Exploring AI Matters
Our mission is to help the policy community understand the breadth and richness of AI and the potential for such technologies, wisely applied, to augment all sorts of human endeavors.
Some AI tools are able to assist humans in performing tasks faster, more accurately, or more efficiently. Some, however, are inaccurate and unreliable. Who or what we hold accountable for these flaws, and what incentives we do or do not create for their correction will influence AI’s hand in how we work.
In this series we will refine, sharpen, and clarify your understanding of AI.
Exploring AI Matters
Episode 24 - Lost in Translation
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Send us comments and questions.
The public has been fascinated by the experience of interacting with large language models, or LLMs, like OpenAI's ChatGPT and Google's Gemini. Inaccuracies, called hallucinations, continue to appear in LLM output. Some lawyers, among others, continue to use these outputs without verifying their accuracy. Doing so damages their work product and reputations.
In this episode of Exploring AI Matters we will be looking at current work with LLMs that plays to their strengths and involves a lower risk of inaccurate outputs. In particular we will look at the use of LLMs to translate between languages.
In this episode we talk with Archie McKenzie, the founder of a Silicon Valley startup that is offering internationalization services to software teams.
Welcome to Exploring AI Matters. This podcast series, previously known as Mind the Gap Dialogues on Artificial Intelligence, will continue to appear in the ABA series to the extent that, in addition, all of the episodes, old and new, will now appear under our new podcast name, Exploring AI Matters. Thank you.
SPEAKER_04The public has been fascinated by the experience of interacting with large language models or LLMs, like OpenAI's Chat GPT and Google's Gemini. Inaccuracies, called hallucinations, continue to appear in LLM outputs. Some lawyers, among others, continue to use these outputs without verifying their accuracy. Doing so damages their work product and reputations. Today we'll be looking at current work with LLMs that plays to their strengths and involves a lower risk of inaccurate outputs. In particular, we will be looking at the use of LLMs to translate between languages. Software teams generally operate in their native language. Once they have finished building their system, they often want to make it available in other languages to access other markets. The process of making a program that was originally written for one language, usable by people who speak other languages, is called internationalization. Historically, internationalization has been a slow and expensive process. Welcome to Mind the Gap Dialogues on Artificial Intelligence, Archie. I am Charles Palmer, a computer scientist.
SPEAKER_01And I'm Ama Adams, a national security lawyer. We are your hosts for this episode of Mind the Gap: Dialogues on Artificial Intelligence. In addition, we have two more hosts.
SPEAKER_03Hi, I'm Mark Donner, another computer scientist.
SPEAKER_05And I am Roland Trope, a national security lawyer.
SPEAKER_04Each episode will be led by two of us, with the others adding impromptu questions and comments as the spirit moves them. Today we'll be talking with Archie Mackenzie, the founder of a Silicon Valley startup that is offering internationalization services to software teams. Archie is atypical in various ways. Of Britain, Archie came to the U.S. to study classics at Princeton. He ventured into a course taught by a famous computer scientist, Brian Kernigan, whose teaching inspired Archie to switch from ancient Greek and Latin to computer science. One of Archie's student projects, a website called histories.ai, foreshadowed his current startup. Histories.ai offers the ancient Greek text of Herodotus's histories, long out of copyright, and pairs it with LLM-generated translations. After graduating from Princeton in 2024, Archie started a company called General Translation to develop and commercialize internationalization automation for software development projects. Full disclosure, one of our hosts, Mark Donner, is an investor in general translation.
SPEAKER_01But thank you for joining us very much this afternoon, Archie. As I was listening to Charles kind of walk through your background, you know, one thing that struck me is that it is not uncommon for students to kind of change or pivot their course of study when they begin sort of their higher education career. But certainly a shift from classics to computer science is quite a turn. So I think just to start us off and to introduce you to our audience, it'd be really fascinating to get a better understanding of how you made that leap and what it felt like.
SPEAKER_00My pleasure, Arma. It's great to be here. I so I don't think it was that scary at all. I think switching majors even from a humanities subject to a STEM subject like computer science is definitely not as scary as choosing to study in the US versus at home in Britain. Really, for me, it was that I arrived at Princeton ready to kind of lock in and study Latin and Greek for four years. But they had these pesky things called distribution requirements, which meant that we have to take some math courses, you know, some science courses, and at least one course in computer science. And so I thought I'd get it out the way in my first semester. And I wandered completely unawares into Brian Koenigan's course, Princeton, called Computers in Our World. And I just absolutely fell in love with the subject. I enjoyed it more than my 300-level Latin courses, and so I kind of pivoted pretty quickly to work on software.
SPEAKER_01That's definitely one of the benefits of sort of a core curriculum model here in the United States, right? Students come in, they get an opportunity to try different things because they're required to do so. And in your case, certain opened up the door and lightened you to another path that sort of brought you to where you are today. Sort of fast-forwarding, you know, 2,500 years from classics to computer science. You know, you said it wasn't scary. I do wonder if you ever sat back and sort of questioned the choice at any time as you learned more about computer science, or was this the kind of situation where you sort of never looked back?
SPEAKER_00Well, there were definitely points when I thought, like, I could be a decent computer scientist with all this work, or like a really good classicist. So yeah, I there were definitely doubts. I will say that taking math classes at Princeton is the fastest way to learn exactly how much smarter other people are than you. I certainly felt that because although I had a lot of fun doing math, I didn't really do badly at all. I'm not good at math by Princeton standards. The Princeton math department is just on a completely different level. So yeah, I had doubts. I think ultimately software has such a gravitational pull that I really couldn't do anything else once I was introduced to it. And I do really credit Brian Koenighan for that. I want to put it on record that were it not for Professor Cernahan, I would likely still be a classicist.
SPEAKER_01That's wonderful tribute, right? Like meeting someone in an experience that sort of changed the trajectory of your academic and sort of professional career.
SPEAKER_00Software is so has such kind of gravitational pull. Yeah. And you can apply it to anything. I mean, I came into Princeton with an interest in different languages. At school, I'd been an exchange student in Germany and China. I'd learned Latin and Greek. I'd also learned German and Mandarin. And so now I'm kind of applying software to my original interest in translation.
unknownYeah.
SPEAKER_01So it's just so versatile. Yeah, kind of coming full circle, almost in a different way. You know, you mentioned the term gravitational pull. And, you know, you're still sort of evolving, right? You've now moved from sort of focusing on computer science and evolving to sort of artificial intelligence. And so what kind of led you on that further journey?
SPEAKER_00Right. Well, like I said, I came into college interested in languages. And in some sense, AI is the field where computers intersect with natural language. And of course, I with the transition from kind of straight computer science to artificial intelligence, I was very lucky on the timing. I was in college between 2020 and 2024. So literally right in the middle of my college experience, ChatGPT came out. AI has these like winter and summer periods where research and excitement and funding for AI companies accelerates and stagnates. And I was very lucky to go to college during probably the hottest summer on record, which is a kind of post-Chat GPT AI boom, which is still ongoing.
SPEAKER_05What did you actually mean by the gravitational pull of software? Because in physics, one object doesn't exert a gravitational pull on the other without the other having, you know, a gravitational pull on the first object. Did software pull you? Did you pull yourself into it? What did you actually mean by that?
SPEAKER_00Right. Well, I don't think I thought that that deeply about the metaphor. What I will say about software is that it's so great because the feedback loop on what you build is so tight. And it's always really been that way. So, you know, since Brian Kernahan was writing programs in the 1960s, you know, you you write a program, you run it overnight, maybe on this huge computer in your university, and you see an error and you can maybe fix it in the morning. And that just doesn't happen in other engineering disciplines. So you can't build a bridge overnight, and then if it collapses, fix it again the next night. And as computers got faster since then and programming languages got higher level, overnight becomes minutes, and then minutes becomes seconds, and it's even better. So fast forward 30 years to the 1990s, the internet boom, I can write a program, I can fix all the bugs, I could, or at least most of the bugs, I can send it out into the world, and then the world can give me feedback because I've just published it to the internet and I can change it over the air. You know, there aren't really any other engineering fields that are like that. I can't, you know, if a car engine breaks, you can't patch that car engine over the air. You know, and then fast forward 30 years to today, you can turn this into an AI conversation. AI is giving us even tighter feedback loops, or I can write a program so much faster, put it out there, and then that program can potentially use AI to change itself and adapt to new complex information in real time. So that's very exciting to me. I think it's really only possible because of the really tight feedback loop you get on these various generations of software. And that's really what I meant by the gravitational pull of software. And I think it's one of the reasons why I am a computer scientist rather than a classicist today. Well, that's a fast I hadn't thought of the feedback loop as being tighter.
SPEAKER_05It's a fascinating explanation, but it also makes me wonder do you get the feedback from the part of the world and the level of expertise that you want? How do you know that you're getting the kind of feedback you need? Because newspaper writers, they get a f they have a feedback loop. Poets, God, they have a very slow feedback loop, if any. Do you get do you reach the audience you want? And how do you know that you're getting the feedback that you really want from the people you want?
SPEAKER_00I mean, that's a that's a very complicated question. I would I would say that in the in what I just laid out for you, I was talking about lots of different types of feedback. So I was talking about the feedback you get from simply running your program and watching it fail. That's the feedback you get from tests, the feedback you can get from users in the real world, and then now the feedback you can get from AI. And I think all of those kind of feedback mechanisms are subtly different. Combined or individually, they make your programs much better than they would be. You know, you iterate much faster on what you're doing. How do you find the right people to give you feedback? I mean, it it's difficult. It's difficult. I mean, maybe maybe the the software people who host this podcast could comment on that.
SPEAKER_05Well, I will tell you in the legal profession, if you want good feedback, you call a really good lawyer like AMA, sure.
SPEAKER_04And I'm finding my my students, because I teach at Dartmouth, or at least I'm about to finish that. When I was learning computer science, you spend a lot of time thinking about what you're submitting. You did desk checking, you thought more about writing it because you had so few, such a long turnaround, so few shots at it. Now they just say, Well, gee, I that didn't work. Oh, try this, and they just keep trying, trying, trying. And I'm I understand that they've got this tool and they're just gonna beat it to death rather than think it to death, and that's working. What I'm seeing now, though, is a lot of them are trying to use uh the AIs, and that's fine. You can do the hype coding or vibe coding or whatever it is, but now you're having to develop another another skill, which is not only not just learning how to code, which was the whole point of the class or whatever, it's now learning how to look at somebody else's code and understand if that's exactly right. Oh, and by the way, back up and precisely specify what you want. Because if you get that wrong, then it's gonna give you a shoe factory as opposed to a fintech application.
SPEAKER_03But you know, specifying software has been the hard problem since the beginning. And again, so getting what you want from a software project has never been easy. And it's not easy when when it's free AI, free is a funny word, when it's from an AI where theoretically it's very fast and it's in some sense cheap, whereas you know, doing it coming along and saying to a team of programmers here, build this thing. What is this thing? Okay.
SPEAKER_00I think you're exactly right, Mark. I will also say that I always go back to the idea that wherever you work in the stack, you want to understand at least one layer below you. So if you work on an application, you want to understand the system that that application is built on. If you work on systems, you want to understand maybe the network or the operating systems or something like that. You or if you work on the operating system, you want to understand the hardware. And so if you are a real professional vibe coder, you should probably just learn to normal code because it will make you a better vibe coder.
SPEAKER_05Archie, in that regard, sometimes when we give talks to audiences whose first language is not English, and we're being translated by AI, it gets the translation from English into, say, Japanese accurate, but it butchers the transcription of my English into the language it's going to then translate into. And you can see that on the screen as it does it. It'll get my words wrong. But what it can't alert me to, and I don't know whether you run into this in the kind of translations you do, but there are concepts in certain cultures that don't exist in another culture and that are embodied in certain words that may become very popular and lead to profound breaks in the bridge from one language and culture to another. My favorite example is when I was describing situation awareness to a Japanese audience last November, and I had to be it, it had somebody had to come up to me afterwards and said, we don't have the term situation awareness in Japanese. What did you mean by that? Because the translation into Japanese was incomprehensible for them in terms of what I was talking about.
SPEAKER_00Right. Well, I think that's an argument for translating on as high a level as possible. It's an argument that translation is not merely the process of swapping out words, but potentially looking at trying to communicate an idea on a higher level. And we think about that a lot as a company. I think our internationalization tools are really built to deal with things like UI in a way that you can't just do by swapping out strings. And so we think about that a lot. I was originally a translator, I was not a computer scientist, I kind of got into software in college. And so I think I think a lot about these kind of core concepts of translation as much as I do about the core concepts of software.
SPEAKER_05Well, and I think that's one of the things that probably makes you as good as you are at this, because you've worked in been educated in both of those worlds.
SPEAKER_04Definitely. So your company, General Translation, provides these internationalization services to software development teams. Can you help us understand why it's important to do that?
SPEAKER_00Yeah. I mean, internationalization services to software development teams. That there are a lot of keywords in there. How I would simplify that is it's important to bring the world's best products to the entire world. And we help software companies do that. So, like for most internet applications, you want them to be as accessible to as many people as possible. Imagine one you're one of the developers behind, say, Airbnb or Uber or Chat GPT. You would benefit from your product both in terms of your mission as a company that wants to increase access to transportation or you know, spread AI all over the world, but also in terms of your bottom line for reaching people in every country. And, you know, one of my core beliefs as a technologist is that accessibility is really important. You have to meet people where they are. You build products that that help as many people as possible. And if you're building a really great product, I think it should be available to the whole world. And that's what we help the people building the best products do.
SPEAKER_04Oh, that sounds great. But I guess I what did they do before you? I mean, granted, what you're offering sounds like the best thing in a while, but how did they do it in the in the old age?
SPEAKER_00So, how technical do you want this explanation to be? I mean, I mean, we could we could kind of dive in. Like previously, internationalization involved replacing all the strings in your app, that is the words, with with some sort of ID. And then you would try and map that ID, given a certain language, to the translation. So one ID and English would would lead to hello world, and the same ID plus German would lead to Hallo Bet. And what you would do is you would go through your your entire code base. So basically everywhere where words appear in your application, you would replace those words with IDs, and then you would build some system for mapping the IDs to words depending on the language. This is one of the most hated parts of software engineering. You know, nobody signs up to be an engineer to work on internationalization. You sign up to work on whatever product you were originally trying to build. It's a painful refactor, and then it's ongoing technical debt because suddenly you have to manage this large external content source with all your words in it, which is disconnected from the function in which functions in which they appear. So there are there are lots of like large US-based companies that hire multiple US-based employees just to maintain the internal systems they built for doing this. And we are taking a completely different approach. What we do is very simple. You import a component into your existing file. We call this component T. You put it at the top of your content and the bottom of your content, and it translates your UI. So no painful refactor, no maintaining large dictionary files, it just works. And because it's translating UI rather than words, we can do things like rearrange the components on your screen to match the word order in a particular language. Something that's very hard without using our system.
SPEAKER_04Yeah, just making some of the words fit. I mean, you mentioned German. I mean, a lot of the German words seem to be a little longer than they need to be, but that's just the way the German was uh written. And I teach my students why do you have your messages scattered all over your code? You know, 27 source files and text messages all over the place. And I suggest to them, just like you said, what if you want this change to translate it to French? And they it's like it did never occur to them.
SPEAKER_00Well, the beauty of our system is that you can you can use a dictionary pattern or you can use an inline pattern and we will handle both for you. So whichever way you've decided to organize your application, we can help you. And it will be much easier with us than with the kind of existing piecemeal solutions.
SPEAKER_04So it sounds like your system changes the speed and cost for the internationalization process. How would you characterize that? How does it do that?
SPEAKER_00Oh, of course. I mean, look, if you look at incredible companies like Airbnb, they spend millions, if not tens of millions, of dollars on localization every year. And ultimately the mission of my company is to bring that sort of multi-million dollar localization department to every developer and to every company, no matter the size, for as low as, say, $30 a month. Now, that's the kind of ultimate goal of our company. We're still a long way from that, but I'm very optimistic we can get there. And with the customers who have used our product, we've got amazing feedback and testimony from those customers. The the other angle on what we're doing and making it much cheaper and faster than other solutions is the number of languages. So previously, translation itself was expensive. You know, you needed to hire different translators to do the actual work of transforming language A to language B. And so even the companies that bothered to do the work of internationalization and localization could only support a small number of languages. But with AI, I think the limiting factor has shifted from actually producing the words to the developer integration and the maintenance of your internationalization process within your software. And then if you do that work, your site can support basically an unlimited number of languages. So that's tens, hundreds. You can reach kind of the long tail of people in low resource languages, even which and there, I think there are many benefits to that, both in terms of your bottom line, but also from a mission perspective.
SPEAKER_04And unlike other aspects of development, as long as you don't change the message, you're done. One translation and you're done, right?
SPEAKER_00Yes, that's correct. So we we wouldn't um we wouldn't retranslate for every user. We translate once for your whole app. And this definitely I mean, I don't think translating for every user is really sustainable or wise.
SPEAKER_03I'm curious if if the quality of the of The LLM generated translations is up to snuff in terms of what your clients demand. They're used to having, you know, professional translators do that work for them. Are the results uh good or better?
SPEAKER_00Yeah, I mean, I actually think in many cases, LLMs are better than the human contractors that you get to do translations usually. I would say that the the best humans will be better at AI models for a long, long time. But unfortunately, the you know, there there isn't an unlimited supply of the best human translators and they're very, very expensive. And so if you look practically at the people who pay, you know, the translators who are paid to do software internationalization, the average quality is is much, much lower. I would say an interesting conversation I had the other day was with somebody who managed the localization process at WhatsApp in the early days. And when they went into Germany, they they had contractors do their translations, and these contractors translated the word crop as in crop and image, as crop as in what grows in a field. So they took the English word crop, it has two meanings, and they chose the wrong meaning to translate. And so for years their German users were just confused. Like, why are they talking about plants? What is this button? Like, what does it have to do with farming? And only when they hired like dedicated German employees could they figure out what was what was wrong. And so humans make lots of mistakes. I think AI makes fewer mistakes, especially if it has the context to go with it. And it really is a context problem. But I'm very optimistic about LLM translation. I think because they have this ability to understand context, you can easily imagine them being just as good as you know your average human translator at a much cheaper cost and with much more benefit to the end user who's ultimately going to benefit from having more software in their language.
SPEAKER_05Archie, let me pick up on that. I'm not a software code writer, programmer, or anything like that. And when you talk about translations, there in Europe, for example, they distinguish between people who are translators and people who are simultaneous interpreters, which is a much harder skill because it has to be done on the fly. But in either case, you deal, you have at least two problems that keep reoccurring. You mentioned one where you just pick the wrong word that doesn't translate. There are imageries and ironies that don't translate. To what extent is that a function of our the language that we humans speak? And you don't run into that in the software code world. So that therefore machine translation doesn't have to do as much as a human would in translating or interpreting human speech from one language into another.
SPEAKER_00You're completely right, Roland. I mean, I think that the the businesses we work with, they care about quality, of course. We work with the best product teams in the world, but they ultimately have a product and that product has a function. It's very different from a literary translation, where which AI is much worse at, and where you really, really need to think about the resonances of certain words, how certain words sound when read out, things like that. And we we have definitely focused on business translation. It is a much less difficult problem than either translating long literary works or simultaneous interpretation.
SPEAKER_05Let me follow further on that. In human translations, especially of literature, words change their meanings in a few decades. In a T. S. Eliot poem written during the Second World War, the word liberation had a rich resonance having to do with the Second World War. Then the word liberation, at least in the US and probably some other countries, came to have much more to do with women's rights. Now we see the term being somewhat mutilated for its use in tariff invocation. So that a translator would be much more cautious now about a word like liberation. But what I'm getting at is that over time don't some of these translations, at least in literature, tend to break down. And that's why there's always a market for new translations, especially of the ancient classics. Do you worry that that may happen in the kinds of things you have to translate to achieve internationalization?
SPEAKER_00I think for us in particular, it's not really a problem. Not least because the products we translate are usually very new and they will often change over 20 years, but also because, again, we have this kind of function that we want to express. We translate products rather than literary works. I think you do hit on a very good, a good point about LLMs in general. So these large language models, they're trained on data from the internet. Most of that data is from, say, pre-2023. I think the knowledge cutoff for a model like GPT-4 is October 2023. Of course, GPT-4 wouldn't have that same connotation for the word liberation that you mentioned because the data it was trained on hasn't kind of factored that in. And so you could easily imagine, you can almost imagine large language models as some sort of like crystallized model of language, which can't update in the same way that our human brain model of language can. And so you're going to be constantly needing to kind of retrain these large language models and provide them additional context and information. Because yes, the meaning of words shifts over time. But you know, I think that there'll always be a need for translation. I hope that um, yeah, I just want to get as many things in as many languages as possible.
SPEAKER_05Listening to that, what I would worry about is if a new technology comes out with its own vocabulary like quantum computing, when it finally washes into the public discourse more, is it going to take certain words that you have already defined and suddenly rapidly shift them? And from what you're saying, it's so new what you're doing, you're not at that risk point yet.
SPEAKER_00I think that's right. We're not really, for us, it's not, it hasn't been a problem.
SPEAKER_04But different industries are going to have different lingo, shall we say? So if you're doing something for, I don't know, Instacart, it's going to be one collection of important words as opposed to do so doing something for you know Verizon. I I guess the fact that you're translating makes that easier rather than having to actually know what the words mean, you just use the model to produce them, or how do you do that?
SPEAKER_00Well, that's completely right. And actually, if you look at the translation industry, some of the most lucrative parts of it are say legal and medical translation, where you need to get those keywords exactly right. In fact, there's a that it's very high stakes to get those keywords exactly right. We put a lot of work in to make sure that we correctly capture the meaning and keywords of whatever our users are doing in the AI translation. Of course, we let our users provide their own translations as well. So if it really is super important to them, they can kind of plug in their own translators, either human or AI. But we but for example, we maintain a kind of internal glossary of key terms in their project that we use AI to maintain and update. So we always know the key terms when we're translating, so we can ensure consistency throughout the application. Another thing is we allow users to provide context to whatever word or phrase or bit of UI they need translated in natural language. So for example, you can take a button that says enhance and you or let's say actually a button that says crop. You can take a button that says crop and say this is crop referring to cropping an image rather than a crop that grows in a field. And AI will take that into account. And you've just described it in natural language. There's no extra API you have to learn there. And so we we put a lot of work into consistency, we put a lot of work into keywords. I think doing that infrastructure work is something that really sets us apart as a company, and it's something that LLM translation is will increasingly need as the models get better.
SPEAKER_01So, in terms of the infrastructure work that you're doing to get this process as accurate as possible, are you using only one of the LLMs or does your system utilize multiples?
SPEAKER_00We use multiple. And we do pretty extensive testing to make sure they work on the particular problems that we're looking at and in the particular languages we want to serve. At the moment, we use anthropic and open AI models, but we're also actively looking at Gemini. We'd love to use closed source ones as well, but generally they're not good enough at low resource languages to justify it.
SPEAKER_01Sort of driven by general translation considerations, is the choice of LLM driven by your company, general translations considerations, or is it a situation where your customers have contracts or preferences that make you leverage or incorporate particular services?
SPEAKER_00Yeah. So we really see our expertise as figuring out which models to use for which job and then how to get them to do that job. And we want to we want to choose the best models for our customers. And the landscape of models changes a lot. I think you see lots of AI companies like Cursor, for example, following this model, where the a better model comes out and they just auto-switch for their customers. With that said, we would build whatever our customers need. So for a higher-end customer, you know, let's say we're even, let's say OpenAI comes to us and they say we want to internationalize ChatGPT. We would adapt our system so that we only used OpenAI models for translation to do that, if it was something that our customers required. But I think knowing about models and knowing how to get the best performance out of models is our expertise as a company. And so for the vast majority of customers, they should trust us to do that. I think.
SPEAKER_01I think is sort of a nice hybrid approach to kind of set you apart. But you know, that suggests to me, and a lot of what you've been saying suggests to me that you really have a behind-the-scenes view in some ways about sort of LLMs from the work that you've been doing as you've built this company and your product. So for our listeners, I think it would be really fascinating to know what you've gained or sort of what insights you've gained or lessons you've learned on how, you know, the functionality or performance of LLMs could be improved. And has that insight led you to make any changes or adapt based on kind of your real world experiences or insight with these multiple LLMs?
SPEAKER_00Right. Yes, definitely. I mean, in some senses, there are there are many AI companies that are built on these tiny little advantages they get from just being very familiar with model performance in a particular area. And we definitely see ourselves as one of those companies. Just in with regards to like insights into the models, you can kind of tell, I'm just giving an example. You can kind of tell which trade-offs people have made when they train the models. So, for example, the mini models from OpenAI are, I think, in my opinion, too small in the terms of the number of parameters for the amount of compute cycles that have been stuffed into them. And that probably increases performance in some areas, but it means that they can't really think outside of the box on a novel task because they're overtrained. For us in particular, it means that they struggle to say rearrange elements of UI on a page because they just can't conceive of the fact that you might actually want to do that rather than just translating in a normal way. Ultimately, for closed source models like those, there's only so much we can do. I mean, we we bring this up to the labs occasionally and we say, oh, you know, we would really prefer you to to train more on this output format or to to not do so much on this particular problem. So we sometimes give notes to the model providers that we work with, but generally with closed source models, there's not much you can do.
SPEAKER_04Fascinating. So you mentioned training. Who does the training of these things? Are these the companies, your customers, or do you get involved in the training?
SPEAKER_00It's open AI and anthropic. But we'd be happy to yeah, we'd be happy to hand over our docs to any LLM providers that want to train on them. I mean, we we don't really help with training at all. That's completely outside of our area of specialization.
SPEAKER_05Two disruptive questions or questions on disruption. Last week, I think it was the CEO of Anthropic that predicted within about five years there will be a sudden cliff reached, and over 50% of the white-collar workforce will be suddenly recognized as no longer needed because of Gen AI and other AI tools. And, you know, just this precipitous layoff of people with nothing to do about it. Does your introduction of your tool, even though it's replacing drudgery and things that people didn't become engineers to do, is it going to cause a disruption in the workforce? Is there going to be a large layoff of engineers who have made their careers doing what you're doing?
SPEAKER_00Well, I doubt there'll be a large layoff of engineers. I think that in in well, in some senses, if our company is not disruptive, then we will have failed. Because if we're not disruptive, then we won't really have uh added any value to any of the customers we serve. I will say that I read Dario Amade's comments about vast swathes of the workforce being laid off, and I personally don't agree with them. I think, you know, just going back to like a kind of first principles perspective, I think the point of work is not to keep people busy or to distribute money from a faucet. I think the point of work is to solve problems. And until we run out of problems, I don't think we'll run out of work. So, you know, like, okay, well, maybe there aren't any internationalization engineers anymore. Maybe that's all automated. You know, do are there is there still software that needs to be written? Because internationalization engineers are usually very smart and they can work on pretty much any system's problem. I think that, yes, until we run out of problems, we're not going to run out of work. And software is so malleable that I don't see any engineers losing their jobs anytime soon.
SPEAKER_05My other disruption question has to do with who you're doing the work for. Are you ever concerned that you're enabling by the internationalization of a particular company's model? You said to make it available to the world. Well, there's parts of the world in the age bracket below a certain age that you may not want that tool to be available. Do you have that concern that you're helping get product into the wrong hands or in an age group where they shouldn't be using it, or that simply doesn't factor into what you're doing?
SPEAKER_00Well, there are a bunch of young people who speak English. I mean, in fact, I would say more young people speak English. If anything, we're we're helping maybe older people who didn't get the chance to learn English when they were young to experience these products. I mean, I think that if a product shouldn't, you know, if a product is age-gated in some way, like say a social media platform, that's really up to the governments of the world to regulate. And if it's offered in the US and that company wants to follow all applicable laws in other countries, you know, I think it's probably outside of the scope of what we do.
SPEAKER_03I'd like to go back to the vast unemployment fantasies. People have been terrorizing themselves with dreams or or nightmares of unemployment for every technological improvement that the world has ever seen. If you look back at the history of agricultural employment in the US, 150 years ago, as I recall, I may get the numbers wrong, it was 50% of the population was employed in agriculture. Today it's roughly 1%, and we produce far more uh agricultural products, food, and so on, than we did 150 years ago. So, and I don't hear anybody whining about trying to get those damn agricultural jobs back. They were pretty miserable jobs then, and people don't want them. So I don't, I don't, I'm not worried if those internationalization jobs, which as Archie have so rightly points out, are miserable. Had an I had an internationalization team on one of my products some years ago, and those guys were very, very sad people. The the reality is they'll be happy to move on to other things, and most of the time there's a shortage of software engineers, and there's lots of more creative and more valuable things for them to do. So I'm not actually worried about this. I think the the head of Anthropic called it wrong.
SPEAKER_04This is this is too interesting. Okay, so here's here's another curveball. You spend a lot of time as a professional user of these AI services, and you are in the middle of it. And if they produce things, you try them and off you go. Are they going in the right direction? I mean, you're in a position to sort of see that these things are just turning into toys or they're getting better. Where is the industry? Or do you think the industry is going in the right direction? And where do you think they're going?
SPEAKER_00Well, I you know, you could answer that question in so many different ways. Look, maybe this is controversial, but honestly, I think it probably is. Insofar as it's possible to trust any company, I trust OpenAI and Anthropic. I think they make great products, they're real professionals, and they, I mean, they're moving just so fast. People in Silicon Valley already talk about OpenAI like it's a trillion-dollar company. And so, sure, like a lot of money is getting wasted because you have to waste money on different research angles to kind of find what works. And sure, it would be better if they made more of their models open source. We're big fans of open source software at our company. A lot of what we do is open source, and we would appreciate more open source. So, you know, we have we have nitpicks, but uh ultimately I'm excited about where these labs are going. I'm excited about all the competition between them for who has the best model. I think ultimately that benefits the end user. And I think they're professional companies who are serious about winning and what they're building is really valuable. I don't think it's a toy. I think people are finding great value in it already, especially for code generation and what we do translation. And maybe there'll be many, many more use cases in the future. So that's that's maybe controversial because it's optimistic, but that's how I feel.
SPEAKER_03It's interesting that that the sort of the uh speculative fiction, science fiction community has been thinking about AI for a very long time, and there have been a variety of different themes that have emerged from that. One of those themes has been that if you are if you can do what we do, think and talk, you must have the same kind of moral framework underpinning your brain. And so you know, you have books like Heinlein's uh The The Moon is a Harsh Mistress, in which an emergent AI basically becomes an ally of these uh I love that book. I love that book fighting for their for their for their freedom. And it's a wonderful, wonderful story. I love that book too. It's one of the was one of the great science fiction stories of all time, but it it has this fantasy that these AIs will be will be benign and well-intentioned toward us. On the other hand, you get a book like The Adolescence of P1, which is almost a contemporary of uh The Moon is a harsh mistress, in which the same sort of thing happens, an emergent AI emerges, and it it is very concerned about its own survival. And so what it does is when somebody is coming along who thinks they've discovered it and is investigating, the thing takes over the air traffic control system and crashes the plane. So the interesting question about emergent AI, about you know, real artificial intelligence, not the stuff that we have today, which is you know arguably not really intelligent, whatever that really means, is that you know, it will there be a survival urge in those things or not? And you know, we have a tendency to project ourselves onto things that look like us or act like us. And if they if they talk to us, we think they're people. That's a little bizarre and dangerous.
SPEAKER_00I mean, just on that, if they talk to us, we think they're people. I think the human brain is set up to to kind of see things that produce language as innately human. You know, nobody okay, okay, so there was that engineer at Google a couple of years ago who was convinced that these language models were sentient. And in hindsight, the language model he was so afraid of was very kind of primitive compared to. ones we have today, and we don't really think the ones today are sentient in any meaningful sense, or at least the the majority of researchers would have that opinion. I think it's telling that, for example, nobody is worried that the image models are sentient. Nobody is worried that the models that produce images are sentient because it's it's maybe a less human thing to do to like approximate an image than to approximate a sentence.
SPEAKER_04Yeah, I had another question there. Recalling my early days learning German and my wife's French background. I know you if you're just translating a UI, it's usually a word or two here and there, save, file, submit, you know, stuff like that. But if you're translating messages, can you really say that there's one way to translate it? Aren't there sometimes with some languages multiple ways to go about it?
SPEAKER_00Of course. I mean, the thing about translation is that there are many ways to get it right and even more ways to get it wrong. And I mean, there are many tasks like that. I could say running a business. There are many good ways, there are many ways to get it right, and there are even more ways to get it wrong. And so as a company, we really focus on consistency throughout your project. That's one of our core principles when we translate. We think that consistency is almost as important as accuracy. Obviously, accuracy is the most important, but um making sure that, for example, we have consistent use of terminology, that the tone is consistent throughout your project, that's very important to us, and we spend a lot of time thinking about that.
SPEAKER_05Yeah, Mark has constantly reminded us of just how valuable science fiction writers have been to imagining accurately where the world may be going. I have not been a reader of so much science fiction. I've tended to read children's stories as I get older. One of which is Raul Dahl's The Great Automaticate Grammatizator. I don't know if you've read that, but it basically looks at what would happen if machines could write so well that people found it easier to let them write the new works of literature and eventually they were bribed to stop doing so. He has a very dystopian view. What that makes me think of, though, is that it seems your product is very different in output quality than OpenAI's Chat GPT program for writing prose. And there's all this pressure in the legal industry to have lawyers start using this so that they no longer have to write emails, they don't have to write summaries of meetings, they don't have to do initial drafts. And I think that's the wrong direction for generative AI to be going because of the trade-offs that will occur. One of them is that the output that we're increasingly accepting is of a much lower standard. And when you get that output standard low enough, sure, people will be willing to accept a Gen AI output as good enough. The second trade-off is that you not only reduce the quality that people get from lawyers or other service providers that rely heavily on language, but you stop giving people the opportunity to train in using it, because writing is a lifelong apprenticeship. And third, you start to devalue those mundane tasks without realizing that if a machine, for example, summarizes a meeting, it may not omit things that a lawyer would realize you don't want a record of from that meeting. It may represent them in terms that are sufficiently different that it loses the nuance that when it gets litigated will be critical. Am I making a distinction that you find inaccurate given how much confidence you have in what open AI is turning out?
SPEAKER_00Right. Well, I definitely wouldn't say open AI is equivalent to a human. I definitely wouldn't say these large language models are equivalent to a human in any way. I would say that they can be used to automate certain processes. But I think it's difficult to imagine, say, a complete AI lawyer that works in exactly the same way as a normal lawyer would work. I think you'll have to find different form factors that really play to the strengths of these models. I would say that maybe the the best metaphor is LLMs as a sort of calculator. So calculators massively enhanced mathematicians, but they didn't replace mathematicians. I mean, I'm sure they replaced some math-related work. But the the digital calculator is an incredible tool. And I think that the kind of equivalent of the digital calculator for language rather than math is the large language model.
SPEAKER_05It's interesting you would say the lawyer, there was a Department of Defense proposal a few years ago to develop an AI lawyer. And one of the reasons it was withdrawn was the fact that if generals could get an AI lawyer that would deliver a recommendation with a high level of confidence faster than the Army or Air Force or Navy JAG officers, the legal officers could do, they would tend to accept that and not wait for the human who might have more misgivings and qualifiers, but might also be right in the longer term. It was a terrifying prospect.
SPEAKER_00Well, I think that the idea of replacing a single human with a single AI is, I don't know if realistic is the best, the best way of describing this. I I would say that in design there's this there's this design pattern called skewamorphic design, which is where you take you take something that already exists and try and translate it exactly to whatever new technology you're using. For example, in the original iPhone, the Notes app had sort of yellow pages and line rules, just like a a kind of a real life piece of paper, like notebook. And skewer morphism is maybe all right for kind of having a very intuitive interface that people understand, but ultimately the way you get the best performance is you design for the new technology. You don't let yourself be limited by the designs that existed before. And so I am not I'm not very optimistic about just replacing like single human employees with single AI employees. I'm more interested in automating certain processes and then letting humans do what they're really, really good at in a way that has humans and AI complement each other.
SPEAKER_05I I'm going to invoke something that one of our earlier guests mentioned in a telephone call recently, John Blackburn, who's the former deputy chief of the Royal Australian Air Force. He believes that the biggest problem with generative AI is unintelligent use of the tool. And I think that that's what you're also pointing towards is that it's not just the tool, but whether we're using it intelligently. Right.
SPEAKER_00And it's, I mean, the fact that it's so new and you can apply it to so many different problems makes me think that there are many ways in which we're currently using it unintelligently and many intelligent ways yet to be discovered, which will provide great value to people.
SPEAKER_01Well, Archie, thank you very much for your time this afternoon. You know, I think we all really appreciated sort of the intelligent and very refreshing conversation we had about a range of different topics. And I think we can all have a special shout out to Professor Karen again for introducing you to computer science and bringing you to where you are today. So thank you again for joining us.
SPEAKER_04Yes, thank you. And I want to thank our sponsors, the Bar Association, Ben Rosenbloom, our composer and performer for our music, and the Marianne Schneider artist for parts of our website. Thanks again for coming, Archie, and thank you everyone for joining us.
SPEAKER_00Thank you so much for having me.
SPEAKER_04We thank the business law section of the American Bar Association for their generous sponsorship of the production of this podcast. We welcome questions and comments from listeners. Send email to comments at Mind of the Gap Dialogues.com. We read all comments and questions and will try to respond in the letters section of a future episode. If you're writing about a particular episode, please mention the specific episode number. And please also include pronunciation tips to help us properly say your name when we reply in a subsequent episode. See you next time on Mind the Gap Dialogues on AI.
SPEAKER_02Thank you for listening to the ABA Business Law Sections Podcast Series to the extent that the section offers a robust collection of content. To explore more about this topic or to learn about joining the section, visit ambar.org slash bizlaw. That's B-I-Z-L-A-W.