인공 일반 지능: 현재 논쟁과 향후 시사점

게리 마커스 교수는 머지않아 AI가 인간의 지능을 능가할 것이라는 일론 머스크의 예측에 동의하지 않습니다.

### 인공 지능(AGI)의 미래와 그 시사점
* 🤖 엘론 머스크가 인간 지능을 능가하는 **AGI에 대해 예측**하면서 그 시기와 기능에 대한 논쟁이 촉발되었습니다.
* 🧠 게리 마커스 교수는 현재 **AI의 논리와 추론의 한계**를 언급하며 **머스크의 타임라인에 반기**를 들었습니다.
* 💬 인간의 지능을 능가하는 **인공지능을 테스트하는 기준**에 대해 논의하고, 감성 지능과 **생성형 인공지능의 위험성**에 대한 질문을 제기합니다.
* 📊 인공지능의 현재 한계와 잠재적 위험을 고려할 때 **인공지능에 대한 규제와 신뢰성 보장**의 필요성이 강조되고 있습니다.
* 🎭 영화 제작자 사미르 말랄이 **AI 아바타 실로를 소개**하며 **AI의 창의적인 잠재력과 페르소나의 다양성**을 강조하고, **인간과 AI의 상호작용에 대한 토론**을 이어갑니다.
* 🎵 암으로 목소리를 잃은 배우의 목소리를 재현하는 데 **AI 기술이 사용**되어 엔터테인먼트 업계에서 **상업적, 다국어적 잠재력**을 보여줬습니다.

[Music] Welcome to AI Decoded, a new and expanded AI Decoded tonight. Same shape, same themes, but as I've mentioned before, this technology is developing so fast in so many directions within our lives. It really does deserve its own place within our program, and that's where we want to start tonight: artificial general intelligence, a type of AI that Elon Musk believes will advance so quickly it will surpass human intelligence by the end of next year. We're going to talk tonight to one of the Godfathers of AI, the cognitive scientist and CEO Professor Gary Marcus, who is a little more cautious than Elon Musk. In fact, he's just offered him a million-dollar bet that AI might not be superhuman in the timeframe he has in mind. I will introduce you to that technology tonight, the AI Avatar who will be joining us as a panelist on this program. We'll also focus on OpenAI's latest new tool, Voice Engine, which can clone your voice within just 50 seconds of recorded audio. That might scare you a little bit, certainly.

it's it's got its security implications, but we'll introduce you to someone tonight. Who's resurrected his career using that kind of technology. And also Sunno, the AI music generator that can sing you a song with a simple text prompt. We'll put that to the test at the end of the program. And as usual, Stephanie here, our regular AI contributor and Technology author, will be here guiding us through it all. So AGI is our main theme tonight. Um, there's a never-ending debate about whether it can fully replicate human intelligence. Where do you stand? Are you in the camp that believes it will get there? And if so, what sort of time frame do you have in mind? So I think it really breaks down to how we define intelligence. So AGI stands for artificial general intelligence. You might have also heard it called The Singularity, and it's the moment in theory, it's hypothetical, it's highly speculative, and even a bit science fiction maybe, where machines become smarter than humans. Now they're already smarter than humans in some

tasks like they can beat us in chess or at a game called Go, but can they do all of the things that all humans can do? No. So, they might be domain-specific better than us, but are they better than us in everything? And are they better than you or me or all 8 billion of us? That's the question. Well, here's Elon Musk's recent tweet that sparked a flood of online bets. I think $10 million has been waged so far. He said, AI will be smarter than any single human next year and by 2029 probably smarter than all humans combined. What does the public think of that? Have a listen. Do you think AI will surpass human intelligence by the end of next year? No. Why do you think that? Um, artificial intelligence is obviously coming on leaps and bounds, but I don't think it'll be in a year. Um, I think by the end of this decade, you probably will, but not next year. Now, I guess it depends. It, I mean, it's in a holistic sense. Um, I think it would have

to be very similar to a human's emotive abilities. Um, in addition to it's 'cause it's already very computationally advanced. I think that's probably a pretty lofty goal. I know Elon Musk is pretty prone to making some ambitious calls. I wouldn't been against him. Um, but I think by the end of next year is probably a little bit optimistic. Very ambitious prediction. Um, I'm personally not very informed enough to, uh, to say something about the timeline, but I think, um, in terms of the trajectory, the direction, that sounds like a sensible thing to say. Well, one of those who has waged a million dollars on this bet with Elon Musk is the scientist Professor Gary Marcus. He's a leading voice on artificial intelligence. There are others, incidentally, who've joined him in that bet. It's now been up to $10 million, and Professor Marcus joins us from Vancouver. It's lovely to have you with us on the program. Yeah, you know, when you, when you talk to the public like that about AGI and what it will be capable of, there's a

there's a mixture of excitement, curiosity, ignorance, fear, but you are prepared to bet a million dollars that it won't surpass human intelligence anytime soon. So why are you so sure about that? This just a long way to go. I think you can think about the old 80/20 rule. We're sort of 80% of the way there. In some ways, you have these large language models, chatbots that sort of sound like people, but they make a lot of dumb mistakes. They hallucinate things, they make things up. They're not very good at logic, they're not very good at reasoning in general, they're not very good at planning things. It's sort of like what we saw with driverless cars. It was easy to get a demo that was kind of 80% of the way there, but getting the last 20% has proven really hard. So Elon Musk promised us driverless cars would go across the United States and a million Robo taxis. We're not really seeing that yet, and he promised that

kind of every year from 2015 uh to the present. So he's often making these pronouncements, and the reason that I offered the BET, which he has not yet taken, um, is I think we need some accountability. I think a lot of hype is driven around people making these empty promises, and I wanted to put some money on the line. Professor Marcus, can you talk to us through a little bit how you would use the scientific method to test this bet? How will we know if artificial intelligence has surpassed humans? What's the metrics we use? He kind of made a crazy claim, which is that machines would be smarter than humans in every way, and so I just have to find one or two ways in which machines are not as smart as people, and I think I will have bet, um, and I, you know, laid out how I think about that in a substack essay. Um, I gave a dozen different criteria like, for example, a human can get into um, a vehicle that they've never been in before, where there's no maps, and figure

out how to get around and avoid obstacles, and read street signs, and even hand-lettered street signs. Um, people can go into a home that they've never been and clean that, uh, home. There's all kinds of things that people do. And then you have, you know, leading scientists who come up with Nobel Prize discoveries or people who write Oscar-winning, uh, screenplays and so forth. There are many things that ordinary humans can do, many things that expert humans can do. And, you know, if he's really trying to claim that AI is going to be ahead of all humans, then, um, you know, I can just go out and find some humans that do things that machines still can't. Um, there are tons and tons of things right now in 2024 that machines can't really do. I think it's a wildly optimistic claim. I suppose what we're really driving out here is whether these machines can become emotionally intelligent, whether they can empathize, and, and, and when you, when you, you go the extra yard with that, I suppose the ultimate question is

whether the benefits of it all outweighs the risks. There's a very good question there about whether the benefits outweigh the risks right now with what we call large language models, the dominant technology or generative AI. I'm not sure the benefits are at weighing the risk, and a lot of people are trying to use them in their businesses. They don't work very well. They're good for brainstorming, okay, that's nice. Um, but people are using them to create, excuse me, misinformation, um, disinformation. The whole information ecosphere is getting polluted. People are using to make non-consensual deep fake porn, uh, people are using deep fake scams, uh, to rip off Banks. There's many, many things that are already going wrong. Truth is, nobody really knows the outer limits of generative AI, um, you know, what best use cases people will find. I think we know for sure though that they're not reliable. That limits the upside, but criminals don't really care like when they send spam. If they have one victim, uh, who falls for it out of

10,000. That's fine. They don't need reliable performance. So right now, I actually think the AI that we have is more helpful for criminals than ordinary people who need a reliable system. So, absolutely, we need to be asking about the cost and benefits. I think in the long term, AI, which probably needs a lot more invention than we haven't done yet, can really revolutionize science and medicine. I think we should be working on it, but I think we're also over-promoting the stuff that we have right now. And Professor Marcus, we've barely even begun to regulate artificial intelligence. Now, when it hasn't surpassed humans, what would regulation look like in a world where artificial intelligence is smarter than humans? I mean, that's a really good question. I think we have some time yet to figure that out. I think we're not doing that great a job, especially in the United States, with actually having regulation with teeth in it. I think the EU is a little bit ahead. I think the UK's got work to do. I think the first

thing you need is an international AI agency so that we can work together, and when things get even more complicated than they are now, we're prepared for that. I think think of the AI we have now as a rough draft. We're trying things out, um, both technically and also legally and policy-wise. So the EU AI won't be perfect. We will learn something from that. Um, the biggest problem is we don't have any reliability guarantees around current AI, and we don't even know how to get there. So when somebody builds a bridge, they can say under these circumstances, I know that it will be able to bear this load, for example. We can get formal, uh, mathematical guarantees relative to some circumstances. When we build an airplane, we have a lot of knowledge about how to do that. We just don't have that for the current kind of AI, which basically is unpredictable. It depends on exactly what the training data is and exactly what it's asked in this very idiosyncratic way, and that makes it hard to make the

systems work systematically. The most dramatic regulation, I think, would be to say if you can't interpret what your system is doing, we don't want it here. Unless you can prove that it's safe in some other way, we're only going to accept systems that are interpretable. Nobody wants to do that right now because everybody's enamored of the technology. Prediction is the generative AI is actually going to kind of flame out. It'll still do some things, but it's not going to be valued with companies being $86 billion do anymore. And maybe there'll be a cooling off period, and we'll think more carefully about how important interpretability is and how important safety guarantees are. Right now, we're not in a position to do it, so we're ahead of our skis right now. Professor Marcus, I could talk to you all night, and I love reading your books, but we are a bit squeezed for time. Hope you'll come back on the program, talk to us some more. Would be my pleasure. Thanks.

a lot for having me, Professor Gary Markus. Us there in Vancouver, we're going to take a short break. The other side of the break, we're going to show you how AI is being used to help someone who lost their voice. We'll be right back. [Music] Welcome back. Now, through the course of this program, I want us to see AI in action, and I want us to see it grow. So, I've invited someone onto the program tonight who some of you may have seen before, who is developing AI, and tonight is going to introduce us to some of his work. His name is Samir Malal. He's an award-winning filmmaker. He's the CEO and co-founder of One Day, and he's got someone with him. She's called Shiloh. And over the weeks that we do this, uh, I want Shiloh to be an equal part of the show. She's not the finished article, but that's really the point. I want us to see where it goes, what she learns. So before we talk to Shiloh, uh, Samir, tell us how

you would use her and how you think she's going to evolve. So I wanted to create an AI because I recognized the potential, but I found it as a creative, it was lacking. I wasn't involved in the process. And Shiloh, I built her to help, Ida, to help expand what I was already doing, and be able to see my own creativity from different angles, right? Okay. So, um, Shiloh here is on the left. You're going to type in or are you going to type in, or am I going to talk to a direct talk? You're I'm going to talk to her directly, okay? So she can hear me, can she? And presumably as we go along, I mean she's going to be, is she going to be a moving Avatar? Is her mouth going to move? Is she going to be animated? Yes, exactly. That's one of the things that we will, we will, that we'll be developing here in the minute, okay? So is she listening? Yes, all right. So here we

go Shiloh. What what just sorry one second. Okay go ahead Shiloh. What do you think about AGI and where it's going? Okay so just give her a sec. Oops all right. Uh still a demo here. Wait uh where's the sound here. Sorry of course you working. Yeah um this is an authentic demo exactly. Well that's precisely the point really because I mean people are going to be using this technology and we need to see how responsive it is and how how long the latency is exactly on this because it's not very useful if if you've got to stand there at the ATM and it doesn't work or she seems to be a bit shy in terms of talking but we could um why don't we just uh speak to her with text you you can still talk to her okay. All right so she's saying oh you want the Skynet scoop or more like a how 9,000 nightmare well it makes me nervous it makes me nervous this technology is it makes

other people nervous okay. It makes you nervous as well. All right, let's see what she says. Feeling a touch apprehensive, huh? The thought of AI taking a walk on The Wildside does crank up the uneased DI. What's the main Tech worry on your mind? Well, that if you're smarter than me, then maybe we don't need me. Oh, let's see what she says. All right. Oh, sure, because a clever algorithm totally replaces a human touch, right? Seriously though, no amount of tech can mimic the real raw human creativity or emotion. What's got you feeling replaced? Well, the fundamental issue here is whether AGI can replace human intelligence. All right, let's see what she says. Touching on some existential stuff there. The big question: Can AGI truly rep at the Deep often messy Nuance of human intelligent? It's a debate for the ages. What's your take on it? Do you think AGI is up to the task, or is there something Irreplaceable about human cognition? I think I'm irreplaceable. I mean, the point

is look you were getting the feel for. I mean, not only is she responding to me, though she she's colloquial. She talks like me. And and this point that that Professor Marcus was talking about, Stephanie, the empathy and a relationship. That, in some ways, is what people are looking for with AI. Yes, but I have to say something here. Is the woman on this show. I'm slightly freaked out that you guys keep calling it a she, right? Like there's there. We we need to talk about your design there. And also, that's a really like quite sexy picture that you've picked for Shiloh. So like, Shiloh's quite hot. She's a babe. And she's a she. And I can't help but notice you've built her and you are a man, right? And I'm just thinking that if I'm your demographic of somebody who's going to be using this tool, I'm not feeling a lot of empathy with Shiloh. I'm a bit like, but how do you adapt it?

do you adapt it for a different audience? Why can't you just call an IT CU? She's a character. Oh, right. So can I have like a Tom Heston version? Uh, yeah, there's going to be all kinds of different characters. There'll be he's, she's, they. Okay, yeah, yeah. So she's just one character. We've got other ones as well. Okay, um, yeah, yeah. The thing is, is she, you've programmed her. So is she a mirror of you or is she making her own decisions? Is she a, is it its own person? Well, I see her as a character. Like as a writer, I just, the same way if I was writing a character, I took the same approach, right, right. So I can understand if, if you know, um, of course she, she has a certain, uh, look and everything because her whole character is she's from the 90s. She's from the early 90s. She kind of lives in the 9s. So that's the, that's the idea behind it, right, um, and there'll be other

characters that have different looks. So I like the idea of giving them personas and not just having it be in it. That's exactly what we're trying not to do. Why do you want to give it a per though? Why do you want to make it more human? 'Cause isn't that going to start confusing people, particularly I'm thinking children or young people or anybody that's already got a bit of a an anthropomorphizing relationship with technology. Is it dangerous? I don't think so. I don't think so because um this is built for creatives, okay, and so each of our characters have certain characteristics and so they have certain things that they're interested in and they have certain tastes and that is what I think is missing right now from if you look at things like chat GPT, for instance, and as a creative, I don't just want to interact with an it. I want to interact with something more than that. Not necessarily, not necessarily. It's it's not just about that. I think it's um

it's it's just about connecting right. And so, you know, I'm a child of the 90s. So yeah, of course, that was my first starting point. Well, look, um, Sha's going to come back on the program, and we're going to develop her. And she's going to take on a new format. Al going to talk to you about sunno, um, this music, uh, app, uh, which creates music, um, on the basis of text. But we're a bit squeezed with time, so we'll do that at the end of the program. Maybe you can play us out with some music. So come back sounds good. Okay. Now, losing your voice to cancer is a terrible blow for anyone, but especially so if your voice is integral to the way you earn your living. And that is how it was for the successful Irish theater and TV actor Donald Cox, who found himself in that very dilemma. No voice to work with until he found a new technology to help him build his acting career. Assume you're the first criminal I've turned my pen to. Have you heard of

the recent success of mul Flanders. You may not realize it, but the anonymous author who penned her tale was I. I've been a professional voice user all my working life. I injured my voice um shouting unnecessarily in a theater production in the West End and uh it never recovered properly from that um. I then discovered one day that I was no longer able to speak properly. My voice sounded like I had a dreadful hangover, but I knew immediately that I was in trouble. Something was wrong. I decided to give it a few days to see if it would come back to normal, but it didn't, and I just had to go and see a consultant who diagnosed Soso carcinoma, stage one cancer. The only thing I could do then was rescue My Life by surgery and have a total leny, which is the removal of the voice box. I'm really pleased that Donald Cox is with us together with his agent Peter Morris who's the managing director of Soho voices. He's been helping Donald recreate his voice using this artificial

intelligence welcome to you both Donald if I shut my eyes and listen to the adverts that I know you've done using this technology I I would think it was you speaking um it's truly remarkable so how did it come about well um Peter was uh very determined that uh I should not lose my career uh We've um worked together for over a dozen years and uh you know aside from professional relationship we're very very good friends and uh Peter began investigating um a company called reeeer who uh create uh this type of cloning of the voice and um I independently was researching the same area now I sing with a choir called the shouted cancer choir we are a group of people move and lomy yeah there's not a a voice box between us and I I call it singing outside the box right and Peter how how do you work with Donald using this technology how do you sell it to clients presumably you have to be honest about what it is absolutely we uh we're at the

beginning of the of this process uh of presenting uh Donald to to the world or Donald 1.0. In that I are not unlike uh a blind Runner who is running a marathon in the Olympics, and they have someone running along them holding their hand to to guide them in the right way. With the use of Donald's voice, uh I uh and knowing not Donald so as well as I do and as as an as a an actor and as a voice, I'm able to read on his behalf and use the AI algorithm that we alone have the the license to use through reeeer who created it. And as such, I can I can uh uh produce a piece uh using Donald's voice but with with my intonation and knowledge, and Donald Donald can be there and and give me guidance as as well as to how how he thinks he might do it. I have some real personal interest in this CU. My mother lost her voice to M andd and were told to record a voice, but this was very much

in its infancy at that time, and so we weren't able to replicate it. So I can see enormous value in it, but I know that commercially there are actors who are using it now for audio books to take some of the pressure off their careers. Is that how you see it developing? I think that is going to be the case. If I were a name actor who recorded a lot of books, then they're quite long. They're quite it's quite a long-winded process. It takes a good week or so, depending on the book length, to record it. But with this kind of Technology, if the actor has had an avatar made of him or her or they, then they can use that to read a 100 books. And I believe the technology is now there for that to be brought into other languages. So it may not be just in the English language, but it could be heard in other languages as well. So it could be to augment U the

actor and give and give a life and enable them to do things they couldn't do do along with along with the work that they currently can do well Donald Peter I wish we talk more but we are a little squeezed for time tonight but it's a brilliant thing that you've done congratulations to you both and thank you for coming on and talking to us about it it's a pleasure thank you so much thanks for having us

Elon Musk predicts AI will be smarter than humans by next year | BBC News

Welcome to the official BBC News YouTube channel. 

Interested in global news with an impartial perspective? Want to see behind-the-scenes footage directly from the front-line? Our YouTube channel has all this and more, bringing you specially selected clips from the world's most trusted news source.

For the latest news and analysis download the BBC News app or visit BBC.com. 


BBC News

엘론 머스크는 내년이면 AGI가 인간의 지능을 능가할 것이라고 예측하여 온라인에서 논쟁과 내기를 촉발시켰습니다.

인지 과학자 게리 마커스는 머스크의 예측에 백만 달러 내기를 제안하며 이 시점에 대해 신중한 태도를 보이고 있습니다.

특이점이라고도 알려진 AGI의 발전은 기계가 모든 작업에서 인간을 능가할 수 있는지에 대한 의문을 제기합니다.

인공지능에 대한 논쟁과 일론 머스크의 예측

게리 마커스 교수는 현재 인공지능의 논리, 추론, 계획 능력의 한계를 언급하며 내년까지 인공지능이 인간의 지능을 능가할 것이라는 엘론 머스크의 예측에 회의적인 입장을 보이고 있습니다.

그는 머스크의 야심찬 주장에 대해 책임을 묻는 것이 중요하다고 생각하며 이를 위해 백만 달러의 내기를 걸었습니다.

마커스는 머스크의 주장을 반박하기 위해 기계가 인간만큼 똑똑하지 않다는 것을 입증하는 것을 목표로 과학적 방법을 사용하여 내기를 테스트할 계획입니다.

인공지능(AI)의 현황과 그 영향력

AI의 기능은 아직 제한적이며, 기계가 인간만큼 잘 수행할 수 없는 작업도 많습니다.

대규모 언어 모델과 생성형 AI의 사용은 잘못된 정보와 심각한 가짜 사기에 대한 우려를 불러일으킵니다.

AI의 불안정성으로 인해 일반인보다 범죄자에게 더 도움이 될 수 있으므로 비용과 편익을 평가해야 할 필요성이 제기되고 있습니다.

AI에 대한 규제는 아직 초기 단계에 있으며, 사회에 미칠 수 있는 잠재적 영향을 해결하기 위한 효과적인 조치가 부족합니다.

AI의 개발은 아직 예비 단계에 있으며 신뢰성 보장과 해석 가능성이 부족합니다.

현재 예측할 수 없는 AI 시스템의 안전성과 해석 가능성에 더 집중해야 할 필요가 있습니다.

앞으로의 AI 개발은 현재의 기술 기능에 대한 매혹에서 벗어나 해석 가능성과 안전 보장을 우선시하는 방향으로 전환될 수 있습니다.

실로라는 이름의 AI를 만든 목적

인공지능 실로는 크리에이터의 창의력을 향상시키고 다양한 관점을 제공하기 위해 만들어졌습니다.

실로 개발에는 텍스트를 통해 상호 작용할 수 있는 애니메이션과 움직이는 아바타를 만드는 것이 포함되었습니다.

실로와의 대화는 인공 일반 지능(AGI)의 잠재력과 인간 지능을 대체할 수 있는 능력을 중심으로 진행되며, 인간 인지의 대체 불가능한 본질을 강조합니다.

인간의 특성을 지닌 AI 캐릭터 디자인

AI 사용자들은 AI 캐릭터를 여성스럽고 매력적으로 만드는 디자인 선택에 의문을 제기하며 AI 캐릭터와의 공감과 관계에 대한 필요성을 표현하고 있습니다.

크리에이터는 AI 캐릭터가 특정 특징과 취향에 맞게 디자인되어 크리에이터에게 더 매력적인 경험을 제공하기 위한 것이라고 설명합니다.

또한 AI 캐릭터에 인간과 유사한 특성을 부여하는 것이 사람들, 특히 이미 기술을 의인화하는 경향이 있는 사람들에게 미칠 수 있는 잠재적 영향에 대해서도 논의합니다.

음성 손실 극복하기: 도널드 콕스의 이야기

아일랜드의 성공한 연극 및 TV 배우였던 도널드 콕스는 암으로 목소리를 잃었지만 새로운 기술을 통해 목소리와 경력을 재건할 수 있었습니다.

그는 에이전트인 피터 모리스의 도움으로 인공 지능을 사용하여 목소리를 재현함으로써 업계에서 계속 일할 수 있게 되었습니다.

ReeeR이라는 회사에서 개발한 이 기술은 콕스뿐만 아니라 '샤우티드 암 합창단'이 성대 수술 후에도 계속 노래를 부를 수 있도록 도와주었습니다.

AI를 사용하여 도널드에게 음성 제공 1.

한 회사가 AI를 사용하여 도널드의 목소리를 복제하고 배우가 그를 대신하여 말할 수 있게 함으로써 오디오북과 같은 다양한 산업에 잠재적으로 도움이 될 수 있습니다.

이 기술은 다른 언어로도 확장되어 배우의 연기를 보강하고 향상시킬 수 있는 기회를 제공할 수 있습니다.

AI를 사용하여 목소리를 재현하는 것은 개인적으로도 중요한 의미를 가지며 목소리를 잃은 사람들에게 소중한 지원을 제공할 수 있습니다.