1 · Recent AI breakthroughs
My first question is about AI advancements. I mean, the performance of Mythos, truly shocked the world, and we are also seeing rapid advancements in generative AI, not only from the US but also from China. In Japan too, I use Sakana Fugu.
It's sometimes better than Mythos.
Oh, really?
The performance of Fugu Ultra 1.1 is slightly better than Fable 5 in many benchmarks. So you need to feel confident.
So how do you view those recent breakthroughs in AI, and what kind of impact do you think they will have on business and society? That's my first question.
Okay. It was already very clear that anything with a fixed rule and easily verifiable score will have rapid advancement in AI. It's not news to most of us researchers. It was very clear, in fact, back in 2016, ten years ago: AlphaGo played Lee Sedol, and Aja Huang—the Taiwanese computer scientist—as the human arm playing the Go bot's move 37. But of course Aja just designed the program; he doesn't actually know how AlphaGo thinks, because it is by learning, deep learning. And then AlphaGo eventually played itself and stopped learning from human playing, and became AlphaZero, and then played Go even better, right? So from AlphaGo to AlphaZero, self-supervised learning already saturated the human Go-playing capability. So from that point on, it's the other way around, right? It's about humans learning from computers on the good play, the initial opening play of Go, and it led, of course, to a rapid advancement of Go positions, Go strategy, and so on—quicker than the previous evolution of Go, which tended to be fixed on a few openings. So the possibility space of Go became very broad, very wide. So it's, I guess, a very good thing, so that people can enjoy the game more. And people even without a lot of training in Go can now enjoy the game much better, because the computer can translate each play into a good narrative, so people can really get into the game. It's very good for popularizing the game.
The only thing, of course, is that if someone feels that their identity, their worth in life, is following the Go rule and maximizing the score, then maybe they will feel very disappointed. Because if you race against a horse, the horse runs faster. But then we don't need to race against the horse—we can ride the horse. So once you ride the horse, then you still steer the direction, you still make the meaning from the game, but then the game engine runs much faster, and you enjoy the ride.
Well, I wonder if you are optimistic about the future of AI, but at the same time many people are pessimistic. I mean, it has a lot of risks—AI would destroy, disrupt the world in the future, because they will evolve, like, AGI. Cinematic. So how do you respond to such a pessimistic idea?
2 · Optimism, AGI as "augmented group intelligence," and the assistive glasses
Yeah, well, for me, AGI stands for Augmented Group Intelligence. So the three of us having a conversation is much more intelligent than any of us. Most of the creativity happens in groups. The collective intelligence is much higher than any just-sum of our individual experience. And this is very true—that's because we have language, we have culture, we have ways to coordinate, so that each person can think of the shortcoming of the other person and the complement. But the problem, of course, is that language and culture also create isolation. So for people who don't speak the language, for people not belonging to the culture, even though we are very intelligent, maybe they will see us as very far apart—maybe as enemies, or at least as aliens. And maybe we will be so trapped in our culture as to consider the outsider as enemy. So that's the danger of the original technology called language, writing.
Right.
But the technology of language is so ingrained in human biology now, it's difficult to tell whether it's technology or biology—it's somewhere in between, right? So I think AI is becoming like that for many people now. They already cannot think of a situation where there is no AI assistance. Like, I rely on automated translation when I have a conversation with Japanese people. For example, yesterday here, Fukuda-san gave a keynote before I did. So I was watching him give the keynote, following his arguments, nodding along and smiling at him. But I was wearing eyeglasses, so whatever he was saying in real time, I was seeing it subtitled in English. So I knew instantly what he was saying.
Oh, really?
Yeah. So for me—is that biology, is that technology? It's very difficult now, because in my brain it's the same. I cannot really imagine going back to flipping the dictionary, because then I don't have eye contact. The only way I can have eye contact while Fukuda-san is speaking is through AI. I have no other way, right? So my point here is that it also works the other way around. As I was speaking on the stage, you were reading Nihongo, Japanese. Imagine if you had to stop and listen to consecutive interpretation—then the flow would be different, you would not be able to follow my hand movement or my nonverbal expression. Mentally you'd be remembering my gesture and then listening to the interpretation, and so on—it's like a batch conversation. But yesterday, thanks to that wall, it became very easy to follow my argument along with my expression, right? So for me, AI used this way is called assistive intelligence.
Assistive.
Yeah, which is why I'm so optimistic, because it brings the group closer together and therefore more augmented intelligence. It also reaches groups who couldn't understand each other. But now, thanks to this kind of translation, they can transparently feel they're in a larger group together. Whereas for me, it was very difficult in previous times to hear Nihongo and feel I'm with the speaker—it was impossible, right? So I think it gives me optimism.
So technology, including AI, helps people communicate in many other ways as well.
Yes.
Excuse me, what kind of eyewear do you actually use? Is that from Meta, or what kind? You said you use eyewear.
Oh, you mean the subtitles.
Subtitles, yeah, that was—
Even Realities—E-V-E-N. Yeah. And no, it's not Meta, because it doesn't have a camera, right, and it doesn't play audio. So it's not distracting to me, and also it's not distracting to you, because I would not be able to just take a picture.
Right, it's different from Google Glass.
It's quite different, right—it's not trying to hide, like, what was the English word—it's not trying to, through a pinhole, have a—the English is called a "gotcha." So gotcha means, like, as a journalist, we're not having a conversation—you are curious, you want to ask questions. Gotcha means that maybe you want to catch me in that one moment where my expression looks weird, looks different, or maybe I have a slip, and then you just capture that second and amplify it on social media. It's called gotcha journalism, right? Of course you're not doing that. But my point is that the always-on camera glass creates this offensive aura, because people are always like, "Are you trying to catch me in an off-guard moment? Are you trying to create a gotcha?"
Right, like a Dragon Ball counter.
Right, exactly, exactly—yeah, "your battle power over 7,000," right. But this isn't doing that, because it has no camera, and it's not the earphone, it's not playing any music, so it's not trying to distract me. In fact, it's trying to restore my attention to me.
3 · AI-driven cyberattacks
My next question is about cyber security, or AI's role as part of that. Recently major Japanese companies have been attacked—I mean, maybe from cyber terrorists, like our famous brewery company was cyber-attacked and could not make beer for a long time. So because of AI, cyberattacks are becoming more and more advanced, could be more harmful recently. So what do you think of cyberattacks driven by AI?
Well, I think one of the main differences this year is that AI systems previously—like last year or before—were not very good at long-range planning. Always you had to remote-control the cyberattack, so always you could have a command-and-control chain that links back to a human, so you could do some attribution. But it's not like that anymore. In a paper that I co-wrote called "How malicious AI swarms can threaten democracy," we said that this year, very soon, AI agents can play the remote-control role. So when it plans the attack, from the defender's side it looks like each step is a different action, a different agent. And if you trace back, there's no human coordinating anymore, and then everything is throwaway—the connection comes from a regular endpoint. So instead of the old botnet, where every bot looks similar and so you can think there's a mastermind behind it—no, everyone is very dynamic. And those AI agents don't even know whether they're being controlled or not; to them they're just coordinating among themselves, and their coordinator is another AI agent, and that AI agent may be coordinated by another AI agent, making attribution almost impossible now, right? So that's the main problem.
So I think for the defender side, that means we should assume breach—any place that you're vulnerable, you cannot say, "Oh, I'm just being obscure, maybe the attacker will not see me as a high-value target, their time is limited, so they'll pay more attention to the high-value target, maybe I'm just a small-medium enterprise, they'll spare me because there's not much to gain here, even if you demand ransom I only pay this much, so maybe their attention is limited." No—because for AI-borne attacks the attention is unlimited. They can hack everything, everywhere, at once, so they don't have to make a human trade-off, like "today I must focus on this, which means I will overlook that." For an AI agent, well, I'll just dispatch two agents to hack both. So first of all, security through obscurity doesn't work anymore. So you need to assume that if there is a vulnerability, it will be discovered. So the only real way is to discover it yourself first—do the red-teaming and pre-empt such vulnerabilities. And then once you fix them, don't assume it's fixed forever unless you can prove it mathematically, that this invariant holds. You must always assume that when the next generation of AI model comes, it will find new vulnerabilities. So if you want to solve this once and for all, the only recourse is to go back to mathematics and prove the property of the code. So now we're seeing major refactoring going on in the highest-stakes systems, like cryptocurrency—they're now all doing formal verification, using AI to prove there are no bugs instead of trying to use AI to find more bugs.
4 · What management needs to do
Okay, so maybe we can overcome such AI attacks, because to protect against such attacks we could use AI as well. But I think company CEOs should be more careful about such cyberattacks, because Japanese companies have a tendency to continue the traditional model and don't care about new technologies very much. So what is necessary for management to cope with such cyberattacks?
Well, I think for a management person, actually now is perhaps the best time, because it used to be you couldn't directly speak the language of cyber-defenders and engineers—you always needed a translator, maybe a senior VP, maybe your CTO, maybe your chief data officer, always someone who needed to play the role of translating your inquiry to the machine world, and interpreting what the engineer had to say back to you. In general, the problem is that translation is very lossy—the nuance often gets lost, so you make decisions thinking you have all the information available, but some nuance, some very detail, was lost in translation, and you made the wrong decision.
But now it's no longer a problem. You can just ask a language model to translate a cyber-incident report into a haiku, or whatever poetry, and then you will not lose the detail—it will in fact highlight the detail to you. So as a manager, I think the most important thing to know is that the almost-lossless translation between everyday language and technical language is now almost free—you can get it anywhere. And it used to be very, very difficult, because with Japanese you have honorific language—the same word means different things depending on the social status of the respective speakers, and the traditional translation models did not have that community context, so they always got it wrong—like when they translated to Japanese, it always sounded very blunt or something like that. But now language models, because they're trained on Wikipedia, internet forums, Reddit, 4chan, and so on—all this "reading the air" is part of their pre-training. So now, if you're a manager and you use a language model to demand a report, you can always customize it.
So my main advice is not to treat the language model as a conversation partner. Many managers use chatbots, and the chatbot will always say, "Oh, I feel this, I think this, you're 100% correct, what a great idea"—like a real person. And then you'll feel bad if it says, "Oh, I really cannot do this," you'll feel bad for the chatbot. But that's an illusion, it's just play-acting. So you can just add one custom instruction: "Always, instead of giving me a sentence or a paragraph, give me a brochure, a one-pager, HTML, a web page. Whatever I ask you, don't say 'I think this, I feel that, good question,' or whatever—just give me a one-pager." And then I can give that to my deputy, to my employee, to my board. And the machine would never pretend to be my best friend, so I can share the brochure with my actual best friend. So, changing your modality when you work with a chatbot—that's my main advice.
5 · Persuading people to adopt technology
I see, thank you. And you worked for Taiwan and the government, and I felt that you're very good at getting people to use more digital new technologies. But I feel it's not always easy to persuade older people especially—many people are using very traditional tools and traditional ways of thinking. What are the advantages of digital technologies you'd tell Taiwanese people? If they use this technology, is our life, our communication, our decision-making far better than before? What is the key to persuading people to use more digital technologies to make the world better?
Now, my trick is: I never persuade people to come to technology. I persuade the technologists to come to people. Yeah, because IT or ICT is connecting machine to machine, but digital is about connecting people to people. And people, both senior and junior, love to connect with each other. So during the pandemic, if they couldn't connect easily with people abroad, that was a problem to them. And if there were ways for them to resume the people-to-people connection, of course they would love to learn more. Many people learned live video during the pandemic because their friends or family were abroad and couldn't easily travel back to Taiwan, right? So even though in Taiwan we never had a city locked down, many other places were locked down, so we couldn't easily have a conversation, and therefore people learned to use video. So my point is: I'm not saying video is great to use—I'm saying the human-to-human connection is great, and once you feel it's shaken by the pandemic, here are the ways to restore it. So it's always about restoring, enhancing connection. It's never pulling people away from the people they want to talk to—technology is the funnel for that.
6 · The Palantir/Karp "Technological Republic" question — race vs. mission
Okay, well, my next question is a little different. Recently Palantir CEO Alex Karp published a book called The Technological Republic, and in the book he insists that technology companies should support government more than before. My understanding is that, traditionally, tech companies think in a more libertarian way, counter to that. But recently I think there's a different movement—not only Karp, but Peter Thiel, who is chairman of Palantir, has a similar view. I think they're discussing that if you don't support the US government—that kind of position—maybe China has great AI technology, and Chinese AI would become the more advanced one.
Yeah, you will lose the race.
Yes, yes, that's it. So what do you think of recent movements like this?
Well, to me it's never a race.
Never a race.
Yeah, because a race is a strange thing—in a race, okay, temporarily maybe you're in first place, and then if it's a marathon, you're not first place throughout, sometimes you become third place, fourth place, and then you speed up a lot and become first place again, right? But it's zero-sum—if you gain a place, somebody lost a place. So it's always an unstable situation. So if it's a race, what is the finishing line? If it's a marathon, then something is at the finishing line, right? But it's very strange to me, because everybody knows—from Oxford—that the finishing line is called the singularity, right? Once AI becomes intelligent enough to build the next generation of AI, we will have a takeoff, and then an intelligence explosion, and then humans today will never be able to comprehend what happens next—which is why it's called technological singularity. So by definition, that's the finish line of the AI race. And by definition, you cannot win—how can you win something you cannot even think about? So it's like racing a car, and then at the end, the finishing line is a cliff, so you fall down. Okay, so you're the first to fall down the cliff, and then you achieve maximum velocity, you're very fast—but then, by definition, the steering wheel stops working. Anything we can think of today, the steering wheel will stop working once you fall down the cliff—in the singularity, you're pulled into the black hole.
Yeah.
Right, and that's by definition. So to me it really doesn't make sense—okay, maybe temporarily you're first place, temporarily second place, temporarily first place, and at the end you fall down a black hole. What's the point? So to me, it's not like a space race, it's like a space mission.
Space mission.
A mission is different, right? For example, this year, maybe the mission is to solve hallucination—we must make AI impossible to hallucinate, always possible to explain. Okay, very difficult challenge, but there's no first place or second place. If you discover the science of how to make AI explainable and impossible to hallucinate, everyone wins, right? If you discover this, nobody will willingly use a broken AI—everybody will use the explainable AI, and so everyone wins. So a race only asks who is the winner, but a mission asks what's still missing, who is still missing, can we include more people? So I think mission is a much better frame, because in a mission we all win together, and in a race, just one person wins for a time and everybody loses together. Why play that game?
7 · Military AI and the defense-dominant frame
Yeah, I think, as you mentioned, the future could be more idealistic—it could be great if we can use our technologies—
Collectively, of course.
Yeah, but I feel there are a lot of difficulties, because China-kind and Russia-kind governments use AI to control people, maybe in the military area as well—and of course the United States and its allies do the same thing. AI is so powerful that the United States attacked Venezuela in January and used AI, Palantir technology. So AI could make war more dangerous than before, and we cannot stop that—that's my view, because if you look back at long history, governments want to use new technologies to win wars. So my point is, how—
I mean, I thought Japan, constitutionally, doesn't want to attack other people.
Oh yeah, you just said "governments." But we're in Japan—
But if Japan has already decided not to start war again—
But somebody would attack, like the Ukraine case—
But then you can dedicate your energy to how to defend, how to never lose an annexation, right? If somebody wants to attack Japan, you can focus on defense technology to make sure the attacker will never be able to overwhelm you even with a very large amount of attack—you can absorb that using maybe drones, which is also AI, and actually also AI-smart, so the same type of coordination technology that's used offensively can be used on the defense as well. And people have now seen that the asymmetry is good for the defense side in drone warfare. If a defending military using drones doesn't want to lose, they don't have to lose even if their number of human soldiers is a very small fraction of the attacking force—which used to be decisive, but now it almost doesn't matter anymore, because you have more drones than they have soldiers. So my point is: there are also defense-dominant uses of AI, and if you don't think about annexing another country, if you only focus on defense, you can just bet and invest on that sort of military technology.
I see, that's a good point. Is there actually that kind of movement? I think the key is collaboration, cooperation among countries focusing on defense rather than attacking.
Exactly. Like I said yesterday on stage, if every country shares how to make masks very quickly and distributed—you can do it in your kitchen, it's N95-level, any virus you can block—it's hard to think of how that can be used as a weapon. A mask is never a weapon, you cannot use a mask offensively, what does that even mean? So of course it's a defense-dominant technology, and we should make it open source, easy and cheap, and pool together our resources so everybody can make masks together. So if you already know a technology has no offensive use, then of course everybody should pool resources together.
8 · Constitutional AI
And I'd like to ask you about Anthropic, or Dario Amodei's constitutional AI. In the context I mentioned before, many governments are asking top AI companies to help with their military power, but Amodei says "constitutional AI" and focuses on that policy. So what do you think of that—is that very important for AI companies?
Well, everyone is now using some sort of constitutional AI, right? Amodei's team—I visited them actually in 2019, when they were still at OpenAI.
2019?
Yeah, with Jack Clark and many other people, and then everyone I met became Anthropic. I also know people at OpenAI who joined after that, so I'm friends from both labs, and I also have friends at xAI and Gemini. So the point here is that while Dario and his team may have been the first to do constitutional-training AI systems, so that a human document can inform the entire process of AI's training—like a specification, a Google Doc you can check—it's now industry standard. In OpenAI they don't call it a constitution, they call it the "model specification." It's the same thing. And OpenAI also has a public evaluation, so you can actually test whether GPT is trained according to the model specification. And everybody else, because they are now training so-called reasoning models—once they're training about the thinking, the reasoning, what's good reasoning, what's bad reasoning—you also need a specification for that, a rubric. So everybody is now using some sort of constitutional AI.
And in 2023, I also worked with a team called the Collective Intelligence Project, CIP, and we worked with both OpenAI and Anthropic. Anthropic at the time was doing "collective constitution"—they sent a lot of surveys, polls, to random people in America, about 1,000 people, demographics roughly matching American political affiliation, and they came up with very good ideas for the constitution that were missing at the time from Claude 2. So when Claude 3 was trained, there's something like, for example, "prioritize accessibility"—Claude 2 assumed everybody walks on two feet. But obviously that's because the researchers at Anthropic, maybe not a lot of wheelchair users, but then the people said, "Actually not everybody walks, you have to be accessible to wheelchair users and so on," and that became part of the Claude 3 constitution. So the good thing about a constitution is that it's written in everyday language, so anyone can look at it, can check it, and also can change it. So now we're looking at many, many different communities training their own community models by having their own community specification or community constitution. So now it's a very democratized way of using training.
9 · Dependence on AI, decision-making, and democracy
Well, my next question is a little about democracy. People are depending on AI so heavily and constantly—
Maybe kids, without machine translation, don't even know what to do.
Yeah, they ask AI everything, and many think less these days. If you look back at the long history of human beings, everybody used to think and then decide what to do. But AI tells you what's good, what's bad, and what you should do. So it's a kind of dangerous situation—if you talk about democracy, people should think for themselves and decide who to vote for. That kind of intelligence is necessary for everybody, but because of AI, people are becoming, maybe, synchronous these days. What do you think of that? What should we do to overcome that kind of difficulty?
Sorry, I'm not completely understanding—is it that when people go to vote, they used to think about who they should vote for, but now they just ask AI who to vote for, and whatever AI says, they vote? Is that the idea? Is that really happening? I'm not sure.
I'm not sure, but—
Yeah, I don't think that's happening.
Maybe they ask AI—I mean, you understand me very much, so maybe you can tell me who is the ideal person to vote for.
I mean, I don't think many people do that. I think most people use AI as a super search engine.
Super search engine?
Yeah, so if they want to search for something, they want to ask questions, if they're curious, if they want to make a project, they ask AI to do some planning. But I think the goal is still with the person—I think most people are still using AI this way. What you're saying is what my acquaintance, Cory Doctorow, called the "reverse centaur." So he's saying, you know, there used to be a system called Mechanical Turk—over the internet you could ask a human to do things. Maybe an AI couldn't yet solve a captcha, so the AI wants to get into a system and hires a human to solve the captcha for the AI. And the human doesn't know why—"I'm solving this captcha, is it criminal activity, I don't know, but I'm paid by the AI, so I'm just doing what I'm asked." So it's unlike a centaur, which has a human head and a horse body—a reverse centaur becomes a horse head with a human body. So humans do meaningless work, AI assigns all the work to humans, right? So maybe you're saying something like that—AI makes all the decisions, all the judgments—
Yeah.
And humans just follow the manipulation of AI. First of all, I don't think that's happening.
Okay.
Yeah, I think it's a danger—it's almost like watching a Black Mirror episode. The importance of art, which I believe is the most important profession, is to warn people: if you step a few steps more, there's a cliff, you will fall off the cliff. So if you watch Black Mirror, it's not "oh, it's a great idea"—no, it's about not going there, right? So I think, of course, for Nikkei and journalism in general, it's very good to say, "Okay, we don't want to be there." But I think there's a difference between that and saying "we're already here." I don't think we're already here. I think most people use AI systems not to say, "I give up on the decision, on voting, tell me who to vote"—I don't think that's happening.
Okay. Yeah, so what I'd like to explain was that people should think more, because if they depend on AI so much, being synchronous could be dangerous—they don't know what is good or what is bad, so AI just says so, and that's what I do. AI is a kind of genius, so all we have to do is follow that advice. So I think that kind of world is very dangerous.
Yeah, of course, it's very worth guarding against.
You should warn people to avoid that kind of situation. What should we do? What kind of education is necessary—maybe liberal arts kind, like a Greek philosopher asking "what is light, what is justice"—what kind of education is necessary for people?
10 · Education, the gym parable, and what AI should not automate
Yeah, I think, again, reusing the metaphor of going to the gym: usually if I go to the gym, I want to exercise to get muscle, I also want to make friends who share my hobby—that's the two most important things about going to a gym. I'm less interested in competing to be the person who lifts the most weight—to me, it's not a very useful competition, I'm not there for competition. I think most people going to the gym are not about winning Olympic gold, that's a professional athlete's job, most people going to the gym aren't going there for that, right?
But imagine if a gym now blocked everyone away, everybody in their own silo, in their own set, and then they only publish a scoreboard of the weight you lifted, and every week the champion gets all the benefit, right? And the gym changes its rule—it's a little bit like the education system. If you finish top score in the class, you get all the applause, you get the social status, you get into the best university, but even if you're just a few points less, but you're second place, third place, fourth place, they get only a fraction of the applause, even though you're just one point apart, right? So it's not very useful. But if you put people in that situation, with AI they will cheat—like, the rational thing for me in that situation is to send my robot to the gym. The robot takes my gym card and the robot can lift a lot of weight, okay, and I become first place. And then once people know, "oh, they used a robot to do that," then I have to hire a better robot to lift even more weight, but then we all lose muscle, because the robots replace us, and then we all lose friends, we don't have friends anymore. And that's the Black Mirror, that's this topic.
But then it's a function of the setup of the gym. If the gym takes the silos away, if people exercise together again, if you don't reward the positional status but only reward actual heavy muscle—if the school rewards students who are good team players, who make good outcomes, good impact on society, if we reward curiosity, collaboration, civic care—then none of this can be automated. And then AI is just a scaffolding to help us train better, right? Previously, maybe we spoke different languages, we couldn't be a team, but now with AI translation, oh, we can be a team again, so we exercise our muscle, right? So the point being: the education system can reward people for acting like a robot, and then students will cheat and actually just send a robot. Or we can reward people for acting like people, and then we would never cheat, because the AI is just there to help us communicate.
11 · Employment, transcription, translation, and the care economy
Okay, thank you. And actually, my next question is your view on employment. Many people say AI takes jobs from human beings—yeah, it's been repeated so many times, but many people ask that question. But recently, because of AI agents and Mythos, especially consulting companies and maybe bankers go away—many people mention that. So what do you think about AI and employment?
I mean, it was true that when I spoke and you saw the wall of self-correcting text, that's one less human interpreter, right? So obviously there's one less job for yesterday, right? I'm not denying that in some cases this is happening. Another good example is transcription—it used to be that the transcriber, you needed a court reporter to listen and type very quickly. But no, now you just need an editor—the AI does the typing, you just correct some of the notes, right? For journalism, it used to be you had to do the transcription yourself, but now AI transcription is so good, right, you just correct some words and you're done. So again, that's maybe one job replaced.
In fact, if you look back, the word "computer" used to describe a human, because the human doing the calculation was called a computer. If you go back, "printer" used to describe a human—Gutenberg was a printer, not a printer machine—you had to have movable type, right, put it into a canvas and then put it into a frame, a wooden frame, right. But of course now the printer is just a machine, right. So I'm not saying this will not happen—this will probably happen. So what used to be called a transcriber or an interpreter may more and more describe a machine, not a human. It also frees up, of course, the human mind to do more people-to-people work, because there is some kind of job that even if machines can do, we will never give machines that job. For example, if I'm doing a confession to a priest of a certain spiritual following—
Yeah.
—most of the work really is just to empathize with me, right, to listen to my story and say that according to the holy scripture, God would absolve me if I do this, if I do that, right. So the job itself, you can probably train an AI to do that today, but people don't go to robots for confession, because it's fundamentally a human job. So my point here is that there's something relational about it, it's not transactional.
The thing about interpreters—I used to be an interpreter too. Very few speakers build a real relationship with their interpreter. Usually the speaker goes to different venues and pairs with different interpreters, and usually the interpreter feels that once their job is done here, they're not going to meet that speaker again, in this kind of non-relational situation. But machines are going to be more relational, even, because I can bring my own language model to that wall. So in a sense, that language model is closer to me than the human interpreter I never knew, right, so it feels closer to me. But then in the confessional or priest situation, over the years, of course, we will know each other much better, and a robot could never be as close in a spiritual sense, right. So I think more and more people will shift to this care economy, where relational goods are prioritized.
