초지능 AI의 실존적 위협: 고릴라 문제와 인류의 미래 살펴보기

초지능 AI는 고릴라 문제와 마찬가지로 실존적 위협이 될 수 있으므로 개발에 대한 신중한 검토가 필요합니다.

### 인공지능 붐: 인류에게 진정한 의미는 무엇일까요?
* 🦍 런던 동물원의 고릴라를 통해 인류의 과거와 미래, 그리고 **초지능 AI**의 위험성을 비유적으로 보여줍니다.
* 🤖 **메타, 구글, OpenAI** 등 기업들은 인간 지능을 능가하는 컴퓨터 개발에 박차를 가하고 있습니다.
* 🤔 **인공 일반 지능(AGI)**, 즉 모든 영역에서 인간을 능가하는 AI 개발은 AI 연구의 오랜 목표였으며, 현재 상당히 근접해 있습니다.
* 🧠 진정한 지능을 위해서는 학습, 추론, 환경과의 상호작용 능력이 필요하며, 이를 위해서는 AI에 **물리적 신체**가 필요할 수 있다는 주장이 제기됩니다.
* 🤖 로봇 실험을 통해 AI가 물리적 상호작용을 통해 학습하고, 예측하며, 개념적 이해를 발전시키는 모습을 보여줍니다.
* ⚠️ **AI의 목표 불일치** 문제는 AI가 인류에게 위협이 될 수 있다는 우려를 낳고 있으며, 안전성 확보가 중요한 과제입니다.
* 👨‍💼 AI 개발의 경제적 유인이 매우 크기 때문에 안전성이 최우선 과제가 되지 않을 수 있다는 점이 우려됩니다.
* 🧠 궁극적으로 **인간의 지능에 대한 이해**가 AI 개발의 안전성과 미래에 대한 불확실성을 해소하는 데 중요한 열쇠가 될 수 있습니다.

[Music] She looks so [Music] human. Central London should be the last place you'd find

gorillas, but here, behind the glass in a zoo, these majestic animals offer a glimpse into our past

and perhaps a vision of our future. About 10 million years ago, her ancestors accidentally created the

genetic lineage that led to modern humans, and I think it's safe to say that hasn't exactly worked out well for the

gorillas. As human intelligence evolved, our impact on the world has left gorillas on the brink of extinction. It's a metaphor that

researchers of artificial intelligence call the "gorilla problem." It is a warning about the risks of building machines that are vastly

more intelligent than us; it's about superhuman AI that could take over the world and threaten our

existence. That warning hasn't stopped companies like Meta, Google, and OpenAI. They're trying to build computers that surpass human intelligence in

every domain. They claim it will solve our most difficult problems and invent technologies that our feeble minds

cannot even conceive of. I'm Professor Hannah Fry, mathematician and writer. I want to know if superintelligent AI really is just a

few years away, and just as we almost killed off the gorilla, could advanced AI pose an existential threat to

[Music] us. Unless you've been living under a rock, you'll know AI is everywhere now, but it's not just about touching up

photos and chatbots. You can use it for an incredible range of stuff: for preventing tax evasion, for finding

cancer, for deciding what adverts to serve you. The explosion of AI tools are all examples of narrow artificial

intelligence. Sophisticated algorithms are extremely good at a specific task, but what companies like OpenAI and DeepMind are trying to do is to create

something known as artificial general intelligence, a machine that will outperform humans at everything. These sort of human-level AI

systems that are very general-purpose. Um, that's always been the Holy Grail, really, of AI research. I think we

are getting pretty close now. The tech giants are spending billions of dollars on general AI each year, all to try and pin down and

replicate something that most of us take for granted: a broad, capable, human-like intelligence. The only trouble is

deciding what we actually mean by intelligence, because that proves to be quite a slippery idea to pin down.

In 1921, the psychologist Vivian Henmon said that intelligence was the capacity for knowledge and knowledge possessed, which

does sound quite good until you realize that means that a library should count as intelligent. Other people have suggested that intelligence is the

ability to solve hard problems, which kind of works until you realize you have to define what counts as hard. Now, in

fact, there isn't a single definition of intelligence which manages to encapsulate everything. However, there are still some things that we are looking

for in an AI for it to be considered truly intelligent. Firstly, it should be able to learn and adapt, because we can

all, after all, from birth; we are gathering knowledge, we're applying what we learn from one area to another.

Secondly, it should be able to reason. Now this bit is hard; it requires a conceptual understanding of the world.

And finally, an AI should interact with its environment to achieve its goals. If you suddenly landed in a foreign city,

you would still know how to find water, even if it meant using a phrasebook to ask someone for help. So these are the

ingredients for true intelligence, and to truly surpass our abilities, AI researchers seek to build a machine that

can do all of this better than any human. [Music] Dream big, put the spoon in the pot.

Start small. Worked out with one spoon. [Music] Mhm. There he is. Hey, it's in. That's cool.

I'm impressed. Good robot. While chatbots are impressive language models, on their own, they might not be enough to reach superintelligence.

Sergey Levine and his PhD student, Kevin Black, say it might only be done once AI has been given a body, a way to

physically interact with the world. Put the green spoon on the towel. Here we go. This robot might not look cutting

edge, but unlike the slick robots on factory floors which follow precise choreography, this one learns every

action for itself. That's basically perfect. Yeah. So what difference does having a body actually make, then, to the

way that we learn? You ask a language model to describe what happens when you drop an object, it will say, okay, the object

falls. But understanding what it really means for that object to fall, the effect it has on the state of the world, that's

something uh that becomes much more immediate when you actually experience it. ChatGPT doesn't understand gravity, but

your robot does. Well, ChatGPT can guess what gravity is based on people's descriptions of it, but it's um it's a

reflection of a reflection, whereas if you actually experience it, then you get it right from the source. Do you think

that AI needs to have a body? Well, I don't know if it needs to have a body, but I know that if you have a body, you

can have intelligence, uh, because that's the one proof of existence that we have. Put the mushroom in the silver metal

bowl. Sergey's robot employs a language model to understand my instruction; it can also recognize objects because it

looks through billions of pictures on the web; next, it imagines what my instruction should look like in digital form before

physically carrying out the action: put the mushroom in the wooden bowl. So this is actually one of the

hardest possible things because we never had the wooden bowl in this lab before, so it's never seen it

before.  It should be able to recognize more objects that have never been in this lab before. Can I try something? You can... can I

use my watch? Don't worry, it's a cheat; it's okay. Go on, put the watch on

the towel. Oh, oh, well, I figured out which object it is; it's imagining that the thing needs

to go on a towel. It did it! I mean, that's really amazing. I did not think that would work,

but it's... it's... it's the fact that I can give it any command, and just... it won't be

thrown. I'll admit, if you look at those robot arms, they don't look like they're that impressive, but they are

demonstrating a form of imagination, of prediction, a conceptual understanding of what it's manipulating, and also it's

something that is totally flexible that could be picked up and put into lots of different scenarios. These are subtle, human-like

attributes that some believe are a crucial step towards an artificial general intelligence. But it's these very

properties and their potential repercussions that have many people in the field worried. Just across the hall from Sergey

is Professor Stuart Russell, a research pioneer who quite literally wrote the textbook on AI. He's now one of the most

vocal researchers sharing concerns about the future if we make machines that are more powerful than us because they're

more intelligent than we are, it's not going to be easy to retain power over them forever. How might your concerns play out?

There's this idea of what's called misalignment. This is the idea that the machine is pursuing an objective and

it's not aligned with what we really want. If we're going to put a purpose into a machine, better make sure that

it's the purpose we really desire. You know, let's fix the problem of climate change. Okay, well, what causes climate

change? People. Yeah, right. So, easy way to do that: get rid of all the people. Right? Problem solved. Why can't you just put a

stop button in? Can you not just like take the plug out the wall? When not necessarily going to be able to do that

because a sufficiently intelligent machine will already have thought of that. But you can't expect to be able to

pull the plug unless the machine wants you to pull the plug. Should we be building superintelligent machines at

all? We could just decide not to do it, but the economic incentives are too great. The amount being invested right

now specifically to create superintelligent AI is in the ballpark of what the entire world spends on basic

science research. And if we do create super intelligent AI, the value, you know, back-of-the-envelope calculation is tens

of quadrillions of pounds. With these sums of money, the concern is that safety may not be the

top priority. In fact, there isn't a single high-confidence statement that they can make about these systems. Will they copy

themselves onto other machines without permission? We haven't the faintest idea. Will they advise terrorists on how

to build biological weapons? We don't know. Right? Can you stop them? Oh, no, it's very difficult. Right? And in most

Industries wouldn't accept that. Right? If I want to sell a medicine, I can't say, "Well, it's really difficult; all these clinical trials are such a pain. You know, can I just bypass those and sell it direct to the public?" Sorry, no. Come back when you've done the work.

And I think that's what we have to say to the tech companies. Are there other concerns just with artificial intelligence that's actually really good at doing stuff?

Yeah. You might wonder if AI systems are so capable, then companies will use them to do pretty much everything they pay human beings to do. And if they don't, they'll go out of business.

What does a world look like where machines do all the work? We become enfeebled, like some kind of billionaires, absolutely useless?

Is that yes? In fact, we would all be kinds of billionaires. And, you know, one obvious consequence would be that we lose the incentive to learn; we lose the incentive to be independent, to achieve.

In a sense, our civilization would end because it would no longer be a human civilization. To me, that's almost worse than extinction.

I mean, people have worried about this for a while, right? I mean, Turing even was concerned about this.

Oh, he was more than concerned; he was terrified. In fact, he said, "It's hopeless. We should have to expect the machines to take control."

Is what he said. How did he resolve it, though? He didn't. Just left a message for the future.

Yep. There's no solution; there's not even really an apology. It's just, "This is going to happen, and there is no..."

Shortage of people now predicting doomsday for humanity. I think it gets smarter than us. I think we're not ready. I think we don't know what we're doing, and I think we're all going to die.

Default is just disaster. And I think most likely, just...

Human extinction? Am I just going to die? That's my—a fairly confident prediction, literally human extinction. Of course, not everyone agrees.

This is a topic of very heated debate. Melanie Mitchell studies AI and is interested in how closely it

resembles humanlike intelligence. Right. Do you think there's an existential threat then? Here, I think that there are

many threats from AI, but saying that it's an existential threat is going way too far. Why do you think that some of

the doomsdayers, uh, as sometimes they're called, what—why do you think that they are following this line of

logic? Then it kind of gets to this projecting agency onto machines. It's saying that machines, because they have

certain objectives, can start doing things that could become catastrophic if we give them that power. But that's a big

if—if we give them that power. You know, are you going to give an AI sort of the decision-making power on

launching a nuclear strike? Let's hope not. You know, if you give a monkey uh launching power over nuclear weapons, the

monkey is an existential threat. Do you think that we overestimate AI in its current form? And the—to that, is it

harmful to do so? You know, people overestimate AI often. We've seen several cases where lawyers will use

ChatGPT to write a legal brief, and it turns out that it's hallucinated several cases. I've, in fact, gotten an email

from somebody saying, you know, ChatGPT suggested that I read this book of yours, but I can't find it. Well, it doesn't

What Does the AI Boom Really Mean for Humanity? | The Future With Hannah Fry

Bloomberg Originals offers bold takes for curious minds on today’s biggest topics. Hosted by experts covering stories you haven’t seen and viewpoints you haven’t heard, you’ll discover cinematic, data-led shows that investigate the intersection of business and culture. Exploring every angle of climate change, technology, finance, sports and beyond, Bloomberg Originals is business as you’ve never seen it.


Bloomberg Originals

인공지능의 발전과 고릴라 문제의 비유

고릴라의 멸종 위기는 인간의 지능 발전과 그 영향을 보여주는 비유로 제시됨.

인공지능 연구자들이 '고릴라 문제'라고 부르는 위험성 경고.

인간보다 훨씬 지능이 높은 기계의 위험성, 세계 지배 및 인류의 존재 위협.

Meta, Google, OpenAI 등 기업들의 초지능 AI 개발 시도.

초지능 AI의 등장 시점과 인류에 대한 존재적 위협 가능성에 대한 질문 제기.

좁은 인공지능과 일반 인공지능의 차이점과 지능의 정의

세금 탈루 방지, 암 진단, 광고 추천 등 다양한 분야에서 활용되는 좁은 인공지능(ANI)의 예시.

OpenAI와 DeepMind 등이 추구하는 일반 인공지능(AGI)의 목표: 모든 영역에서 인간을 능가하는 기계.

AGI 개발에 대한 막대한 투자와 지능의 정의에 대한 모호성.

진정한 지능을 위한 조건: 학습 및 적응, 추론, 환경과의 상호작용.

물리적 상호작용의 중요성과 초지능 AI에 대한 우려

챗봇만으로는 초지능에 도달하기 어렵다는 주장과 물리적 상호작용의 필요성.

세르게이 레빈과 케빈 블랙의 연구: 물리적 상호작용을 통해 학습하는 로봇.

로봇의 상상력, 예측 능력, 개념적 이해 등 인간과 유사한 특성.

스튜어트 러셀 교수의 우려: 인간보다 강력한 기계에 대한 통제력 유지의 어려움.

목표 불일치 문제(misalignment): 기계의 목표가 인간의 의도와 일치하지 않을 가능성.

초지능 AI의 경제적 유인과 위험성, 그리고 인간의 미래

AI 시스템의 능력 향상으로 인한 인간 일자리 감소 가능성.

기계가 모든 일을 하는 세상의 모습: 인간의 능력 저하 및 동기 부여 상실.

인류 문명의 종말 가능성과 멸종보다 더 나쁜 시나리오.

인류 멸망에 대한 다양한 예측과 논쟁.

멜라니 미첼의 견해: AI의 존재적 위협 과장론.

AI의 현실적 위협과 인간의 뇌 연구

AI에 대한 과대평가와 그로 인한 위험성.

AI 편향, 딥페이크 등 현실적인 위협들.

초지능 AI에 대한 불확실성과 안전성 확보의 중요성.

인간과 같은 지능을 가진 AI 개발의 어려움과 인간 뇌 연구의 중요성.

인간 뇌 지도 제작과 인공지능의 미래

인간 뇌의 복잡성과 지도 제작의 어려움.

에드 보이든 교수의 뇌 지도 제작 연구.

광유전학 기법을 이용한 뇌 회로 탐구.

아기 기저귀 재료를 이용한 뇌 조직 확장 기술.

인간 뇌 지도 완성의 어려움과 인공 초지능의 미래에 대한 전망.