On the latest NPR Politics podcast, we look back on the life and legacy of Jimmy Carter. The former president died Sunday at the age of 100. And unlike other presidents, his accomplishments after leaving the Oval Office are arguably what most define his story. Listen to our remembrance of the 39th president on the NPR Politics podcast.
Hi, this is Mary Childs. There's the saying about artificial intelligence that right now AI is the worst it will ever be. It screws up a lot now, but it is only going to get better. So how to even answer questions like will AI replace jobs or change jobs? We talked about this on the show last year, and especially in the Planet Money newsletter, which our beloved Greg Rosalski writes. Subscribe if you haven't already. It's very good.
So today we are sharing something that Greg uncovered for the newsletter this summer. He figured out a clever way to investigate AI progress for jobs. Look at something AI is already good at. Translation. Greg interviewed an AI innovator who also is the head of the language learning app Duolingo.
We regularly publish Greg's newsletter chats as bonus episodes for our Planet Money Plus subscribers. So if that is you, you already got a chance to hear this. Bonus content is just one perk of signing up for Planet Money Plus. You also get every episode of Planet Money without sponsor messages and you get exclusive Planet Money merch at the NPR shop. So while you listen today, consider supporting us by signing up. Just go to plus.npr.org. Okay, here's Greg.
Recently, I've been trying to figure out how to measure if AI is over-hyped right now, or maybe it's appropriately hyped, I don't know. And one of the people I talked to you about this is an entrepreneur and AI innovator. My name is Luis Fanan, and I am the CEO and co-founder of Duolingo.
Duolingo, it's the popular language learning app. Luis has a pretty fascinating background. He was born and raised in Guatemala, and he came to the United States for college to study mathematics and computer science. In his early 20s, Luis co-created CAPTCHA. You're probably familiar with it when a website forces you to prove you're human and ask you to do things like identify stoplights or bicycles or whatever in a grid of photographs.
i didn't know this until i spoke to him but capture is actually an acronym completely automated public turning test to tell computers and humans apart it's a mouthful. It's basically a test to distinguish whether you're human or computer. Louisa's work on capture is one reason why he was awarded a MacArthur genius grant back in two thousand six at the time he was only twenty eight how do you how do you stay humble after that.
Uh, I don't know. I don't know if I am humbled, but well, that's honest. You know, I don't know about staying humble, but I mean, I mean, there's a lot left to do. Back when he first co-created CAPTCHA, Louis says the test was pretty simple. It was, you know, easier to trip up machines. At first, all it took was identifying squiggly letters and numbers.
Nowadays, computers are about as good as humans at reading the story of characters, which is why captures have moved onto basically really blurry images of bicycles or whatever. Interestingly, Louis says, these are images that are often taken by self-driving cars. The computer guiding these cars sometimes has a hard time deciphering what these images are, which is why capture crowd sources help from humans. This data is, in other words, kind of slowly making AI smarter.
So yeah, Luis is a pretty interesting guy to talk to you about AI, not only because of his work on CAPTCHA, but also because earlier this year, there were a series of headlines suggesting that his company, Duolingo, laid off human workers and replaced them with AI. Okay, so here is today's bonus episode, my conversation with Luis Fanon.
I want to start with jobs in language translation. So I've seen a lot of these popular listicles in the media that highlight the 10 most likely jobs to be like killed by AI. And basically on all of these lists are the jobs of translators, interpreters. And yet the US Bureau of Labor Statistics projects that over the next decade, the number of American translators, interpreters will actually grow by 4%, which is about equal to the average growth of all occupations.
And I also took a look at job sites and there's still tons of companies hiring translators and interpreters right now. So what's up here? If AI is so good, why are translating jobs actually still growing? Well, so, okay, a couple of things. This is not new. Translation by computers between, for example, English and Spanish or English and German, like between the large languages has been
really good, you know, starting with Google Translate for about 10 years. And it keeps getting better and better and better. And, you know, what I don't think is going to happen is one day, suddenly, you know, translation was complete. And then the next day is perfect and then I'll translate this to get it fired. This is a gradual process, which started out years and years ago. And today, a lot of translation is already done by computers. I mean, that's already happening. Now, at the same time, I think there's also more and more demand for translation
Because translations are a lot cheaper, a lot faster because computers are there, there's a lot more demand. So what you're seeing today, at least for translation in particular, is this combo, this hybrid between humans and computers. I mean, maybe the computer takes the first pass and then a human kind of fixes it up a little bit, et cetera. And over time is probably the case that computers will be as good as humans at language translation, maybe even better because they're faster.
We're not quite there yet. There's still the case that computers make some mistakes. And I think companies are hiring humans or governments are hiring humans when you want somebody to actually have a real opinion because it may be a life or death situation. For example, I don't think you want to fully rely on a computer if you're a translator for the army or something like that. And you're talking to an enemy combatant. I don't think you want to fully do that yet.
Or companies who just want to make sure that no mistakes are made. But even translators that are hired, most of them use computer translation as a first step. And I think you're going to see something like that. So does Duolingo employ translators and interpreters? And if so, how many or about how many?
Interpreters, we don't because we don't do like, you know, what, what you would see like on the UN, like, you know, doing real time interpretation like that. But translators, we do not very many. We have been reducing that number for a while. And at this point, it's, it's, I know, dozens. It's not a lot. What do they do at your company? Generally, they're, they're going over the things that computers have done.
Probably double check. Most of the stuff we do with computers. Yeah, double checking, making sure. Now, it depends on what for, by the way, I should say. There's a difference in importance in text. For example, if it's our content as in like our learning content, there's so much of that. Thousands and thousands and thousands of kind of sentences and words and paradox. That is mostly done by computers and we probably spot check it. But if it's things like the user interface of Duolingo, where we say like, you know, the buttons as quick.
and we have to translate that is all done with humans and we spend a lot of effort on that. But that's because each one of those is highly valuable. We cannot have a mistake in the quit button or in the button that says purchase now or whatever. We just cannot have a mistake. And not only it's not just about mistakes, we want to make sure the voice is consistent throughout the app, et cetera. So in other words, because AI can make mistakes, because it's not perfect. You need humans. It's worth the extra cost.
Yeah, for example, in our app, we have made the decision that in Spanish, we use the informal second person pronoun. So when we refer to somebody in the app, we refer to the user informally, when they say you, we say the un-informal you. It's like a stud versus two. Two is informal. We don't use a stud, we use two. The people that do our user interface in Spanish know that and know that very well. If you were to just have a computer to do it,
It may be consistent in different screens. It may be something. So we just want to make sure that it's all consistent, that it has the same voice.
been playful, voice, etc. And so for that, we still employ humans. So earlier this year, there were a number of articles, actually quite a few, including in the Washington Post. They were published and they made a really big deal out of the fact that Duolingo laid off, I guess, some contractors who specialized in translation and replaced them with AI. So first off, what can you tell us about that episode? Did you lay off translators and replace them with AI?
Yeah, there's a lot of exaggeration that happened there. I mean, there was, there was all these articles that said we made the major layoffs. This is not true. We did not, we did not lay out 10% of your workforce or something. Yeah. No, this is, this is just not true. We did not, first of all, no, no full-time employees were affected here. We've had a contractor force of a few hundred contractors. And what happened at the end of last year is we did not renew the contract for some of them because we looked at, you know, the work that was going to be required over time. And we just didn't need as many of them.
It is true that one of the reasons we didn't need as many of them is because some of the stuff that they were doing, we could now automate. But you've got to understand this type of worker. I mean, these were not all of them, but a lot of them were people who were working a couple of hours a week from a very remote location. That type of work is probably a lot more susceptible.
or being substituted with AI, then somebody who is kind of in the office every day, you know, doing more creative stuff or anything like that. This was, you know, a lot of that was just kind of wrote stuff. So I got to ask, do you have any plans to lay off more contractors or employees and replace them with AI? And why or why not? No plans. No plans for that. What is true is that we're, you know, we're going to, as a company, and I think most tech companies, we're leaning into AI quite a bit.
And we're going to continue doing that. And you know what we're seeing, I'll tell you, I'll give you a really good example of what we're seeing with AI. About five years ago, somebody pitched a feature idea to me, which was basically in the app, some of the lessons, you were basically going to see a little animated, kind of like a little two-minute cartoon where you got to listen in the language that you were learning. It was a really cool feature. And then I asked, how long is the data for this going to take to create? Because we need to make the data, like basically all the episodes,
They said five years. And I said, no, you're crazy. We're not going to make that feature. I don't want to spend five years on this. This is ridiculous. About a year ago, the same person came back to me and basically said, hey, we can do that. But the data now takes like three months to make.
with AI. There's some human involvement, but it takes like three months to make with AI. And then I'm like, sure, just do it. And you know what's beautiful about it? Not only can you do it in three months, as opposed to five years, even if you mess up and you do it in three months and at the end, you don't love the outcome, you can redo it. And it'll only take three months. So you can do it and mess up and do it again and still do it way faster than five years. That's the type of stuff that you're seeing where things just, by the way, we're still going to need the humans.
We're just going to be able to do way more and way cheaper the things that we just couldn't do before. Supercharging productivity. Yeah, the productivity is like 10x. And that's more the direction we're going to go where we're going to lean into AI quite a bit. But we're just going to be doing things that before we're just prohibitive because they were going to require a thousand people working for years, whereas now 20 people working for a few months can do what was required before. So I think on our end, that's probably more what you're going to see.
You care about what's happening in the world. Let's state of the world from NPR keep you informed. Each day we transport you to a different point on the globe and introduce you to the people living world events. We don't just tell you world news, we take you there and you can make this journey while you're doing the dishes or driving your car. State of the world podcast from NPR. Vital international stories every day. You care about what's happening in the world.
Let's state of the world from NPR keep you informed. Each day we transport you to a different point on the globe and introduce you to the people living world events. We don't just tell you world news, we take you there and you can make this journey while you're doing the dishes or driving your car. State of the world podcast from NPR. Vital international stories every day. When did Duolingo start using AI and how have you been using it?
or since the beginning, I mean, when we launched Duolingo, we knew that we wanted to have a computer that was going to teach you. So not, we didn't want to have humans teach you. We wanted to have computer. So from the beginning, we were using AI now. Of course, AI has gotten a completely different tone and definition in the last couple of years because of large language models, but AI has been around for, you know, the concept of AI has been around for whatever 60 years, maybe longer than that. So when did you adopt large language models?
With large language models, almost two years ago, less than a little less than two years ago, we started playing with GPT4. We got early access to it. And so we started playing with it. We officially launched the first features related to it last year. Is this an official partnership with OpenAI or is it like? We have an official partnership with OpenAI. It allows you to chat in real time on the app or? Yeah, conversation. That's the biggest thing.
You know, learning a language requires a number of different aspects. You've got to learn vocabulary, you've got to learn how to read, you've got to learn a bunch of stuff. We historically have been really good at teaching you all aspects of language except for one, which is conversation. And in particular, the kind of multi-turn, you know, think on your feet kind of conversation. We just couldn't do that with technology up until large language models, but now we can.
And so that's what we're doing where we're basically really adding a lot of features to teach you how to converse a lot better. So a couple weeks ago when OpenAI recently released GPT-40, they really highlighted to the world how great it is at translation. It was really kind of front row and center of how they were selling this new
Yeah. Version of it. And, you know, they showed a smartphone using GPT-40, translating a conversation between people in real time. Are you concerned at all that it will reduce the demand for Duolingo? The fact that, like, the bees are getting so good? We are not seeing a decline in people wanting to learn a language with Duolingo, even though this has been true. If you look at our users, we have two big chunks of users. One is actually Americans or British people who are doing so as a hobby. It's just a hobby.
And when you ask them, like, why are you learning on Duolingo? They're like, well, you know, I used to play Candy Crush. And now at least I'm getting pretty good at Spanish.
Those people are going to continue doing so because it's hobby again. I mean, you know, people are doing chess as a hobby and low computers have been better than humans at chess for the last 20 years. So that's kind of one big group of people. And the other big group of people is people learning English for a number of reasons. Not only is it for, you know, possible education opportunities or employment opportunities, but it's even, it's just even to get around. I mean, if you go to many other countries, this is true in many Latin American countries, a lot of stuff is just in English.
and like ads are in English and stuff like that. And people just want to say, well, you know, yeah, music is in its depth. That's so English. It's a unique language in that respect, because it is the lingua franca, et cetera. People actually honest to that want to learn English. And that's the other big chunk of our users. So, you know, historically, we just have not been concerned about this. So last question. Do you have worries about AI? Are you optimistic? Like, how would you characterize your feelings, your belief about AI?
I'm mostly optimistic, I'm mostly optimistic, but I am anxious. And I think the anxiety comes from the fact that it's just really hard to know what will happen and how fast it will happen. And in addition to that, I can see a place where maybe the end thing will be really good. Like, you know, imagine we have AI can do everything, etc, whatever you want. And the end thing is, well, we just, you know, we just have a life of leisure. Maybe that's the end thing. It's great.
I worry that the transition may not be so good, especially if it's very fast, because it may be the case that a lot of people could lose their job. It said, I just don't know. My sense is that we probably want this probably a slower transition rather than a faster transition. And it seems to be happening that way with translators, interpreters. No, you said like, right, this has been one of the frontiers of AI research and development. And it's taken a long time. The good news is that even with things that are clearly obvious societies take a while to change,
Even with things that I like, we just all agree that this is the right thing to do. It just takes years. And so I think in general, if you give humanity, you know, a couple of decades of notice.
I think we'll probably figure it out, but I do worry about a short transition because, you know, if a lot of people lose their job at the same time, particularly, you know, this is kind of this is all this research in a large groups of unemployed young males leads to basically war. You kind of don't want that. You know, as long as that doesn't happen, I think we'll be okay. All right. Well, thank you so much for taking the time to speak with me. Thank you, Greg. Thank you.
If you liked this, you can hear more of Greg's newsletter interviews and other bonus episodes like the Planet Money Movie Club by signing up for Planet Money Plus. In these times of uncertain advertising revenue, subscriptions are the steadiest, best way to support our work and help keep us going. Again, just go to plus.npr.org for details. We are very grateful. I'm Mary Childs. This is NPR.
NPR brings you the updates you need on the day's biggest headlines. The Senate narrowly passed the debt ceiling bill that will prevent the country from defaulting on its loans. Stories from across the world. Knowing how to forage and to live with the land is integral to a nice culture. And down your block. From CPR news, this is Colorado Matters. And you can find all of that and more in your pocket. Download the NPR app today.
