Abstract

Adal traces who decides what machines can read through the history of character encoding, from national standards such as ASCII (American Standard Code for Information Interchange) and four large East Asian standards to Unicode and its 1991 merger with ISO/IEC 10646. He describes the shift from a logic of scarcity to a logic of expansion in the number of sinographs, and uses the Frederica Wang problem and China's attempt to encode 897 rare sinographs in its own Guobiao standard to argue that states have returned as primary actors alongside big tech. A character that Unicode has not encoded cannot be searched or indexed and cannot reliably pass through an optical character recognition (OCR) pipeline or an AI model's tokenizer, so he proposes a digital Zomia of marks, such as marginalia and graffiti, that remain outside the standard. In questions he covers fonts for simplified and traditional characters, emoji, code breaking, and the 172 scripts in Unicode 18.

Chapters

Transcript

AI-generated transcript. It was produced from the recording by speech recognition and cleaned up with AI. It has not been edited by the speaker and may contain errors. The captions on YouTube are AI-generated in the same way. Each timestamp links to that moment in the video.

[0:07] Now, let me start by talking about states and encoding standards before Unicode. So let's start by encoding standards. What are encoding standards? So a barcode requires a standard. Otherwise, barcodes will not work, right? Same thing for Wi-Fi, same thing for electric plugs, although arguably that has only succeeded that much. Unicode is the universal encoding standard that almost all devices, pretty much all devices, digital devices use nowadays. It's on your phone, it's on your computer, it's everywhere. And it's what makes characters interoperable. What does that mean? It means that when you type an A with your left pinky, if you're on the QWERTY keyboard, the A is transferred into a certain sequence of zeros and ones, which is readable

[1:08] to the person who is receiving your email, for example, on another computer. It makes sure that when you type an A, the other person doesn't read it as a B. So who creates these standards? it's often been national governments who have created these standards. The first one is ASCII, the American Standard Code for Information Exchange, which was created in, I believe, 1963 by ANSI, which is the American National Standards Institute. The size of the code space is pretty limited. It fits on one slide. I mean, that's all. That's the whole thing. It's 128 characters. It's 7-bit. What does 7-bit mean? It means that it's 2 to the power of 7, which is equal to 128. And it has the Latin alphabet used in English. And then a few control characters, things like delete and

[2:10] stuff like that. now you can if you're sitting at a computer you can try this yourself this is how you read it so that before I say that this is how you read it so you start with here it's like it's row four the a and here it's column one so that means that it's you U standing for Unicode, because everything is Unicode nowadays, 0041. And you can try it yourself by, if you're on Windows, it's very easy. You just type 0041 and then Alt-X. If you're on a Mac, you actually have to enable it before doing this. But then the letter A will appear. Now, ASCII only covered the letters in the English Latin alphabet. There was also French and German, right?

[3:11] All the accents, accents aigües, accents circumflexes, all the CCD, all that. There's German, right? All the umlauts, right? And back then there was also the set, right? And then there were standards for non-European languages. There was a standard that was called ISO IEC 8859. Now, what does that mean? ISO, as I mentioned earlier, is the international standard organization that standardizes everything from baby seats to photographic film. Anyone who remembers the days of photographic film, it was ISO, right? That was the sensitivity of the emulsion. And so this is one of their standards. And they do it in conjunction with IEC, which is the International Electrotechnical Commission.

[4:13] And it covered several scripts in addition to the Latin script used in the English language. things like accents on top of Latin characters, Cyrillic, Arabic, Greek, Hebrew. This is not an exhaustive list. It's a bit of an example. And at the same time, there were other standards that were not organized by ISO, IEC, like ISCII. ISCII is the Indian script code for information exchange. It sounds very similar to ASCII, but it's for South Asia. And that encoded, I think, nine Indian languages, South Asian languages that you can see here. Well, you can see some of them here, but there are others also, which are mentioned on the bottom here. And then in East Asia, national governments also created their own encoding

[5:14] standards. There was the Guobiao in China. There was the Japan Industrial Standard, JIS, in Japan. There was the Korean Standard in Korea and the Chinese National Standard in Taiwan. There were also others, but those were known as the big four. Now, the big difference with standards in Asia, as you can see here, are these numbers. Right? They're enormous. The character set is huge. We're talking thousands and thousands of characters in each of these four character sets. That's because they all employ sinographs. And as you know, sinographs, Chinese characters, have a huge character set. now the standards that we've seen so far have either been set up by national states

[6:17] right like ASCII and by the US ISCII by India and the various East Asian standards by China the Republic of China Korea Japan or by ISO but ISO is essentially an international body that's based in Geneva and whose members are national representatives. So in that sense, I think it's justifiable to consider this first period really about being states determining and coding standards. This is before Unicode. In the 1980s, a few people in Silicon Valley were grappling with all of these different standards. in what was known as multilingual computing. The most famous of these individuals was somebody known as Joe Becker. And Joe Becker, this is the back of a photograph that I found in

[7:21] the Silicon collection at Stanford University. Joe Becker, the father of Unicode, right? And this is, I think, justifiable in a lot of ways. He was an employee at Xerox. He was working on Xerox Star on multilingual computing. And in 1988, Joe Becker, whom you can see here, wrote a very famous pamphlet, which is known as Unicode 88. And this essentially set the agenda for Unicode. It was both, it's a mix of a manifesto and a very detailed technical paper. So what is Unicode? Now, Unicode is a universal standard. And these are just some examples. In no way is it exhaustive. But Unicode, by being an international standard, it essentially aspires, and

[8:22] actually does, because it aspires to including every single standard in the whole world. And so it includes ASCII that we saw earlier, it includes ISCII, it includes ISO 8859, it includes all of the East Asian standards, and many, many more for Thai, Vietnamese, lots and lots of others. Now, the bottom part is actually crucial, because what it's doing here is it's encoding each of these four national standards separately. Now this is, I'm sorry, I'm sorry.

[9:24] So Unicode does not encode them separately. So what's the difference between this slide and this slide? In the first slide, these four national standards for East Asian characters have been made into one. In this slide, which is the ISO standard, they are encoded separately. So let me back up to say, so Unicode introduced its own universal standard. But when it introduced its universal standard, there was already a standard that was being pushed out, that was in the works. And that standard was by ISO, right? We saw ISO earlier, right? representatives. And so that standard was known as ISO IEC 10646. And like the Unicode standard, it was meant to include all of the characters in the world. So you essentially had two

[10:27] competing universal standards. The first one is actually ISO. That had been in the works for a while. It make sense. ISO had like several regional standards. There were several that were outside of its purview, but it wanted to create a single standard. And then it came out with a proposal for 106.46, which was still in the process of being reviewed. And then Unicode comes up and has a competing standard. The big difference between them, and I jumped the gun a bit earlier, is that in the ISO version, each of these four East Asian standards is encoded separately. So you end up with, I don't know, like tens of thousands of characters. In Unicode, there's a device by which the sinographs are encoded together, which means that a common

[11:29] sinograph that is used in China, Korea, Japan, Vietnam, Taiwan, any number of places, is actually encoded as a single code point. Now, it looks a bit different in Chinese, in Japanese, in Korean, but that difference in appearance is supported by the font. So you have a different Chinese font, a different Japanese font, different Korean font, but they're all representing the same basic character. And so this is a distinction between what Unicode calls characters and glyphs. The character is the abstract concept of a character. The glyph is its physical appearance as it appears on a screen or a page or whatever. okay um and so now why was this why was this um necessary this was necessary

[12:32] um because the unicode proposal was 16-bit what does 16-bit mean it means that it's two to the 16th power whereas the ISO proposal was 32-bit, which is 2 to the 32nd power. I have this in a later slide. The difference is that it's between over 4 million possibilities in the case of ISO IEC and 65,000 code points, right, possibilities in the case of Unicode. So in order to fit characters into 65,000 characters, Unicode had to differentiate characters from glyphs in East Asia and bring together all of these sinographs that are similar across East Asia instead

[13:35] of encoding them separately. Okay. So there was a big, what do you want to call it? I do, I do. So there was a big, what shall we say? There was a lot of animosity between the Unicode camps and the ISO camps. And it lasted for a year or two. The article goes into great depth into this. It's very entertaining because all of these individuals are very colorful and have very, very colorful, you know, express what they what they you know their positions this he was the Japanese representative at ISO his name is Masami Hasegawa this here is is is is Jürgen Bettels Jürgen Bettels of DEC then here you have Michel Suignard of Microsoft.

[14:41] And here you have Ed Hart. So all these people were on different camps. These were definitely, you know, and so they were eventually brought together. The problem was that if this went on, you would have had two universal standards. So the problem is that if you have two universal standards, what happens? Well, software companies still have to create two versions of their software, one for one standard and the other for the other standards. It's better than creating 20 or 30 versions of the software, which is what happened when you had a bunch of national standards. But it's still not as good as creating one version of their software, which is what you have when you have a single universal standard. The other problem was that if you had communication between digital devices in different standards, you could very well have garbled characters in a lot of cases. Because the code points would not be the same. And or else even worse, you would

[15:41] have the wrong character show up. Right. So so having two standards was was just was just bad for everyone, as it often is. And then there was a moment in 1991 when everything was kind of brought together. This guy who's here, you can barely see him at heart is the one who did it. And he when he wrote, he reminisced about it. He wrote it was like trying to get the Hatfields and the McCoys together. it, I can tell you. I asked to my manager about it. We realized that we couldn't even begin to talk about a merger until there was a trust created between the two parties. We had to build the trust, even while appearing to make progress on the technical issues. In the end, it was a success. There was a real incentive to have just one code. Really, it was just a common sense thing. Once people put aside their egos and realized the other side also had reasonable people, And so you have the unification.

[16:43] And the unification is also called the merger. I mean, there were certain concessions made to ISO, but the big, big, big concession that was made to Unicode is that all sinographs were brought together. And one sinograph across all of East Asia had a single code point. now the thing with sinographs is not only is the character set very large but it's actually an open character set which means that there are old characters that are perpetually being discovered. There are new characters that can be invented. It's, you know, changing a sinograph, you can invent one yourself by, you know, changing one of these like little parts to achieve a slightly meaning, slightly different, subtle meaning. And so this character set,

[17:49] there's always the potential to have more characters. And there are always, There's always more characters that people want to use. Now, having said that, I'd like to, I've argued that there are really two main actors here, big tech and the big state. Big tech is, I think, the Unicode consortium. This is, in some ways, a simplification because, yes, software companies did have a huge financial interest in having a universal standard. and Unicode really suited them. The 16-bit size of Unicode worked really well on digital devices back then. It made a lot of sense for them. But at the same time, the participants often acted at their own behest. I mean, they often believed that a universal standard was a good thing to have. They wanted to solve the problem of mojibake, of garbled characters. And later on, we'll see that there are other

[18:49] ideals are added to this and which are present in the Unicode consortium today. So the question is, how do they handle the difference between simplified characters in mainland China and non-simplified characters in Taiwan? in the same way that they handle the difference between Japanese, Korean, Taiwanese, Chinese, Vietnamese, Chinese characters, the character is the same. So the concept of the character is the same. The font is handling. The font is used to render the simplified character. Thank you. That's a very good point. Now, let's jump a bit from 1991 to 1996.

[19:49] By then, Unicode is expanding. And you know how I said that Unicode could only contain 65,000 characters? Well, that actually changes. Because in 1996, with version 2 of Unicode, which by then is Unicode 2.0-ISO-IEC 10646, because there's been a merger, there are 17 planes of 16 -bit that are created. And so if you take 2 to the 16th power and you multiply it by 17, you get just over 1.1 million possibilities. So the character space increases dramatically. and this, I think, encourages Unicode in 1989 to move from what I've called a logic of scarcity to a logic of expansion.

[20:51] A lot of the early founders of Unicode, like Joe Becker, really, and others also, saw the proliferation, the large number of sinographs as a problem. I mean, Joe Becker talks about these sinographs, I mean, he uses words like they're garbage, they're junk, they're suspicious, they're useless, they're nano-useful characters. And I mean, he's arguing, of course, because he wants everything to fit within 65,000 characters. But in so doing, he's arguing that a lot of these sinographs, just forget about them. We don't need them on digital devices. It's okay. And I mean, it's true that 99.9% of texts can be written with more than 99.9%. 99% of texts that are written are actually written with the initial 20,000 character sinographs that were encoded.

[21:53] So arguably, that 0.01 really matters. And so this is, again, making the point that I already made, which is that the more efficient solution is to move to a full fixed area. So he doesn't want any junk characters in there. Now, in 1999, what happens with Unicode is that this is Unicode 3.0 in 1999, is that there is a jump from 20,914 characters in the first version of Unicode, which grows slightly in the second version. And then suddenly it expands dramatically to 70,000 characters. So a lot of these quote-unquote nano useful characters are being included in the Unicode character set. And it continues growing to this day. And now we have 101,997 sinographs in Unicode.

[22:58] And that number will continue to grow. I mean, there's an entire institutional apparatus that's been created in order to bring these characters into Unicode. In fact, today, if I wasn't at this conference, I would really seriously consider being in Tokyo for the meeting of the Ideographic Research Group, the IRG, which is the institution that does this, that accepts proposals for new characters, reviews them, experts review them, and then it's suggested to the Unicode Technical Committee, the UTC, which usually meets in Silicon Valley. Yes. care much about the history angle? Are they invested in knowing where they've been to understand where they're going? Some of them do. I would even say that many of them do. None of my research would be possible if Ken Whistler, who is one of the founders of Unicode, had not donated the Unicode collection to Stanford libraries.

[24:03] Because a lot of... There's obviously significant of Unicode. But people like Ken Whistler, he's also a linguist, and he has a PhD in linguistics, so he cares deeply about the research aspect of this. Okay, so now in 19... I'd like to give an example that illustrates what happens with this and what these nano useful characters can mean and which I think really gets to my core point. And to do that, I'd like to briefly go back to 1990 when a number of the Unicode founders were discussing something that they called the Frederica -Wang problem. Who is Frederica Wang?

[25:07] Well, it's the idea that when Frederica Wang was born, her parents, and they're saying that the parents are, you know, semi -literate. And so instead of writing the standard character for Wang, right, they use maybe like a slightly different character. But actually, in a lot of cases, parents often go back to ancient texts to like poems that they love, which have very rare characters. So they could very well have used one of these other characters, which are also pronounced Wang, but which were not encoded in the first version of Unicode. So you could not actually type them on a digital device. And so Frederica is a completely fictional character. But the idea is that her name is and therefore it cannot be used on a digital device. Her name just does not exist on digital devices.

[26:10] She has to use another character. She has to use this name instead of using her real name. And so they're discussing this and they're asking, you know, should her characters be encoded? And there are various perspectives. On the one hand, there is Joe Becker, who doesn't like large numbers of sinographs, who says putting her unique name into Unicode won't make vendors draw it in their fonts. It's only going to sync Unicode. And on the other hand, there's Rick McGowan, who says, Little Frederica Wang has a name and she will not vanish merely because there is no place in the code set for her special character handicap. And so he's offering other solutions. I won't get into the specifics of this. You're welcome to ask me about them later on. It's a different, it's what both of these people agree is that,

[27:10] is that Frederica is that the encoding is essentially, I mean, she is at the center, right? And making her name interoperable across devices is something that she wants. Both of them, the starting point is that she really wants to do that. He's like, it's impossible. He's like, let's try to make it possible. But the idea is that she as an individual wants to do that. Now, let's jump to 2022, 2023, when there's a really interesting episode that challenges this. this idea that it's the individual who wants to encode their name and to make it available on digital devices. In late 2022, the Standardization Administration of China, which is the national body that is in charge of script encoding in China, did what Ken Lunde, Ken Lunde is currently the chair of the Ideographic Research

[28:12] Group that's meeting in Tokyo right now. He's currently, I'm sure, in the mountaineering, not now, because it's nighttime. But he was earlier in a meeting. And he called this an incredibly bad thing. Today, Ken Lunde is an employee at Apple. He has deep knowledge of sinographs. And so what is this incredibly bad thing? Well, this is the incredibly bad thing. list of 897 rare sinographs that China sought to unilaterally encode on the unallocated plane number 10. So China of the unicode. So China essentially, instead of going through the process, which usually takes three to four years, you can try to fast track it in a couple of years. But anyway, it takes years. Instead of going through the process of proposing

[29:13] characters to the Ideographic Research Group, which then forwards them to the which goes to two more committees before they're encoded. And anyway, this is a very large number of characters. Usually you don't come with like, you know, almost 900 characters. You come with like a smaller number. What China did, the Chinese state decided to do was simply to encode them within Guobiao, which is its own standard, the Chinese standard. which is a disastrous thing for Unicode. It's disastrous because it's essentially breaking the standard, right? It means that software companies once again have to create a version of Office or Windows or whatnot, one for China and one for the rest of the world. And it's even worse because it could encourage other states like Japan or Taiwan to do the same thing. And then suddenly you're, you know, potentially back to where you started from, which is no universal standard or a very loose universal standard.

[30:16] And so to avert this, Ken Lunde and other experts did, you know, they vetted the repertoire of proposed sinographs, which is a very laborious process. In record time, they fast tracked it in ways that, you know, had never been done before. They approved it through the UTC. They pushed it through ISO. They pushed it to the Unicode Technical Committee. And after several rounds of revisions, it was eventually adjusted to 622 characters. And then it was published on September 12, 2023. So it was published less than a year after it was proposed, which is unheard of. So this did avert disaster. But let me explain China's side of the story. Why did China do this? Why did the Chinese state not just follow the regular procedure? Now, the reason, and there's considerable, I think there's good evidence to support

[31:19] this, is that proper names are rather non-standard proper names, like those of Frederica Wang, are a headache for modern states. every individual has to be identifiable in an electronic database. Otherwise, I mean, how can you, you know, match a fingerprint or, you know, a photo from a face recognition engine to a name, right? It's critical to have names in order to identify people in population databases. And so when a character cannot be written on a digital device, that's a huge problem. So there were a small number, maybe a few million Chinese, whose names could not be written. Maybe it was a few hundred thousand. And so in order to do that, the Chinese state kind of went ahead and did this unilateral encoding. The Chinese state is not the only one that...

[32:22] And so there are two ways that states usually deal with this question. One is by encoding new characters, right? that we already saw. The other way is by codifying the characters that the names of babies can use. So in general, to write Chinese, you can use 6,500 characters. That's what newspapers, magazines can do. There's an additional 1,605 characters that can be used for proper names only. But this was only published in 2013. So anyone who was born before 2013 was not subject to this, which is why in 2022, the Chinese state had this considerable population of people whose names could not be written on a digital device. and they wanted to fast track it.

[33:23] They didn't have to, but they decided to. I guess it was an emergency. And so they did this. And by the way, other states also have these kinds of, Japan has a very similar system set up, but it's much older. It is from 1946 and 1951. So what we have here is that first in the age before Unicode, it was national states that had encoding, that created encoding standards, right? ASCII, ISCII, the Japanese, Chinese, Korean, Taiwanese, you know, Indian standards. You know, ISO was about to publish a universal standard, which would have included all of these different standards. up. And Unicode is born in Silicon Valley, and it represents the interests of large technologies companies primarily.

[34:25] And I've called the rise of Unicode, a very rapid rise, the ascendance of big tech. So I see it as the big state and then the big tech. And then in 2022, 2023, the Chinese state really is putting its foot down and forcing Unicode to extremely rapidly encode the characters that it wants to encode. Otherwise, it's just going to break the system. And this, in some ways, is, I think, the re-imposition, the return of the big state as the primary actor. Now, the question I'd like to ask is, you know, who is Unicode for? Is it for large software companies that save money by only publishing one version of their software? Is it for individuals to write their names, even for somebody like Frederica Wang, whose name has rare characters? Is it for minority scripts that may have a very small community of users, but

[35:31] whose users want to use them on their phones and computers? Is it for scholars studying ancient scripts? Is it for states who are creating population databases to render their populations legible in order to watch over them? The story of Frederica Wang has an iconic and ironic coda. but its encoding was the very instrument by which she fell under the yoke of the surveillance state. Now, this points to, I think, a larger issue. We're acutely aware, I think, of languages that historians need to learn in order to go to the archive, right? You're not going to go into an archive where the documents are all in Spanish, unless you either know Spanish or else you have a friend or an assistant who can come with you. And then you go there with some translation

[36:31] tools. But we need to be, I think, also aware of the languages that machines can and cannot read. There is an infinite supply of marks that are unreadable in the world. It's not just sinographs. It's really any mark. to borrow, I think a term that was popularized by James Scott, we could, I think, talk, talk about a digital Zomia, this whole world, right? So Zomia for, for James Scott and others also is this area that is outside of the purview of the state, right? It's kind of like hill country where the state doesn't, doesn't reach much. You could think of a digital Zomia. A digital Zomia is this whole kind of like realm where you have characters, marks, you have symbols, you have squiggles that the state cannot, that, that, I'm sorry, that, that, that machines cannot read, right?

[37:32] And if machines cannot read them, you can't transmit them on a digital device. You can't put them in a database. They, they're just, they're unreadable by the machine. And so here, this points, I think, to a question that will become increasingly important in the age of AI, as long as a mark is not encoded, it remains largely outside of the purview of AI. Once it is encoded, it can be OCR'd and become, or recognized in some other way, and become one of the characters that AI can read and employ. The frontier of encoding is where the politics of legibility are being played out. If legibility is a form of power, the question is, agency who has the power to make something machine readable and for what purpose? Yes, all Latin characters are encoded, of course, and we can even grant that almost

[38:32] all sinographs used by almost all people are already encoded. The remaining ones are very, very rare. The same goes for the 170 other scripts currently in Unicode. But what about emoji? There's an infinite possibility in the numbers of emoji. What about marks invented by people like, which somebody is actually in the process of trying to encode them, but by others, right? What about the countless marks and symbols on the sources that we use? What about the marks and squiggles that, you know, children write on their homework? What about, you know, are they going to become legible for machines? Do we want to make them legible for machines? And so the question is, what do we want machines to read and to not read? And it's this boundary between machine legibility and machine illegibility, which I think is a highly political space, but which I think we haven't really thought much about thus far. And this is kind of a depiction of

[39:33] this. I just give the example of states because it's prominent, because they are prominent in the encoding process. The state is actually the one that is kind of pushing out the line between illegibility and legibility. And these are two versions of, you know, say Unicode, right? But there are many more than two, right? And then there are others. Who are the others? It could be scholars, right? It could be just individuals like Frederica Wang. It could be, you know, it could be technology companies. It could be anyone, right? Who is really, you know, deciding what the line between legibility and illegibility is. So thank you very much. And I think we have about just under maybe 15 minutes for questions. So I guess I'm confused by something.

[40:34] I understand that if something isn't in the code, It might not be legible through OCR, but AI is able to create and interpret images. It's able to create and interpret sounds. None of those are coded by Unicode. why is it not possible to transmit those digitally i mean it is possible they may not map to a particular set of characters but i can upload a picture to you and it can give me an interpretation of what's happening in that picture yeah maybe correct me and be incorrect yeah

[41:36] yeah yeah no that's that's that's a really good point yeah no thank you um so i mean yeah i'm supposed to remind you that's right that's right thank you thanks thanks for reminding me yeah um yeah so the question is um you know what's the difference between really text and graphics or sound right and since AI is definitely able to recognize images and to recognize sound in case in fact you can you can have a fairly precise what is that called like a fairly precise signature to a sound clip You can also do a reverse image search, like TinEye, for example, where you can upload an image and then Google will look for a similar image online. So to some extent, yes.

[42:37] But the big difference is that with text, first of all, text takes infinitely less space. You know, if you were going to have AI recognize every character on a standard page, right, that the size of the file would be like, I don't know, enormous compared to a text, right? Second of all, there is a certain certainty to text, right? I mean, A is an A, right? As long as you have a universal standard and you have the appropriate font, you know, and a few other things like rendering, you're fine. With sound clips, there's quite a bit of, but there's always like, it's much more fuzzy. And images even more so. I mean, image recognition, there's not that

[43:42] one-to-one correspondence that you have with text. But that's a good point. I mean, that is something that is, yeah, I mean, how could AI change Unicode, right? That's a good question. Yes. I have a curiosity question more than about application. When did the emojis get added into this whole system? The emojis came about, oh my God. I should know this. Oh, it's okay. It came about, I think it's about, not that long. I mean, like maybe like a dozen years ago or so. And then it grew like at an incredible speed. Now there's like a specialized emoji person at the Unicode Consortium. And there's a lot of politics around emojis, which are really interesting in terms of race, gender, and a lot of disability, right?

[44:42] I mean, they have a policy. of like no firearms, but it's, yeah, it's interesting. I'm interested when they do the new release and I think a little one else, it's like, okay, there's 20 emoji are coming out. Here's what we can now add. It's like, it's giving a lot of sense and sometimes I'm like, who's that? Yes, yes, yes, yes, yes. Yeah. Does anyone online have any questions? No. I asked. OK. OK. Thank you. Yeah. If you think about the implication of this, I really appreciate your talk. One, the reason why I attended this particular talk, and I understand this is a pretty hard work feedback from the audience. Just bear with me for a second. Because I know nothing about any of this. Zero. So this is really important. I've never, I don't think I've been able to talk to coding before.

[45:45] This is a very, like, it kind of reminds me, however, of code breaking and how different sides of World War II communicate with their forces throughout the war and how important North Atlantic, Pacific, things like that. Given the proliferation of AI in the corporate finance and other security apparatus around the world, they're going to leverage AI to analyze and they need a huge amount of information so that they can make decisions faster. And you mentioned how, unless I'm originally corrected, unless something is encoded somehow, I might not be able to read it correctly. Absolutely. Is it possible that we're going to see something similar to what we saw in World

[46:47] War II, where we're going to come up with these nonsense words and nonsense figures to communicate across, I guess, a communication space in order to fool these AI, these LLMs? in order to get to this, just as a house here. You see something like that happening, it's just like a venting of things. And then changing what those characteristics look like five minutes later just to keep secrets out of the list. Yeah, thank you. So the question is about code breaking. And during World War II, there's a whole industry around code breaking. And how is this affected by this? I think you're absolutely right. I mean, during World War II, you were scrambling letters, That is extremely easy for AI to solve. I mean, AI will go through zillions of permutations extremely fast and solve that in the blink of an eye. So I think what remains

[47:47] more of a secret code is this part. All the squiggles, lines, whatnot that are not encoded. Right. This is not to say that, you know, if you put an image recognition engine, if you trained it and like, you know, put it on a set of marks, that image recognition engine wouldn't be able to figure out what the code was. But then this is a big but AI is functioning at enormous scales. It's not looking at one text. It's looking at, you know, billions of characters and millions of texts. Right. And it is the power of AI is that it's extracting from a huge amount of data, more data than any human or any team of human could ever fathom or read. And that's the power of AI. And if something is not readable to AI, then AI doesn't have access to it. Right. It remains an image.

[48:47] And if it's an image, yeah, I mean, if you if you actually know that one thing. Yeah, but that's not how a lot of, I think, that's not how I think surveillance technologies are in a lot of cases being deployed nowadays. They're being deployed at huge scales, right? We're not looking at like one text. We're looking at like millions and millions of texts that have been produced by millions and millions of people, right? And it's going through them at a scale that is unimaginable for any human or even team of humans to do. So, yeah, so I think that reinforces. That's, I think, where the struggle is. It's this boundary between the readable and the unreadable. When something is encoded, it becomes readable to machines. Yeah. So getting back to the community session yesterday. I was thinking about that, too. The big bird. the big bird and the Latin script, it also a very specific type of Latin script.

[49:49] Someone encoded that. And that's why AI could read it very well. Yes. Yes. It's still using Latin letters, which would be, you know, would spit out in different words. It's just that it looks different from our eyes. Somebody had to train it to impress that. When a Q looks like this, it's a Q. And when a U looks like this, it's a Q. And his abbreviation would be God. yeah I was thinking the exact same thing like machine is learning how to apply these in different ways so exactly recognize the verb as a verb yeah exactly so he's just coming to it after someone had already encoded it right because he had said in the beginning it wasn't doing a good job of it and then a few months later yeah yeah so yeah so the question is about Professor Kalenza's keynote address yesterday where he showed texts in medieval Latin scripts, right, that had been recognized by AI. So, yeah, so what is encoded is the Latin script that you could see on a

[50:55] screen. Now, the AI has been trained to, as you said, right, when there's a bird, they know that it means this, when there's this. And AI is very good at handwriting recognition. So with Japanese handwriting recognition, for example, it takes a lot of training to be able to read early modern handwritten Japanese manuscripts and AI is becoming really good at it. And very, I mean, it's, it's at the point where, you know, in not so long, it's going to be better than really, you know, the almost all humans at reading it just because it's able to, to remember like a huge number of permutations of like, and to map those permutations, those visual images, right, onto an encoded character. Right? And so it looks at a handwritten text and it spits out something that's printed, that's in, you know, printed characters.

[51:56] One last question? Yeah. One of the reasons I came to this talk Very interesting, the idea that Chinese characters are open characters, it can be modified. But I was also wondering about more languages that have more limited groups of speakers. And the issue of World War II and code came up. Is Navajo encoded in this system? Are Bantu languages that maybe don't have a writing system or have a writing system that isn't Indigenous but was invented by missionaries were there to figure out how to print the Bible in every language in the world.

[53:01] Are they encoded? Is pretty much every language encoded? Another one that occurred to me was Georgian. Georgian is, yeah. Georgian is a Navajo I suspect, yes, but I'm not sure. You could look it up very easily. I think they they use the Latin script. Versus like the Cherokee syllabary, which was really special. It's not Latin script. That might be a counter mistake. Yeah. The question is about Navajo and other rarely used scripts. Have they been encoded? And what are the implications of that? And I think, I mean, now you have 172 scripts in version 18 of Unicode, the latest version, that have been encoded. that's like about 160,000 characters in all and it's increasing all the time I mean every year there's a new version of Unicode and every version of Unicode has more encoded scripts and more characters so it's completely and

[54:06] there's the script encoding initiative that's based at Berkeley, California which is especially kind of dedicated to bringing forward new proposals for encoding new scripts So it's increasing. And it's, yeah, I mean, they're not secret anymore. In fact, they're not just human readable, but they're machine readable, which is the big thing. All right, thank you so much, everyone.