12:46.200 you're translating a novel. But what happens 00:12:46.200 --> 00:12:49.580 when that exact same statistical blind spot occurs 00:12:49.580 --> 00:12:52.700 in a hospital emergency room? That is exactly 00:12:52.700 --> 00:12:55.539 where the academic puzzle becomes a severe high 00:12:55.539 --> 00:12:58.919 -stakes liability. Using tools like Google Translate 00:12:58.919 --> 00:13:01.259 in a medical setting is increasingly common because 00:13:01.259 --> 00:13:03.059 it helps doctors communicate with patients in 00:13:03.059 --> 00:13:05.299 day -to -day activities when a human translator 00:13:05.299 --> 00:13:07.500 isn't available. Makes sense on a practical level. 00:13:07.679 --> 00:13:10.500 It does, but researchers are aggressively cautioning 00:13:10.500 --> 00:13:12.240 against relying on it for anything critical. 00:13:12.559 --> 00:13:16.279 Let's play out a scenario. A patient uses a regional 00:13:16.279 --> 00:13:18.799 colloquialism to describe a sharp pain in their 00:13:18.799 --> 00:13:21.840 chest. The machine translation app, lacking the 00:13:21.840 --> 00:13:24.700 cultural context of that specific slang, translates 00:13:24.700 --> 00:13:27.200 it into a mild word like heartburn. Which happens 00:13:27.200 --> 00:13:30.039 all the time. The doctor, reading the app, skips 00:13:30.039 --> 00:13:33.000 the cardiac workup and hands the patient antacids. 00:13:33.559 --> 00:13:36.679 A pure mathematical error just resulted in a 00:13:36.679 --> 00:13:39.100 misdiagnosis. That's why the medical community 00:13:39.100 --> 00:13:41.620 stresses that machine translated medical texts 00:13:41.620 --> 00:13:44.919 must be reviewed by a human. And the legal field 00:13:44.919 --> 00:13:48.279 is facing similar crises. Oh, I bet. Legal language 00:13:48.279 --> 00:13:51.460 is so specific. Incredibly precise. Yeah. It 00:13:51.460 --> 00:13:54.500 uses normal words in very atypical ways. If a 00:13:54.500 --> 00:13:57.220 lawyer uses a free online translation tool to 00:13:57.220 --> 00:13:59.659 decipher a client's foreign contract, they aren't 00:13:59.659 --> 00:14:01.980 just risking a mistranslation. They might actually 00:14:01.980 --> 00:14:04.740 be violating client confidentiality. Wait, how? 00:14:04.860 --> 00:14:07.740 Because they are exposing private sensitive information 00:14:07.740 --> 00:14:10.220 to the remote servers of the software providers. 00:14:10.360 --> 00:14:12.460 Oh, of course. You're literally sending the private 00:14:12.460 --> 00:14:15.159 document to Google or whoever. Exactly. And the 00:14:15.159 --> 00:14:17.950 sources mention actual court debate. over police 00:14:17.950 --> 00:14:20.909 searches, too. How so? If a police officer obtains 00:14:20.909 --> 00:14:23.730 consent to search a vehicle using a machine translation 00:14:23.730 --> 00:14:26.889 app on a smartphone, is that legally valid? Did 00:14:26.889 --> 00:14:29.029 the suspect actually understand the legal parameters 00:14:29.029 --> 00:14:31.750 of what they were consenting to? Or did the machine 00:14:31.750 --> 00:14:36.509 hallucinate a softer phrasing? Wow. It is a massive 00:14:36.509 --> 00:14:39.539 legal gray area. Naturally, the military and 00:14:39.539 --> 00:14:41.639 surveillance sectors are heavily invested in 00:14:41.639 --> 00:14:43.840 navigating these exact high -stakes environments. 00:14:44.240 --> 00:14:46.360 Absolutely. Following the 9 -11 attacks, agencies 00:14:46.360 --> 00:14:48.799 like DARPA -funded programs like TISE and the 00:14:48.799 --> 00:14:51.179 Babylon Translator are specifically focusing 00:14:51.179 --> 00:14:55.019 on two -way mobile translations for Arabic, Pashto, 00:14:55.179 --> 00:14:57.980 and Dari to facilitate rapid communication in 00:14:57.980 --> 00:15:00.259 the field. If we connect this to the bigger picture, 00:15:00.440 --> 00:15:02.519 the central theme really seems to be the environment. 00:15:02.669 --> 00:15:05.990 In fluid, high -risk human environments, like 00:15:05.990 --> 00:15:08.490 a hospital, a police stop, or a military checkpoint 00:15:08.490 --> 00:15:11.049 machine translation, is a dangerous liability. 00:15:11.129 --> 00:15:14.009 Right. But in highly controlled, specific environments, 00:15:14.389 --> 00:15:17.049 it is an absolute miracle tool. Oh, the triumphs 00:15:17.049 --> 00:15:19.029 of the technology in controlled settings are 00:15:19.029 --> 00:15:21.889 staggering. Neural networks recently successfully 00:15:21.889 --> 00:15:24.409 translated ancient Akkadian and its dialects 00:15:24.409 --> 00:15:26.970 Babylonian and Assyrian. That's just amazing 00:15:26.970 --> 00:15:29.769 to me. There are hundreds of thousands of clay 00:15:29.769 --> 00:15:32.539 tablets from ancient Mesopotamia sitting in museums, 00:15:32.919 --> 00:15:34.600 completely untranslated because there simply 00:15:34.600 --> 00:15:36.980 aren't enough human experts. Machine translation 00:15:36.980 --> 00:15:39.639 is mass processing these texts, literally unlocking 00:15:39.639 --> 00:15:42.899 ancient human history. Or look at augmented reality 00:15:42.899 --> 00:15:45.659 for travelers. The Google Translate camera feature 00:15:45.659 --> 00:15:48.159 we mentioned at the start, when it works, is 00:15:48.159 --> 00:15:51.059 brilliant. It's like magic. It overlays the translated 00:15:51.059 --> 00:15:53.059 text right onto your environment, preserving 00:15:53.059 --> 00:15:56.549 the font and background. And in gaming, the title 00:15:56.549 --> 00:15:59.570 Lineage W gained massive popularity in Japan, 00:15:59.889 --> 00:16:02.649 specifically because its built -in machine translation 00:16:02.649 --> 00:16:05.210 features allowed players from different countries 00:16:05.210 --> 00:16:08.230 to seamlessly strategize and communicate in real 00:16:08.230 --> 00:16:11.309 time. That's so cool. And there's also incredible 00:16:11.309 --> 00:16:14.230 complex work being done with signed languages, 00:16:14.549 --> 00:16:18.049 right? Yes. A prototype called Team was developed 00:16:18.049 --> 00:16:21.019 to translate English text into American Sign 00:16:21.019 --> 00:16:22.600 Language. And this isn't just swapping words 00:16:22.600 --> 00:16:24.620 for hand gestures. It's way more complex than 00:16:24.620 --> 00:16:26.740 that. Right. Stress, pitch, and timing are conveyed 00:16:26.740 --> 00:16:29.379 entirely differently in sign languages. The system 00:16:29.379 --> 00:16:31.899 analyzes the English grammar, accesses a sign 00:16:31.899 --> 00:16:34.120 synthesizer, and a computer -generated human 00:16:34.120 --> 00:16:36.639 appears on screen to sign the text, replicating 00:16:36.639 --> 00:16:38.820 the spatial and temporal nuances required for 00:16:38.820 --> 00:16:40.700 comprehension. And we absolutely have to talk 00:16:40.700 --> 00:16:43.340 about how Wikipedia uses this technology. Wikipedia 00:16:43.340 --> 00:16:46.539 is available in 85 languages, but there is a 00:16:46.539 --> 00:16:49.509 massive imbalance. How big of an imbalance? While 00:16:49.509 --> 00:16:52.730 the English Wikipedia has over 6 .5 million articles, 00:16:53.190 --> 00:16:55.429 the German and Swedish versions only have around 00:16:55.429 --> 00:16:59.429 2 .5 million. Wow, that's a huge gap. Yeah. Volunteer 00:16:59.429 --> 00:17:02.070 editors are heavily utilizing Wikipedia's content 00:17:02.070 --> 00:17:04.809 translation tool to quickly draft articles from 00:17:04.809 --> 00:17:07.170 English into those other languages, bridging 00:17:07.170 --> 00:17:09.750 the global knowledge gap at a speed humans could 00:17:09.750 --> 00:17:12.309 never achieve alone. But this brings us to a 00:17:12.309 --> 00:17:15.069 really difficult question. If this technology 00:17:15.069 --> 00:17:17.849 is volatile enough to alter a medical diagnosis, 00:17:18.670 --> 00:17:21.089 but powerful enough to translate hundreds of 00:17:21.089 --> 00:17:24.410 thousands of Acadian clay tablets, how do we 00:17:24.410 --> 00:17:26.049 actually grade it? Right. How do you measure 00:17:26.049 --> 00:17:28.670 success? Exactly. How do we measure the success 00:17:28.670 --> 00:17:31.109 of a translation? I know there are automated 00:17:31.109 --> 00:17:33.289 metrics to do this. The sources mentioned programs 00:17:33.289 --> 00:17:37.589 with acronyms like BLEU, NIST, and METEOR. But 00:17:37.589 --> 00:17:40.029 how does a mathematical formula actually grade 00:17:40.029 --> 00:17:43.049 a fluid language translation? Let's look at BLEU, 00:17:43.170 --> 00:17:46.349 which stands for Bilingual Evaluation Understudy. 00:17:46.470 --> 00:17:48.250 It basically takes the machine's translation 00:17:48.250 --> 00:17:50.930 and overlays it onto a professional human translation 00:17:50.930 --> 00:17:53.710 of the same text. It then mathematically counts 00:17:53.710 --> 00:17:56.869 how many overlapping word sequences or engrams 00:17:56.869 --> 00:17:59.960 they share. The more overlapping sequences, the 00:17:59.960 --> 00:18:03.539 higher the score. It is incredibly useful for 00:18:03.539 --> 00:18:05.740 rapid testing. There are other methods too, right? 00:18:05.859 --> 00:18:07.759 Like example -based machine translation, which 00:18:07.759 --> 00:18:09.869 we haven't touched on yet. Yeah, example -based 00:18:09.869 --> 00:18:12.569 MT is an alternative approach that doesn't just 00:18:12.569 --> 00:18:15.089 look at statistical probabilities, but actually 00:18:15.089 --> 00:18:18.250 searches a massive database for similar past 00:18:18.250 --> 00:18:20.829 sentences and uses those as templates. Like finding 00:18:20.829 --> 00:18:23.829 a similar puzzle piece. Exactly. And some metrics 00:18:23.829 --> 00:18:25.970 found it actually performed better specifically 00:18:25.970 --> 00:18:28.589 for English to French translations compared to 00:18:28.589 --> 00:18:31.250 purely statistical models. But regardless of 00:18:31.250 --> 00:18:33.849 the metric or the method, all the sources emphasize 00:18:33.849 --> 00:18:37.250 one core truth. Which is? Human judges are still 00:18:37.250 --> 00:18:39.569 the absolute most reliable method of evaluation. 00:18:39.970 --> 00:18:43.049 Because BLEU can count overlapping words, but 00:18:43.049 --> 00:18:45.410 it can't matter whether a sentence actually makes 00:18:45.410 --> 00:18:47.970 logical sense. Precisely. It brings us back to 00:18:47.970 --> 00:18:50.630 Claude Piram. He summarized this entire dynamic 00:18:50.630 --> 00:18:53.130 perfectly with what I'll call the 90 -10 rule. 00:18:53.690 --> 00:18:56.210 He noted that machine translation, at its absolute 00:18:56.210 --> 00:18:59.690 best, only automates the easy 90 % of a translator's 00:18:59.690 --> 00:19:01.750 job, the vocabulary hauling. Just moving the 00:19:01.750 --> 00:19:04.809 words over. Right. But that final 10%, that's 00:19:04.809 --> 00:19:07.410 the part that requires six hours of intense human 00:19:07.410 --> 00:19:10.190 research to resolve ambiguities and context. 00:19:10.730 --> 00:19:13.130 The machine lays the bricks, but the human has 00:19:13.130 --> 00:19:15.589 to build the actual architecture of comprehension. 00:19:16.430 --> 00:19:18.890 And when you remove the human completely, and 00:19:18.890 --> 00:19:21.349 just let the machine loop in on its own mathematical 00:19:21.349 --> 00:19:24.509 logic, well, the system begins to hallucinate. 00:19:24.630 --> 00:19:27.210 Oh, we have to talk about the 2017 YouTube glitch. 00:19:27.289 --> 00:19:29.529 This is legendary. Oh yeah, this is so weird. 00:19:29.720 --> 00:19:31.720 Someone figured out that if you went to Google 00:19:31.720 --> 00:19:34.099 Translate and repeatedly typed in the Japanese 00:19:34.099 --> 00:19:37.420 hiragana characters A, which is just the phonetic 00:19:37.420 --> 00:19:40.400 sounds E and GU, the system's neural network 00:19:40.400 --> 00:19:42.920 completely lost its mind trying to find a pattern. 00:19:43.319 --> 00:19:45.640 It desperately tried to map those nonsense syllables 00:19:45.640 --> 00:19:48.440 to real -world concepts, spitting out completely 00:19:48.440 --> 00:19:51.079 absurd English phrases like deep sea squeeze 00:19:51.079 --> 00:19:55.079 trees. And the most famous output, desiring egg. 00:19:55.369 --> 00:19:58.450 Just all caps. Deceering egg. People made videos 00:19:58.450 --> 00:20:00.430 reading these outputs in dramatic voices and 00:20:00.430 --> 00:20:03.109 they got millions of views. So what does this 00:20:03.109 --> 00:20:06.009 all mean? Doesn't the deceering egg glitch perfectly 00:20:06.009 --> 00:20:09.269 prove that the machine has absolutely no inner 00:20:09.269 --> 00:20:10.970 understanding of the world? Oh, without a doubt. 00:20:11.029 --> 00:20:13.210 It's just hallucinating patterns when the math 00:20:13.210 --> 00:20:15.630 breaks down? It absolutely proves it. The machine 00:20:15.630 --> 00:20:18.690 has no conceptual understanding of an egg or 00:20:18.690 --> 00:20:22.529 the deep sea or a tree. It is purely mathematical 00:20:22.529 --> 00:20:25.680 probability. failing in real time. Just totally 00:20:25.680 --> 00:20:28.099 breaking. Right. Which raises an important question. 00:20:28.519 --> 00:20:30.440 If these machines are just predicting spatial 00:20:30.440 --> 00:20:33.099 relationships on a map of vocabulary, they possess 00:20:33.099 --> 00:20:36.059 no actual creativity. And that lack of creativity 00:20:36.059 --> 00:20:38.180 has sparked a massive legal debate regarding 00:20:38.180 --> 00:20:40.480 copyright. Because according to the law, only 00:20:40.480 --> 00:20:42.880 works that demonstrate original human creativity 00:20:42.880 --> 00:20:45.460 are subject to copyright protection. Multiple 00:20:45.460 --> 00:20:47.619 legal scholars are currently arguing that machine 00:20:47.619 --> 00:20:50.480 translation results are not entitled to any copyright 00:20:50.480 --> 00:20:52.750 protection whatsoever. The original author of 00:20:52.750 --> 00:20:54.690 the text retains their copyright, of course, 00:20:55.170 --> 00:20:57.950 but the actual translated output generated by 00:20:57.950 --> 00:21:00.690 the machine. Because it is just the result of 00:21:00.690 --> 00:21:03.410 a mathematical algorithm, it is legally considered 00:21:03.410 --> 00:21:05.769 devoid of the human spark required for ownership. 00:21:06.069 --> 00:21:08.950 Which leaves us with a truly wild final thought 00:21:08.950 --> 00:21:11.539 to consider. We learned today that machine translations 00:21:11.539 --> 00:21:13.500 may not be eligible for copyright protection 00:21:13.500 --> 00:21:16.660 because they lack human creativity. Yet, as we 00:21:16.660 --> 00:21:18.500 discussed with the Swedish and German examples, 00:21:19.019 --> 00:21:21.259 these exact same machines are increasingly being 00:21:21.259 --> 00:21:24.200 used to translate massive global repositories 00:21:24.200 --> 00:21:26.859 of knowledge like Wikipedia. They're basically 00:21:26.859 --> 00:21:29.519 building the modern library of Alexandria, crossing 00:21:29.519 --> 00:21:31.920 language barriers at an unprecedented scale, 00:21:32.339 --> 00:21:34.779 using code that lacks comprehension. So as you 00:21:34.779 --> 00:21:36.960 go about your week, and maybe the next time you 00:21:36.960 --> 00:21:39.440 use an app to translate a menu on vacation, ask 00:21:39.440 --> 00:21:42.140 yourself this. If machines are translating the 00:21:42.140 --> 00:21:44.359 bulk of the world's shared information, but those 00:21:44.359 --> 00:21:47.240 machines possess absolutely no creativity and 00:21:47.240 --> 00:21:49.859 cannot legally own their words, who will truly 00:21:49.859 --> 00:21:52.039 own the cross -cultural knowledge of our future? 00:21:52.480 --> 00:21:55.460 Is our shared human history becoming an uncopyrightable, 00:21:55.720 --> 00:21:58.200 machine -generated average? Thank you for taking 00:21:58.200 --> 00:22:00.960 this deep dive with us today. Keep asking questions, 00:22:01.059 --> 00:22:03.160 keep looking past the magic, and keep exploring 00:22:03.160 --> 00:22:05.200 the visible systems shaping your world.