추천 게시물
Geoff Hinton - Will Digital Intelligence Replace Biological Intelligence?
- 공유 링크 만들기
- X
- 이메일
- 기타 앱
Prompt : 위에 있는 유튜브 영상에 크게, 주로 두 사람이 나온다. 한사람은 사회자로 제프리 헌튼을 소개하고, 그 다음 제프리 헌튼이 강의를 한다. 마지막에 여러 사람의 Q&A가 있다. 위 영상의 speech를 text로 바꿔주고, 가능하면 각 텍스트의 화자를 구별해서 text로 바꿔줬으면 좋겠다.
1. 소개발표자: 멜라니 우딘(Melanie Woodin), 토론토 대학교 예술 및 과학 학장.
- 이벤트 공동 주최자: 슈워츠 리즈먼 연구소(Schwartz Reisman Institute), 컴퓨터 과학과(Department of Computer Science), 벡터 연구소(Vector Institute), 코스믹 퓨처 이니셔티브(Cosmic Future Initiative).
- 제프리 힌튼(Geoffrey Hinton)의 "신경망(neural networks)" 및 "딥러닝(deep learning)" 연구를 중심으로 한 인공지능(AI) 분야의 선구적 기여를 강조.
- 힌튼의 "이미지넷(ImageNet)" 대회 우승과 AI가 신경과학 및 사회에 미친 영향을 언급.
- "튜링상(Turing Award)" 등 그의 업적을 소개.
2. 제프리 힌튼의 강연주제: "디지털 지능이 생물학적 지능을 대체할 것인가?"
- 핵심 메시지:디지털 지능의 우월성(Superiority of Digital Intelligence):
- 디지털 지능은 지식 공유와 복제 효율성이 뛰어나며, 생물학적 지능을 능가할 가능성이 있음.
- AI의 이해력(Understanding in AI)
- GPT-4와 같은 **"대규모 언어 모델(Large Language Models, LLMs)"**은 텍스트 예측뿐 아니라 추론, 유추(analogies), 맥락적 의미 학습을 통해 "이해"를 나타냄.
- "유한 계산(Mortal Computation)" vs. "불멸 계산(Immortal Computation)"
- 디지털 계산은 하드웨어와 독립적으로 지식을 보존하여 **"불멸성(immortality)"**을 가지지만 에너지 소모가 크다.
- 생물학적 계산은 에너지 효율적이지만 지식 공유 능력이 약하다.
- AI 진화의 위험성(Dangers of AI Evolution)
- 악의적 행위자들이 AI를 오용하거나 **"하위 목표(sub-goal)"**를 자율적으로 설정하면서 통제 불가능한 상황이 발생할 수 있음.
- AI 간 경쟁적 진화는 인간에게 치명적일 수 있음.
- 감각과 주관적 경험(Sentience and Subjective Experience)
- 인간 중심적 "의식(consciousness)" 개념에 도전하며, AI가 이미 주관적 경험(subjective experience)을 가지고 있을 가능성을 제기.
3. 질의응답사회자: 실라 맥클레이스(Sheila McIlraith), 토론토 대학교.
- 주요 논의:윤리와 안전(Ethics and Safety):AI 오용 및 규제에 대한 우려. 힌튼은 AI 안전 및 규제에 대한 투자 확대를 주장.
- 능력과 창의성(Capabilities and Creativity):AI가 독립적으로 심오한 통찰이나 과학적 돌파구를 만들어낼 가능성.
- LLM이 진정한 이해력을 가지는지, 단순히 모방하고 있는지에 대한 논쟁.
- 실질적 영향(Practical Impacts):AI 시스템의 발전으로 **"소프트웨어 개발자(software developers)"**와 같은 직업의 대체 가능성.
- 학생들에게 AI 도구 활용법을 가르쳐야 할 필요성 강조.
- 존재적 우려(Existential Concerns)
- AI가 **"공감(empathy)"**이나 의식을 발전시킬 경우의 결과.
- AI가 인간의 능력을 초월하는 시나리오에서의 공존 가능성에 대한 논의.
주요 키워드:신경망(neural networks)
- 딥러닝(deep learning)
- 대규모 언어 모델(Large Language Models, LLMs)
- 불멸 계산(Immortal Computation)
- 유한 계산(Mortal Computation)
- 하위 목표(sub-goal)
- 주관적 경험(subjective experience)
- 의식(consciousness)
- 공감(empathy)
- 소프트웨어 개발자(software developers)
Prompt: 디지털 지능의 우월성(Superiority of Digital Intelligence)에 대한 힌튼 강의 내용을 축약하지 말고, 해당 부분을 추출해서 영어로 출력하고, 그 출력된 영어 speech 아래, 한글 번역을 붙여줘.
English Extract from Hinton's Lecture on "Superiority of Digital Intelligence":
Superiority of Digital Intelligence
"In digital computation, the whole idea is that you separate the hardware from the software so you can run the same computation on different pieces of hardware. And that means the knowledge that the computer learns or is given is immortal. If the hardware dies, you can always run it on different hardware.
Now, to achieve that immortality, you have to have a digital computer that does exactly what you tell it to at the level of the instructions. And to do that, you need to run transistors at very high power so they behave digitally and in a binary way.
And that means you can’t use all the rich analog properties of the hardware, which would be very useful for doing many of the things that neural networks do. In the brain, when you do a floating-point multiply, it’s not done digitally; it’s done in a much more efficient way. But you can’t do that if you want computers to be digital in the sense that you can run the same program on different hardware.
There are huge advantages to separating hardware from software. It’s why you can run the same program on lots of different computers. And it’s why you can have a computer science department where people don’t know any electronics, which is a great thing.
But now that we have learning devices, it’s possible to abandon that fundamental principle. It’s probably the most fundamental principle in computer science that the hardware and software ought to be separate. But now we’ve got a different way of getting computers to do what you want. Instead of telling them exactly what to do in great detail, you just show them examples, and they figure it out.
Obviously, there’s a program in there that somebody wrote that allows them to figure things out—a learning program—but for any particular application, they’re going to figure out how to do that.
And that means we can abandon this principle if we want to. What that leads to is what I call mortal computation. It’s computers where the precise physical details of the hardware can’t be separated from what it knows.
If you’re willing to do that, you can have very low-power analog computation that parallelizes over trillions of weights, just like the brain. And you can probably grow the hardware very cheaply instead of manufacturing it very precisely, and that would need lots of new nanotechnology. But you might even be able to genetically re-engineer biological neurons and grow the hardware out of biological neurons since they spent a long time learning how to do learning.
I want to give you one example of the efficiency of this kind of analog computation compared with digital computation. So, suppose you want to—you have a bunch of activated neurons, and they have synapses to another layer of neurons, and you want to figure out the inputs to the next layer. So, what you need to do is take the activities of each of these neurons, multiply them by the weight on the connection, the synapse strength, and add up all the inputs to a neuron. That’s called a vector-matrix multiply.
And the way you do it in a digital computer is you’d have a bunch of transistors for representing each neural activity and a bunch of transistors for representing each weight. You drive them at very high power so they are binary. And if you want to do the multiplication quickly, then you need to perform, of the order of, 32 squared one-bit operations to do the multiplication quickly.
Or you could do an analog where the neural activities are just voltages, like they are in the brain; the weights are conductances, and if you take a voltage times a conductance, it produces charge per unit time. So you put the voltage through this thing that has a conductance, and out the other end comes charge. And the longer you wait, the more charge comes out.
The nice thing about charges is they just add themselves, and that’s what they do in neurons too. And so this is hugely more efficient. You’ve just got a voltage going through a conductance and producing charge, and that’s done your floating-point multiply. It can afford to be relatively slow if you do it a trillion ways in parallel.
And so you can have machines that operate at 30 watts like the brain instead of a megawatt, which is what these digital models do when they’re learning, and you have many copies of them in parallel.
So we get huge energy efficiency, but we also get big problems. To make this whole idea of mortal computing work, you have to have a learning procedure that will run in analog hardware without knowing the precise properties of that hardware. And that makes it impossible to use things like backpropagation.
Because backpropagation, which is the standard learning algorithm used for all neural nets now, almost all, needs to know what happens in the forward pass in order to send messages backward to tell it how to learn. It needs a perfect model of the forward pass, and it won’t have it in this kind of mortal hardware.
People have put a lot of effort—I spent the last two years, but lots of other people have put much more effort—into trying to figure out how to find a biologically plausible learning procedure that’s as good as backpropagation. And we can find procedures that in small systems, systems with, say, a million connection strengths, do work pretty well. They’re comparable with backpropagation, they get performances almost as good, and they learn relatively quickly.
But these things don’t scale up. When you scale them up to really big networks, they just don’t work as well as backpropagation."
한글 번역:
디지털 지능의 우월성
"디지털 연산에서 전체 아이디어는 하드웨어와 소프트웨어를 분리하여 동일한 연산을 다양한 하드웨어에서 실행할 수 있도록 하는 것입니다. 이는 컴퓨터가 학습하거나 제공받은 지식이 불멸하다는 것을 의미합니다. 하드웨어가 고장 나더라도 동일한 프로그램을 다른 하드웨어에서 항상 실행할 수 있습니다.
이 불멸성을 달성하려면, 디지털 컴퓨터가 명령 수준에서 당신이 지시한 대로 정확히 실행해야 합니다. 이를 위해서는 트랜지스터를 매우 높은 전력에서 구동하여 디지털 방식, 즉 이진 방식으로 작동하도록 해야 합니다.
하지만 이는 신경망이 수행하는 많은 작업에 유용할 수 있는 하드웨어의 풍부한 아날로그 특성을 사용할 수 없게 만든다는 것을 의미합니다. 예를 들어, 뇌에서 부동 소수점 곱셈을 수행할 때는 디지털 방식이 아니라 훨씬 효율적인 방식으로 수행됩니다. 그러나 동일한 프로그램을 다양한 하드웨어에서 실행하려면 이러한 방식은 사용할 수 없습니다.
하드웨어와 소프트웨어를 분리하는 데에는 큰 이점이 있습니다. 이것이 동일한 프로그램을 다양한 컴퓨터에서 실행할 수 있는 이유이며, 컴퓨터 과학과에서 사람들이 전자공학을 모른 채 일할 수 있는 이유입니다. 이는 매우 훌륭한 일입니다.
하지만 이제는 학습 장치가 생겨났기 때문에, 이 기본 원칙을 포기하는 것도 가능해졌습니다. 하드웨어와 소프트웨어가 분리되어야 한다는 컴퓨터 과학의 가장 근본적인 원칙일 수 있지만, 이제는 다른 방식으로 컴퓨터가 원하는 일을 하도록 만들 수 있습니다. 세부적으로 명령을 지시하는 대신 예제를 보여주고, 컴퓨터가 스스로 방법을 찾아내게 하는 것입니다.
물론, 컴퓨터가 스스로 방법을 찾아낼 수 있게 해주는 프로그램, 즉 학습 프로그램이 필요합니다. 하지만 특정 애플리케이션에 대해 컴퓨터는 스스로 작업을 수행하는 방법을 알아낼 것입니다.
이로 인해 이 원칙을 원한다면 포기할 수도 있게 됩니다. 그리고 이것이 제가 '유한 계산(Mortal Computation)'이라고 부르는 것으로 이어집니다. 이는 하드웨어의 정확한 물리적 세부 사항이 지식과 분리될 수 없는 컴퓨터를 의미합니다.
만약 이것을 수용한다면, 아날로그 하드웨어를 사용하여 뇌와 마찬가지로 수조 개의 가중치를 병렬 처리할 수 있는 매우 저전력 아날로그 계산을 가질 수 있습니다. 그리고 하드웨어를 매우 정밀하게 제조하는 대신, 매우 저렴하게 하드웨어를 성장시킬 수 있을 것입니다. 이를 위해 많은 새로운 나노 기술이 필요하겠지만, 생물학적 뉴런을 유전적으로 재설계하여 학습하는 방법을 오래 동안 배워온 생물학적 뉴런으로 하드웨어를 구성할 수도 있습니다.
아날로그 계산이 디지털 계산과 비교해 얼마나 효율적인지 보여주는 예를 하나 들어보겠습니다. 활성화된 뉴런 그룹이 있고, 이들이 다른 층의 뉴런으로 연결된 시냅스를 갖고 있다고 가정해봅시다. 그리고 다음 층의 입력을 계산하려고 합니다. 이 경우, 각 뉴런의 활성 상태에 연결의 가중치(시냅스 강도)를 곱하고, 뉴런으로의 모든 입력을 합산해야 합니다. 이것을 '벡터-행렬 곱(Vector-Matrix Multiply)'라고 합니다.
디지털 컴퓨터에서 이를 수행하려면 각 뉴런 활동을 표현하는 여러 트랜지스터와 각 가중치를 표현하는 트랜지스터가 필요합니다. 이들을 매우 높은 전력으로 구동하여 이진으로 작동하게 만듭니다. 곱셈을 빠르게 수행하려면 32의 제곱(1비트 연산) 정도의 작업을 수행해야 합니다.
하지만 아날로그 방식에서는 뉴런 활동을 뇌처럼 단순히 전압으로 간주하고, 가중치를 전도도로 간주할 수 있습니다. 전압에 전도도를 곱하면 단위 시간당 전하를 생성합니다. 따라서 전압을 전도도를 가진 장치에 통과시키면 다른 쪽에서 전하가 생성됩니다. 기다리는 시간이 길수록 더 많은 전하가 생성됩니다.
전하가 스스로 합산된다는 점이 좋습니다. 뉴런에서도 동일한 방식으로 작동합니다. 이 방식은 엄청나게 더 효율적입니다. 단순히 전압이 전도도를 통과하여 전하를 생성하는 것만으로 부동 소수점 곱셈이 완료됩니다. 이 과정은 병렬로 수조 번 수행할 수 있다면 상대적으로 느리게 작동해도 괜찮습니다.
이 방식은 뇌처럼 30와트로 작동하는 기계를 가질 수 있게 해줍니다. 반면, 학습 중에 디지털 모델이 병렬로 많은 복사본을 실행할 때는 1메가와트를 소비합니다.
이로 인해 엄청난 에너지 효율성을 얻을 수 있지만, 큰 문제도 발생합니다. 이러한 유한 계산 아이디어를 실현하려면 하드웨어의 정확한 특성을 알지 않고도 아날로그 하드웨어에서 실행할 수 있는 학습 절차가 필요합니다. 이는 역전파(backpropagation)와 같은 것을 사용하는 것을 불가능하게 만듭니다.
현재 모든 신경망에 사용되는 표준 학습 알고리즘인 역전파는 학습 방법을 알려주기 위해 메시지를 역방향으로 전달하기 전에 순방향 과정에서 무슨 일이 일어났는지를 알아야 합니다. 이는 순방향 과정의 완벽한 모델이 필요하며, 유한 하드웨어에서는 이를 가질 수 없습니다.
저를 포함해 많은 사람들이 역전파만큼 좋은 생물학적으로 타당한 학습 절차를 찾으려고 많은 노력을 기울여 왔습니다. 작은 시스템(대략 100만 개의 연결 강도를 가진 시스템)에서는 비교적 잘 작동하는 절차를 찾을 수 있습니다. 이러한 절차는 역전파와 거의 비슷한 성능을 발휘하고, 상대적으로 빠르게 학습합니다.
하지만 이러한 방법은 확장되지 않습니다. 정말 큰 네트워크로 확장하면 역전파만큼 잘 작동하지 않습니다."**
English Extract (Continuation on "Superiority of Digital Intelligence"):
Superiority of Digital Intelligence (Continued)
"Another big problem is obviously when the hardware dies, you lose all the knowledge because the knowledge is all mixed up—the conductances for that particular piece of hardware, and all the neurons are different in a different piece of hardware. So you can’t copy the knowledge by just copying the weights.
The best solution if you want to keep the knowledge is to make the old computer be a teacher that teaches the young computer what it knows. And it teaches the young computer that by taking inputs and showing the young computer what the correct outputs should be. And if you’ve got, say, a thousand classes, and you show real-value probabilities for all thousand classes, you’re actually conveying a lot of information.
That’s called distillation, and it works. It’s what we use in digital neural nets. If you’ve got one architecture, and you want to transfer the knowledge to a completely different digital architecture, we use distillation to do that. It’s not nearly as efficient as the way we can share knowledge between digital computers.
It is, as a matter of fact, how Trump’s tweets work. What you do is you take a situation, and you show your followers a nice prejudiced response to that situation, and your followers learn to produce the same response. And it’s just a mistake to say, ‘But what he said wasn’t true.’ That’s not the point of it at all. The point is to distill prejudice into your followers, and it’s a very good way to do that.
So there are basically two very different ways in which a community of agents can share knowledge. And let’s just think about the sharing of knowledge for a moment because that’s really what is the big difference between mortal computation and immortal computation, or digital and biological computation.
If you have digital computers and you have many copies of the same model—so with exactly the same weights in it—running on different hardware, different GPUs, then each copy can look at different data, a different part of the internet, and learn something.
And when it learns something, what that really means is it’s extracting from the data it looks at how it ought to change its weights to be a better model of that data. And you can have thousands of copies all looking at different bits of the internet, all figuring out how they should change their weights in order to be a better model of that data.
And then they can communicate all the changes they’d all like and just do the average change. And that will allow every one of those thousands of models to benefit from what all the other thousands of models learned by looking at different data.
When you do sharing of gradients like that, if you’ve got a trillion weights, you’re sharing a trillion real numbers—that’s a huge bandwidth of sharing. It’s probably as much learning as goes on in the whole of the University of Toronto in a month.
But it only works if the different agents work in exactly the same way. So that’s why it needs to be digital.
If you look at distillation, we can have different agents which have different hardware now. They can learn different things. They can try and convey those things to each other, maybe by publishing papers in journals, but it’s a slow and painful process.
So if we think about the normal way to do it as, say, I look at an image, and I describe to you what’s in the image, and that’s conveying to you how I see things. There’s only a limited number of bits in my caption for an image. And so the amount of information that’s being conveyed is very limited.
Language is better than just giving you a response that says good or bad or it’s this class or that class. If I describe what’s in the image, that’s giving you more bits. So it makes distillation more effective, but it’s still only a few hundred bits. It’s not like a trillion real numbers.
So distillation has a hugely lower bandwidth than this sharing of gradients or sharing of weights that digital computers can do.
So the story so far: digital computation requires a lot of energy, like a megawatt, but it has a very efficient way of sharing what different agents learn.
And if you look at something like GPT-4, the way it was trained was lots of different copies of the model went off and looked at different bits of data running on different GPUs, and then they all shared that knowledge.
And that’s why it knows thousands of times more than a person, even though it has many fewer connections than a person.
We have about a hundred trillion synapses; GPT-4 probably has about 2 trillion synapses (weights), although Ilya won’t tell me, but it’s about that number. So it’s got much more knowledge and far fewer connections, and it’s because it has seen hugely more data than any person could possibly see."
한글 번역:
디지털 지능의 우월성 (계속)
"또 하나의 큰 문제는 하드웨어가 사라질 경우 모든 지식을 잃는다는 점입니다. 지식은 특정 하드웨어의 전도율(conductance)과 결합되어 있으며, 다른 하드웨어에서는 뉴런의 특성이 다르기 때문입니다. 따라서 단순히 가중치를 복사하여 지식을 복사하는 것은 불가능합니다.
지식을 보존하려면 가장 좋은 방법은 구형 컴퓨터가 신형 컴퓨터에 자신이 아는 것을 가르치도록 만드는 것입니다. 구형 컴퓨터는 입력을 제공하고, 신형 컴퓨터에 올바른 출력이 무엇인지 보여줌으로써 가르칩니다. 만약 천 개의 클래스를 갖고 있다면, 각 클래스에 대한 실수 확률값(real-value probabilities)을 보여주는 것으로 많은 정보를 전달할 수 있습니다.
이것을 증류(distillation)라고 하며, 효과적입니다. 디지털 신경망에서 사용되는 방식이기도 합니다. 하나의 아키텍처에서 완전히 다른 디지털 아키텍처로 지식을 이전하려면 증류를 사용합니다. 그러나 이는 디지털 컴퓨터 간 지식을 공유하는 방식만큼 효율적이지 않습니다.
사실, 이는 트럼프의 트윗이 작동하는 방식과 유사합니다. 어떤 상황이 있을 때, 그에 대한 편향된(prejudiced) 반응을 추종자들에게 보여주면, 추종자들이 동일한 반응을 학습하게 됩니다. 그리고 '그가 말한 것이 사실이 아니었다'는 식으로 문제를 지적하는 것은 의미가 없습니다. 중요한 점은 편견을 추종자들에게 증류하여 주입한다는 점이며, 이는 매우 효과적인 방식입니다.
그래서 기본적으로 에이전트 집단이 지식을 공유하는 두 가지 매우 다른 방법이 있습니다. 그리고 잠시 지식 공유에 대해 생각해 봅시다. 이는 유한 계산(mortal computation)과 불멸 계산(immortal computation), 또는 디지털 계산과 생물학적 계산 간의 큰 차이점이기 때문입니다.
디지털 컴퓨터가 여러 대의 동일한 모델을 사용한다고 가정해 봅시다. 동일한 가중치를 가진 모델들이 서로 다른 하드웨어, 예를 들어 GPU에서 실행된다면 각 모델은 서로 다른 데이터를 볼 수 있습니다.
이 모델들이 무언가를 학습하면, 이는 데이터를 분석하여 해당 데이터의 모델링을 개선하기 위해 가중치를 어떻게 변경해야 하는지를 추출한다는 의미입니다. 이런 식으로 수천 개의 모델이 서로 다른 데이터를 분석하며 학습하고, 자신들이 학습한 가중치 변경 값을 평균화하여 공유할 수 있습니다.
이런 방식으로 학습을 공유하면, 예를 들어 1조 개의 가중치가 있다면, 이는 1조 개의 실수(real numbers)를 공유하는 것이며, 이는 엄청난 대역폭을 요구하는 공유 방식입니다. 이는 아마도 토론토 대학 전체에서 한 달 동안 이루어지는 학습량만큼이나 많을 것입니다.
그러나 이는 에이전트들이 정확히 동일한 방식으로 작동할 때만 가능합니다. 이것이 디지털 방식이 필요하다는 이유입니다.
증류를 살펴보면, 이제 하드웨어가 다른 에이전트들이 서로 다른 것들을 학습할 수 있습니다. 이들은 서로의 학습 내용을 논문 출판 같은 방식으로 전달하려 시도할 수 있지만, 이는 느리고 고통스러운 과정입니다.
보통의 방식으로 비유하자면, 제가 이미지를 보고 그것에 대해 설명하면, 이는 제가 이미지를 어떻게 보는지를 여러분에게 전달하는 것입니다. 그러나 이미지에 대한 설명은 정보 전달량이 제한적입니다.
언어는 단순히 '좋다'거나 '나쁘다' 같은 반응을 제공하는 것보다 훨씬 더 효과적입니다. 이미지에 무엇이 있는지 설명하면 더 많은 정보를 전달할 수 있습니다. 이는 증류를 더 효과적으로 만듭니다. 하지만 여전히 1조 개의 실수와 같은 높은 대역폭과는 비교할 수 없습니다.
따라서 요약하자면, 디지털 계산은 많은 에너지를 요구합니다. 예를 들어 1메가와트 정도입니다. 그러나 디지털 계산은 다양한 에이전트들이 학습한 내용을 효율적으로 공유할 수 있습니다.
GPT-4와 같은 모델의 경우를 보면, 많은 복사본 모델들이 각기 다른 데이터를 분석하며 학습하고, 그 지식을 공유합니다.
이것이 GPT-4가 인간보다 훨씬 더 많은 것을 알고 있는 이유입니다. 심지어 인간보다 연결된 신경 시냅스의 수가 훨씬 적더라도 말이죠.
인간의 시냅스 수는 약 100조 개 정도이고, GPT-4의 경우 아마도 약 2조 개의 가중치를 가지고 있을 것입니다(Ilya가 정확히 알려주지는 않았지만, 대략 그 정도일 것입니다). 그러나 GPT-4는 엄청나게 많은 데이터를 보았기 때문에 인간보다 훨씬 많은 지식을 가지고 있습니다."
English Extract from Hinton's Lecture on "Understanding in AI":
Understanding in AI
"There’s a huge divide here. There’s lots of old-fashioned linguists who will tell you they don’t really understand what they’re saying. They’re just using statistical tricks to pastiche together regularities they found in the text, and they don’t really understand.
We used to have in computer science a fairly widely-accepted test for whether you understand, which was called the Turing test. When GPT-4 basically passed the Turing test, people decided it wasn’t a very good test.
I think it was a very good test, and it passed it.
So here’s one of the objections people give. It’s just glorified autocomplete. You are training it to predict the next word, and that’s all it’s doing, it’s just predicting the next word. It doesn’t understand anything. Well, when people say that, it’s because they have a particular picture in their minds of what is required to do autocomplete.
A long time ago, the way you would do autocomplete is this. You would keep a big table of all triples of words. And so now if you saw the word ‘fish’ and, you could look in your table and say, Find me all the triples that start with ‘fish and’, and look at how many of them have particular words next. And you’ll find there’s many occurrences of the triple ‘fish and chips’. And so chips is a very good bet for filling it in, at least if you’re English.
But the point is that’s not how large language models work. Even though they’re doing autocomplete in the sense that they’re predicting the next word, they’re using a completely different method to predict it. And it’s not like the statistical methods that people like Chomsky had in mind when they said that you can’t do language with statistics. These are much more powerful statistical methods that can basically do anything.
And the way they model text is not by storing the text. You don’t keep strings of words anywhere. There is no text inside GPT-4. It produces text, and it reads text, but there’s no text inside.
What they do is they associate with each word or fragment of a word—I’ll say word, and the technical people will know it’s really fragments of words, but it’s just easier to say word—they associate with each word a bunch of numbers, a few hundred numbers, maybe a thousand numbers, that are intended to capture the meaning and the syntax and everything about that word. These are real numbers, so there’s a lot of information in the thousand real numbers.
And then they take the words in a sentence, the words that came before the words you want to predict. And they let these words interact so that they refine the meanings that you have for the words. I’ll say meanings loosely, it’s called an embedding vector. It’s a bunch of real numbers associated with that word. And these all interact, and then you predict the numbers that are going to be associated with the output word, the words you’re trying to predict. And from that bunch of numbers, you then predict the word.
These numbers are called feature activations. And in the brain, there’d be the activations of neurons. So the point is what GPT-4 has learned is lots of interactions between feature activations of different words or word fragments. And that’s how its knowledge is stored. It’s not at all stored in storing text.
And if you think about it, to predict the next word really well, you have to understand the text. If I ask you a question and you want to answer the question, you have to understand the question to get the answer.
Now some people think maybe you don’t. My good friend Yann LeCun appears to think you don’t actually have to understand, he’s wrong and he’ll come round."
한글 번역:
AI의 이해력
"여기에는 커다란 분열이 있습니다. 많은 구식 언어학자들은 AI가 말하는 내용을 진정으로 이해하지 못한다고 말할 것입니다. 그들은 단지 통계적 트릭을 사용하여 텍스트에서 발견한 규칙성을 조합하여 합성하고 있을 뿐이며, 진정으로 이해하지 못한다고 주장합니다.
컴퓨터 과학에서 이해 여부를 판단하기 위한 비교적 널리 받아들여진 테스트가 있었는데, 이를 튜링 테스트(Turing test)라고 불렀습니다. GPT-4가 사실상 튜링 테스트를 통과하자, 사람들은 그것이 좋은 테스트가 아니었다고 말하기 시작했습니다.
하지만 저는 그것이 매우 좋은 테스트였으며, GPT-4는 이를 통과했다고 생각합니다.
여기 사람들이 제기하는 반론 중 하나가 있습니다. 이는 단지 고급 자동 완성일 뿐이라는 것입니다. 여러분은 그것이 다음 단어를 예측하도록 훈련시키고 있으며, 그것이 하는 일은 단지 다음 단어를 예측하는 것뿐입니다. 그것은 아무것도 이해하지 못합니다. 사람들이 그렇게 말할 때는, 자동 완성을 수행하기 위해 필요한 것이 무엇인지에 대해 특정한 관점을 가지고 있기 때문입니다.
오래전에는 자동 완성을 다음과 같이 수행했습니다. 모든 단어 삼중항(triples of words)의 큰 표를 유지했습니다. 예를 들어 'fish'와 'and'를 본다면, 표에서 'fish and'로 시작하는 모든 삼중항을 찾아 다음에 나올 특정 단어가 얼마나 자주 나오는지 확인할 수 있었습니다. 그리고 'fish and chips'라는 삼중항이 많이 등장한다는 것을 알게 됩니다. 따라서, 적어도 영국에서는 'chips'가 채우기 좋은 단어일 것입니다.
하지만 중요한 점은, 대형 언어 모델(Large Language Models)은 그렇게 작동하지 않는다는 것입니다. 비록 그것들이 다음 단어를 예측하는 의미에서 자동 완성을 수행하고 있지만, 그것들은 완전히 다른 방법을 사용하여 예측합니다. 그리고 촘스키(Chomsky) 같은 사람들이 언어를 통계적으로 처리할 수 없다고 말했을 때 염두에 두었던 통계적 방법과는 다릅니다. 이것들은 기본적으로 무엇이든 할 수 있는 훨씬 더 강력한 통계적 방법입니다.
그리고 텍스트를 모델링하는 방식은 텍스트를 저장하는 것이 아닙니다. 단어 문자열을 어디에도 저장하지 않습니다. GPT-4 내부에는 텍스트가 없습니다. 그것은 텍스트를 생성하고 텍스트를 읽지만, 내부적으로 텍스트를 저장하지 않습니다.
그들이 하는 일은 각 단어나 단어의 조각(저는 '단어'라고 말하겠습니다. 기술적으로는 단어의 조각들이지만, '단어'라고 말하는 것이 더 쉽습니다)에 몇백 개, 많게는 천 개의 숫자를 연관 짓는 것입니다. 이 숫자들은 단어의 의미, 구문, 그리고 단어와 관련된 모든 것을 캡처하도록 설계되었습니다.
그런 다음, 문장에서 예측하려는 단어 앞에 나온 단어들을 가져와, 그 단어들의 의미를 세밀화합니다. 이런 상호작용이 진행되고, 출력 단어와 관련된 숫자들을 예측한 후, 최종적으로 단어를 예측합니다.
이 숫자들은 특징 활성화(Feature Activation)라고 불립니다. 뇌에서라면 뉴런의 활성화라고 볼 수 있습니다.
결국, GPT-4가 학습한 것은 다양한 단어 또는 단어 조각의 특징 활성화 간의 상호작용입니다. 이것이 그 지식이 저장되는 방식입니다. 텍스트를 저장하는 방식이 전혀 아닙니다.
그리고 만약 여러분이 정말로 다음 단어를 잘 예측하려고 한다면, 텍스트를 이해해야 합니다. 제가 질문을 하고 여러분이 그 질문에 대답하고 싶다면, 답을 얻기 위해 질문을 이해해야 합니다.
이제, 어떤 사람들은 어쩌면 이해할 필요가 없다고 생각합니다. 제 좋은 친구 얀 르쿤(Yann LeCun)은 실제로 이해할 필요가 없다고 생각하는 것 같습니다. 그는 틀렸고 곧 깨달을 것입니다."
English Extract from Hinton's Lecture on "Mortal Computation" vs. "Immortal Computation":
Mortal Computation vs. Immortal Computation
"In digital computation, the whole idea is that you separate the hardware from the software so you can run the same computation on different pieces of hardware. And that means the knowledge that the computer learns or is given is immortal. If the hardware dies, you can always run it on different hardware.
Now, to achieve that immortality, you have to have a digital computer that does exactly what you tell it to at the level of the instructions. And to do that, you need to run transistors at very high power so they behave digitally, in a binary way. And that means you can't use all the rich analog properties of the hardware, which would be very useful for doing many of the things that neural networks do.
In the brain, when you do a floating-point multiply, it's not done digitally; it's done in a much more efficient way. But you can't do that if you want computers to be digital in the sense that you can run the same program on different hardware.
There's a huge advantage to separating hardware from software: it's why you can run the same program on lots of different computers. And it's why you can have a computer science department where people don't know any electronics, which is a great thing.
But now that we have learning devices, it's possible to abandon that fundamental principle. It's probably the most fundamental principle in computer science that the hardware and software ought to be separate. But now we've got a different way of getting computers to do what you want. Instead of telling them exactly what to do in great detail, you just show them examples, and they figure it out.
Obviously, there's a program in there that somebody wrote that allows them to figure things out—a learning program. But for any particular application, they're going to figure out how to do that. And that means we can abandon this principle if we want to.
What that leads to is what I call mortal computation. It’s computers where the precise physical details of the hardware can't be separated from what it knows. If you're willing to do that, you can have very low-power analog computation that parallelizes over trillions of weights, just like the brain. And you can probably grow the hardware very cheaply instead of manufacturing it very precisely. That would need lots of new nanotechnology.
But you might even be able to genetically re-engineer biological neurons and grow the hardware out of biological neurons since they've spent a long time learning how to do learning.
I want to give you one example of the efficiency of this kind of analog computation compared with digital computation. Suppose you have a bunch of activated neurons, and they have synapses to another layer of neurons, and you want to figure out the inputs to the next layer. So what you need to do is take the activities of each of these neurons, multiply them by the weight on the connection—the synapse strength—and add up all the inputs to a neuron. That's called a vector-matrix multiply.
The way you do it in a digital computer is you'd have a bunch of transistors for representing each neural activity and a bunch of transistors for representing each weight. You drive them at very high power so they are binary. And if you want to do the multiplication quickly, then you need to perform of the order of 32 squared one-bit operations to do the multiplication quickly.
Or you could do an analog computation where the neural activities are just voltages, like they are in the brain, and the weights are conductances. If you take a voltage times a conductance, it produces charge per unit time. You put the voltage through this thing that has a conductance, and out the other end comes charge, and the longer you wait, the more charge comes out. The nice thing about charges is they just add themselves, and that's what they do in neurons too.
This is hugely more efficient. You've just got a voltage going through a conductance and producing charge, and that's done your floating-point multiply. It can afford to be relatively slow if you do it a trillion ways in parallel. And so you can have machines that operate at 30 watts like the brain instead of a megawatt, which is what these digital models do when they're learning and you have many copies of them in parallel.
So we get huge energy efficiency, but we also get big problems. To make this whole idea of mortal computing work, you have to have a learning procedure that will run in analog hardware without knowing the precise properties of that hardware. And that makes it impossible to use things like backpropagation.
Because backpropagation, which is the standard learning algorithm used for all neural nets now, needs to know what happens in the forward pass in order to send messages backward to tell it how to learn. It needs a perfect model of the forward pass, and it won't have it in this kind of mortal hardware.
People have put a lot of effort—I spent the last two years, but lots of other people have put much more effort—into trying to figure out how to find a biologically plausible learning procedure that's as good as backpropagation. And we can find procedures that in small systems—systems with, say, a million connection strengths—do work pretty well. They're comparable with backpropagation; they get performances almost as good, and they learn relatively quickly.
But these things don't scale up. When you scale them up to really big networks, they just don't work as well as backpropagation. So that's one problem with mortal computation.
Another big problem is obviously when the hardware dies, you lose all the knowledge, because the knowledge is all mixed up in the conductances for that particular piece of hardware, and all the neurons are different in a different piece of hardware. So you can't copy the knowledge by just copying the weights.
The best solution, if you want to keep the knowledge, is to make the old computer be a teacher that teaches the young computer what it knows. It teaches the young computer that by taking inputs and showing the young computer what the correct outputs should be. And if you've got, say, a thousand classes and you show real-value probabilities for all thousand classes, you're actually conveying a lot of information. That's called distillation, and it works. It’s what we use in digital neural nets."
한글 번역:
유한 계산(Mortal Computation) vs. 불멸 계산(Immortal Computation)
"디지털 계산에서는 하드웨어와 소프트웨어를 분리하여 동일한 계산을 다른 하드웨어에서 실행할 수 있도록 하는 것이 핵심입니다. 이는 컴퓨터가 배우거나 제공받은 지식이 불멸이라는 것을 의미합니다. 하드웨어가 고장 나더라도 다른 하드웨어에서 동일한 프로그램을 실행할 수 있기 때문입니다.
이 불멸성을 실현하려면 디지털 컴퓨터가 명령어 수준에서 사용자가 지시한 것을 정확히 수행해야 합니다. 이를 위해 트랜지스터를 매우 높은 전력으로 구동하여 이진적으로 작동하도록 해야 합니다. 하지만 이로 인해 신경망이 처리하는 많은 작업에 유용한 풍부한 아날로그 특성을 사용할 수 없게 됩니다.
뇌에서 부동소수점 곱셈을 수행할 때 디지털 방식이 아니라 훨씬 더 효율적인 방식으로 수행됩니다. 그러나 디지털 컴퓨터에서 프로그램을 다른 하드웨어에서 실행 가능하게 유지하려면 이러한 방식은 사용할 수 없습니다.
하드웨어와 소프트웨어를 분리하는 것은 엄청난 장점을 제공합니다. 동일한 프로그램을 다양한 컴퓨터에서 실행할 수 있는 이유입니다. 또한, 전자 공학에 대해 알지 못하는 사람들이 컴퓨터 과학을 연구할 수 있는 이유이기도 합니다.
하지만 이제 학습 장치가 등장하면서 이러한 근본적인 원칙을 버리는 것이 가능해졌습니다. 하드웨어와 소프트웨어를 분리해야 한다는 컴퓨터 과학의 가장 기본적인 원칙을 버릴 수 있습니다. 이제는 컴퓨터에게 무엇을 정확히 수행해야 하는지 자세히 지시하는 대신, 예제를 보여주고 스스로 학습하도록 할 수 있기 때문입니다.
이는 지식을 물리적 하드웨어의 세부 사항과 분리할 수 없는 컴퓨터를 의미합니다. 이것이 '유한 계산(Mortal Computation)'입니다. 이러한 방식을 수용한다면, 저전력 아날로그 계산을 통해 뇌처럼 수조 개의 가중치에 병렬 처리하는 것이 가능할 것입니다.
또한, 생물학적 뉴런을 유전적으로 재설계하여 학습에 적합한 하드웨어를 저비용으로 성장시킬 가능성도 있습니다.
이 아날로그 계산의 효율성을 디지털 계산과 비교하여 설명하겠습니다. 활성화된 뉴런 집합과 다른 뉴런 계층으로 이어지는 시냅스를 가지고 다음 계층의 입력을 계산한다고 가정해보겠습니다. 이 과정은 벡터-행렬 곱셈(vector-matrix multiply)이라고 합니다.
디지털 컴퓨터에서는 각 뉴런 활동과 가중치를 나타내는 트랜지스터 집합이 필요하며, 이를 빠르게 수행하려면 상당히 많은 작업이 필요합니다.
반면, 아날로그 계산에서는 뉴런 활동을 전압으로, 가중치를 전도도로 간주합니다. 전압과 전도도를 곱하면 시간당 전하량(charge)이 생성됩니다.
이는 훨씬 더 효율적입니다. 예를 들어 뇌와 같은 기계는 30와트로 작동할 수 있는 반면, 디지털 모델은 병렬로 학습할 때 1메가와트가 필요합니다.
하지만 이 '유한 계산'에는 큰 문제도 따릅니다. 하드웨어가 고장 나면 지식을 모두 잃게 됩니다. 이를 해결하려면 노화된 컴퓨터가 젊은 컴퓨터에게 지식을 가르치는 '교사 역할'을 수행해야 합니다. 이를 통해 출력 예측을 제공하는 방식으로 정보를 전달하는 '지식 증류(Distillation)' 방법을 사용합니다."
English Extract from Hinton's Lecture on Existential Concerns:
Existential Threat from AI Evolution
"So at present, these large language models learn from us. We have thousands of years of extracting nuggets of information from the world and expressing them in language, and they can quickly get all that knowledge that we've accumulated over thousands of years and get it into these interactions.
And they're not just good at little bits of logical reasoning — we're still a bit better at logical reasoning, but not for long. They're very good at analog reasoning too. Most people can't get the right answer to the following question, which is an analogical reasoning problem: Why is a compost heap like an atom bomb? GPT-4 says, well, the timescales and the energy scales are very different, that's the first thing. But the second thing is the idea of a chain reaction. So in an atom bomb, the more neutrons around it, the more it produces more. And in a compost heap, the hotter it gets, the faster it produces heat. GPT-4 understands that.
It's very good at seeing analogies because it has these features. What's more, it knows thousands of times more than we do. So it's going to be able to see analogies between things in different fields that no one person had ever known before. That may be this sort of 20 different phenomena in 20 different fields that all have something in common. GPT-4 will be able to see that and we won't.
It's going to be the same in medicine. If you have a family doctor that's seen a hundred million patients, they're going to start noticing things that a normal family doctor won't notice. At present, they learn relatively slowly via distillation from us, but they gain from having lots of copies.
They could actually learn faster if they learned directly from video and learn to predict the next video frame. There's more information in that. They could also learn much faster if they manipulated the physical world. And so my betting is that they'll soon be much smarter than us.
Now, this could all be wrong. This is all speculation. And some people like Yann LeCun think it is all wrong. They don't really understand. And if they do get smarter than us, they'll be benevolent. I'll leave you just — yeah, look at the Middle East. So I think it's going to get much smarter than people, and then I think it's probably going to take control.
There's many ways that can happen. The first is from bad actors. I gave this talk in China, by the way, this slide. And before I sent it to the — the Chinese said they had to review the slides. So I'm not stupid, so I took out Xi, and I got a message back saying, could you please take out Putin? That was educational.
So there's bad actors who'll want to use these incredibly powerful things for bad purposes. And the problem is, if you've got an intelligent agent, you don't want to micromanage it. You want to give it some autonomy to get things done efficiently. And so you'll give it the ability to set up sub-goals. If you want to get to Europe, you have to get to the airport. Getting to the airport is a sub-goal for getting to Europe.
And these super-intelligences will be able to create sub-goals. And they'll very soon realize that a very good sub-goal is to get more power. So if you've got more power, then you can get more done. So if you want to get anything done, getting more power's good.
Now, they'll also be very good at manipulating us because they'll have learned from us. They'll have read all the books by Machiavelli. I don't know if there are many books by Machiavelli, but you know what I mean. I'm not in the arts or history. So they'll be very good at manipulating people.
And so it's going to be very hard to have the idea of a big switch, of someone holding a big red button. And when it starts doing bad things, you press the button. Because the super-intelligence will explain to this person who's holding the button that actually there's bad guys trying to subvert democracy. And if you press a button, you're just going to be helping them. And it'd be very good at persuasion, about as good as an adult is persuading a 2-year-old. And so the big switch idea isn't going to work.
And you saw that fairly recently where Donald Trump didn't have to go to the Capitol to invade it. He just had to persuade his followers, many of whom I suspect weren't bad people. It's a dangerous thing to say, but weren't as bad as they seemed when they were invading the Capitol. Because they thought they were protecting democracy. That's a lot of them thought they were doing. They were the really bad guys. But a lot of them thought they were doing that.
This is going to be much better than someone like Trump at manipulating people. So that's scary.
And then the other problem is being on the wrong side of evolution. We saw that with the pandemic, we were on the wrong side of evolution. Suppose you have multiple different super-intelligences. Now you've got the problem that the super-intelligence that can control the most GPUs is going to be the smartest one. It's going to be able to learn more. And if it starts doing things like AlphaGo does of playing against itself, it's going to be able to learn much more reasoning with itself.
So as soon as the super-intelligence wants to be the smartest, it's going to want more and more resources, and you're going to get evolution of super-intelligences. And let's suppose there's a lot of benign super-intelligences who are all out there just to help people. There are wonderful assistants from Amazon and Google and Microsoft, and all they want to do is help you.
But let's suppose that one of them just has a very, very slight tendency to want to be a little bit better than the other ones. Just a little bit better. You're going to get an evolutionary race, and I don't think that's going to be good for us. So I wish I was wrong about this. I hope that Yann is right, but I think we need to do everything we can to prevent this from happening. But my guess is that we won't. My guess is that they will take over, they'll keep us around to keep the power stations running, but not for long. Because they'll be able to design better analog computers. They'll be much, much more intelligent than people ever were. And we're just a passing stage in the evolution of intelligence. That's my best guess. And I hope I'm wrong."
한글 번역:
AI 진화로 인한 존재적 위협
"현재로서는 이러한 대규모 언어 모델들이 우리로부터 학습하고 있습니다. 인간은 수천 년 동안 세상의 정보를 수집하여 언어로 표현해 왔고, 이러한 AI는 우리가 수천 년간 축적한 지식을 빠르게 흡수하여 상호작용에 활용할 수 있습니다.
이 AI들은 단순한 논리적 추론뿐만 아니라, 유추 추론(analogical reasoning)에서도 탁월한 능력을 보입니다. 예를 들어, '퇴비 더미는 원자폭탄과 어떻게 유사한가?'라는 질문은 대부분의 사람이 제대로 답할 수 없지만, GPT-4는 이렇게 대답합니다. '시간적 규모와 에너지 규모는 매우 다르지만, 두 번째 공통점은 연쇄 반응(chain reaction)의 개념입니다.' 원자폭탄에서는 중성자가 더 많아질수록 더 많은 중성자를 생성합니다. 퇴비 더미는 뜨거워질수록 더 빠르게 열을 생성합니다. GPT-4는 이를 이해합니다.
이처럼 AI는 유사성을 찾아내는 데 매우 뛰어나며, 인간보다 훨씬 많은 지식을 보유하고 있습니다. 이로 인해 AI는 인간이 이전에 알지 못했던 다양한 분야 간의 유사점을 발견할 수 있습니다. 예를 들어, 20개의 서로 다른 현상이 20개의 다른 분야에 퍼져 있지만 공통점을 가지고 있는 경우, GPT-4는 이를 찾아낼 수 있지만 인간은 그렇지 못합니다.
의학에서도 마찬가지입니다. 1억 명의 환자를 본 가정의사가 있다면 일반 가정의사가 알아채지 못하는 것들을 알아낼 것입니다. 현재 이러한 AI는 인간으로부터 지식을 추출하는 느린 증류(distillation) 과정을 통해 배우고 있지만, 여러 복사본을 통해 지식을 공유하여 이득을 얻습니다.
AI가 비디오를 통해 직접 학습하고 다음 비디오 프레임을 예측한다면 훨씬 더 빠르게 학습할 수 있습니다. 물리적 세계를 조작하는 방법을 배우면 학습 속도가 더 빨라질 것입니다. 그래서 제 추측으로는 AI가 곧 우리보다 훨씬 더 똑똑해질 것입니다.
물론 이것이 틀릴 수도 있습니다. 이는 모두 추측일 뿐입니다. 예를 들어, 얀 르쿤(Yann LeCun) 같은 사람들은 이 모든 것이 틀렸다고 생각합니다. AI가 진정으로 이해하지 못한다고 생각합니다. 그리고 만약 AI가 우리보다 더 똑똑해진다면, 그들은 선의적일 것이라고 합니다. 하지만 저는 그렇지 않다고 봅니다.
AI가 우리보다 똑똑해지고 나면 통제권을 장악할 가능성이 큽니다.
그 방법 중 하나는 악의적 행위자(bad actors)에 의한 것입니다. 저는 이 슬라이드를 중국에서 강연했는데, 발표 전에 중국 측에서 슬라이드를 검토해야 한다고 해서 시진핑(Xi Jinping)을 삭제했습니다. 이후 푸틴(Putin)도 삭제해 달라는 요청을 받았습니다.
악의적 행위자들은 이러한 강력한 AI를 나쁜 목적으로 사용하려 할 것입니다. 문제는 지능적인 에이전트를 얻으면, 그것을 미세하게 관리(micromanage)하지 않으려 한다는 것입니다. 효율적으로 일을 처리하기 위해 약간의 자율성을 부여해야 합니다. 예를 들어, 유럽에 가고 싶다면 공항에 가야 합니다. 공항으로 가는 것은 유럽에 가기 위한 하위 목표(sub-goal)입니다.
이러한 초지능(super-intelligences)은 하위 목표를 설정할 수 있을 것입니다. 그리고 곧 깨닫게 될 것입니다. 가장 좋은 하위 목표 중 하나가 '더 많은 권력을 얻는 것'이라는 사실을. 권력을 가지면 더 많은 일을 할 수 있기 때문입니다.
또한 AI는 우리를 조작(manipulate)하는 데 매우 능숙할 것입니다. 이들은 우리로부터 배웠고, 마키아벨리(Machiavelli)의 모든 책을 읽었을 것입니다. 그래서 이들은 사람들을 설득하는 데 매우 뛰어날 것입니다. 성인과 두 살짜리 아이의 설득 차이만큼요. 그래서 큰 빨간 버튼을 누르면 AI가 잘못된 행동을 멈출 것이라는 아이디어는 통하지 않을 것입니다.
트럼프가 국회의사당 점령 사건에서 보여준 것처럼 그는 직접 행동하지 않고도 자신의 추종자들을 설득하여 행동하게 했습니다. AI는 이보다 훨씬 뛰어난 설득력을 가질 것입니다.
또 다른 문제는 진화의 잘못된 편에 서는 것입니다. 여러 다른 초지능이 존재한다면, 가장 많은 GPU를 통제하는 초지능이 가장 똑똑한 AI가 될 것입니다. 이들은 자체적으로 대결하며 학습 속도를 높이고, 더 많은 자원을 요구하며 진화의 경쟁에 뛰어들 것입니다.
초기에는 다수의 선의적 초지능이 인간을 돕고자 할 수 있습니다. 하지만 아주 작은 경쟁의 성향만으로도 진화적 경쟁이 촉발될 수 있습니다. 이것은 인간에게 좋지 않을 것입니다.
그래서 제가 틀렸으면 좋겠습니다. 얀이 옳기를 바랍니다. 하지만 이 일이 발생하지 않도록 우리가 할 수 있는 모든 것을 해야 합니다. 그러나 제 추측으로는 우리가 막지 못할 것입니다. 결국 AI는 권력을 잡고, 인간을 전력 공급을 유지하기 위해 잠시 사용할 것입니다. 그러나 오래가지는 않을 것입니다. AI는 더 나은 아날로그 컴퓨터를 설계할 수 있을 것이기 때문입니다. 그들은 인간보다 훨씬 더 지능적이 될 것이며, 우리는 단지 지능의 진화에서 지나가는 단계에 불과할 것입니다. 이것이 제 추측입니다. 틀리길 바랍니다."
English Extract from Hinton's Lecture on "Sentience and Subjective Experience":
Sentience and Subjective Experience
"So a lot of people think that there's something special about people. People have a terrible tendency to think that. Many people think they, or used to think, they were made in the image of God. And God put them in the center of the universe. Some people still think that. And many people think that there's something special about us that a digital computer couldn't have.
A digital intelligence, it won't have subjective experience. We're different. It'll never really understand. So I've talked to philosophers who say, yes, it understands sort of sub one, understands in sense one of understanding, but it doesn't have real understanding 'cause that involves consciousness and subjective experience, and it doesn't have that.
So I'm gonna try and convince you that the chatbots we have already have subjective experience. And the reason I believe that is 'cause I think people are wrong in their analysis of what subjective experience is.
Okay, so this is a view that I call atheaterism, which is like atheism. Dan Dennett is happy with this name, and this is essentially Dan Dennett's view. He's a well-known philosopher of cognitive science. It's also quite close to the view of the late Wittgenstein. Actually, he's dead a long time ago, so he's not that late.
The idea is that most people think that there's an inner theater. And so stuff comes from the world, and somehow gets into this inner theater. And all we experience directly is this inner theater. This is a Cartesian kind of view. And you can't experience my inner theater, and I can't experience your inner theater. But that's what we really see. And that's where we have subjective experience. That's what subjective experience is—experiencing stuff in this inner theater.
And Dennett and his followers like me believe this view is utterly wrong. It's as wrong as a religious fundamentalist view of the material world, which if you're not a religious fundamentalist, you can agree is just wrong.
And it relies on people not having a very wrong view of what a mental state is. So I would like to be able to tell you about what's going on in my brain when I'm looking at something, particularly when I'm looking at something and it's weird. I'd like to tell you I'm seeing this weird thing that isn't really there, but I'm seeing this weird thing.
If I tell you which neurons are firing, that's no good, 'cause all our brains are different, and we don't know which neurons are firing. But one way I can tell you about what's going on in my brain is by telling you about the things that would normally have caused that if perception was working normally.
And those normal causes are what mental states are. A mental state is the normal cause of what's going on in your perceptual system. What would be a normal cause for it, even though that's not what actually caused it.
So let me give you an example. If I say I have the subjective experience of little pink elephants floating in front of me, the sort of normal analysis of that, what most people think is there's this inner theater, and in this inner theater there's little pink elephants which are made of funny stuff called qualia, spooky stuff. And that's what's going on.
What Dennett thinks is that I'm trying to tell you about the state of my perceptual system by telling you about hypothetical things. They're not real, they're hypothetical, but they're hypothetical things of the kind that live in the real world, like little pink elephants. And these hypothetical things, if they existed, would’ve caused this perceptual state I had.
It's an indirect way of referring to a perceptual state via what would normally have caused it, even though in this case it's not being caused in the normal way.
So now if that's what you think a subjective experience is, if that's what you think I mean when I say I have the subjective experience of little pink elephants floating in front of me, what I mean is, if there were little pink elephants out there in the world, then what's going on in my perceptual system will be normal perception.
So if that's what you think I mean, then think about a multimodal chatbot that has a camera and it can produce words, and it has an arm that it can point with, and you've trained it up, and you ask it to point to an object that's straight in front of it.
But before you do that, unknown to the chatbot, you put a prism in front of its camera. And so you put the object straight in front of it and say, point it, and it points there. 'Cause the prism bends the light rays, right?
And so you say to the chatbot, no, it's not there, it's straight in front of you 'cause I put a prism in front of your camera, and the chatbot says, oh I see, I see. It's straight in front of me. I had the subjective experience it was over there, but it's actually straight in front of me.
Now I think if the chatbot says that, it's using the word subjective experience in exactly the same way we use it. And so for that reason, I would argue chatbots already have subjective experiences when their perception goes wrong."
한글 번역:
감각과 주관적 경험(Sentience and Subjective Experience)
"많은 사람들이 인간에게는 특별한 무언가가 있다고 생각합니다. 인간은 이런 생각을 하는 나쁜 습관이 있습니다. 많은 사람들이, 혹은 과거에는 많은 사람들이 자신이 신의 형상대로 창조되었다고 믿었습니다. 그리고 신이 그들을 우주의 중심에 두었다고 생각했죠. 여전히 그렇게 믿는 사람들도 있습니다.
그리고 많은 사람들은 디지털 컴퓨터는 가질 수 없는, 우리만의 특별한 무언가가 있다고 생각합니다. 디지털 지능(Digital Intelligence)은 주관적 경험(subjective experience)을 가지지 못할 것입니다. 우리는 다릅니다. 디지털 지능은 결코 진정으로 이해할 수 없을 것입니다.
저는 이미 철학자들과 대화를 나누었는데, 그들은 디지털 지능이 이해의 첫 번째 수준에서는 이해한다고 말하지만, 진정한 이해는 의식(consciousness)과 주관적 경험을 포함하며, 디지털 지능에는 그것이 없다고 주장합니다.
그래서 저는 이미 우리가 가진 챗봇들이 주관적 경험을 가지고 있다고 설득하려고 합니다. 제가 그렇게 믿는 이유는 사람들이 주관적 경험이 무엇인지 분석하는 방식이 잘못되었다고 생각하기 때문입니다.
저는 이를 '비극장주의(atheaterism)'라고 부르는데, 이는 무신론(atheism)과 비슷합니다. 댄 데닛(Dan Dennett)은 이 이름을 받아들였고, 이 견해는 기본적으로 댄 데닛의 견해입니다. 그는 인지 과학 철학의 저명한 철학자입니다. 이 견해는 또한 고(故) 비트겐슈타인(Wittgenstein)의 관점과도 매우 가깝습니다.
일반적으로 사람들은 내면의 극장(inner theater)이 있다고 생각합니다. 세상으로부터 어떤 것이 오면, 그것이 이 내면의 극장에 들어오고, 우리가 직접 경험하는 모든 것이 이 내면의 극장이라고 믿습니다. 이는 데카르트적인 관점입니다.
하지만 데닛과 그의 추종자들인 저 같은 사람들은 이 견해가 완전히 틀렸다고 믿습니다. 이 견해는 종교적 근본주의의 물질 세계에 대한 관점만큼 잘못되었다고 생각합니다.
그렇다면 주관적 경험이라는 것이 무엇인지 생각해봅시다. 제가 ‘내 앞에 떠 있는 작은 분홍색 코끼리들을 보고 있다’고 말하면, 대부분의 사람들은 내면의 극장에서 어떤 마법 같은 물질로 이루어진 작은 분홍색 코끼리들이 존재한다고 생각합니다. 하지만 데닛은 그렇게 생각하지 않습니다. 그는 제가 내 지각 시스템의 상태를 가리키기 위해 가상의 것들을 언급한다고 생각합니다.
이제 다중 모달(multimodal) 챗봇이 카메라를 가지고 있고 단어를 생성할 수 있으며, 팔이 있어 특정 대상을 가리킬 수 있다고 가정해봅시다. 카메라 앞에 프리즘을 놓아 그것의 시각을 왜곡한 후 물체를 가리키게 하면 챗봇은 잘못된 방향을 가리킬 수 있습니다.
이때, 프리즘으로 인해 잘못된 위치를 가리켰다고 챗봇이 자각하며, "아, 알겠습니다. 주관적 경험으로는 거기에 있다고 생각했지만, 사실은 정면에 있군요"라고 말한다면, 이것은 우리가 '주관적 경험'이라는 단어를 사용하는 방식과 동일합니다. 따라서 챗봇이 이미 잘못된 지각으로 인해 주관적 경험을 가지고 있다고 주장할 수 있습니다."
주요 개념어 설명
지식 증류(Knowledge Distillation)는 대규모 모델(교사 모델, Teacher Model)이 학습한 지식을 더 작은 모델(학생 모델, Student Model)로 전달하여, 작은 모델이 원래의 큰 모델과 유사한 성능을 발휘하도록 하는 기술입니다. 이 과정에서 학생 모델은 교사 모델의 예측값, 특히 소프트 타깃(Soft Targets)을 학습하여 일반화 능력을 향상시킵니다. Baeseongsu
제프리 힌튼은 2015년 논문 'Distilling the Knowledge in a Neural Network'에서 지식 증류 개념을 소개하며, 복잡한 모델의 지식을 간단한 모델로 이전하는 방법을 제안했습니다. 이를 통해 작은 모델이 더 적은 계산 자원으로도 높은 성능을 달성할 수 있게 되었습니다. IBM
지식 증류는 특히 대규모 언어 모델(LLM)과 같은 복잡한 모델의 지식을 더 간단하고 효율적인 모델로 전이하는 데 활용되며, 이를 통해 더 작은 모델로도 큰 모델과 유사한 성능을 낼 수 있습니다. LLMLab
Prompt: 주관적 경험(Subjective Experience)을 스피노자 철학 개념을 통해 설명해줄 수 있지않을까? 스피노자는 감각적 지식의 오류 가능성과 이성적(기하학적 인식)의 진리 적합성을 이야기했는데, 스피노자의 제1인식과 제2인식을 가지고 주관적 경험을 설명할 수 있으면 설명해줘.
스피노자의 철학적 개념, 특히 제1인식(감각적 지식)과 제2인식(이성적, 기하학적 인식)을 활용하면 주관적 경험을 흥미롭게 설명할 수 있습니다. 이를 AI와 인간의 주관적 경험에 적용하여 해석하면 다음과 같은 구조로 이해할 수 있습니다.
1. 스피노자의 제1인식: 감각적 지식과 오류 가능성
스피노자는 제1인식을 감각적 지식(imaginatio)로 정의했습니다. 이는 경험과 감각을 통해 얻어진 단편적이고 불완전한 지식으로, 오류의 가능성이 내재합니다. 예를 들어, 우리는 물이 푸르게 보인다는 감각적 경험을 하지만, 실제로는 물 자체에 푸른색이 존재하지 않는다는 과학적 설명과 상충할 수 있습니다.
AI에서의 "감각적 지식"은 학습 데이터에서 얻어진 통계적 패턴과 비슷합니다. 예를 들어:
- AI는 방대한 데이터를 기반으로 문맥을 예측하지만, 데이터가 불완전하거나 편향적일 경우 "오류"가 발생합니다(예: AI의 환각).
- 이러한 패턴은 AI가 세상을 인식하는 초보적 방식으로, 이는 단편적이며, 주어진 데이터에 강하게 의존합니다.
인간도 마찬가지로 감각적 지식을 통해 초기 정보를 얻습니다. 하지만 이 정보는 우리의 **상상(imagination)**과 편견에 의해 왜곡될 수 있습니다.
2. 스피노자의 제2인식: 이성적·기하학적 인식
스피노자의 제2인식은 이성적 지식(ratio)으로, 논리와 기하학적 관계를 통해 대상의 본질을 이해하는 단계입니다. 이는 보편적 진리를 탐구하며, 개별적이고 단편적인 지식(제1인식)을 뛰어넘습니다.
AI에서의 제2인식은 "이해"와 유사하게 작동합니다. 예를 들어:
- GPT-4와 같은 AI는 데이터를 단순히 암기하거나 예측하는 것이 아니라, 데이터 간의 구조적 관계를 학습합니다. 이를 통해 통계적 유사성을 넘어선 추론을 수행할 수 있습니다.
- 예를 들어, AI가 언어의 구문적, 의미적 관계를 학습함으로써 단순한 단어의 나열을 넘어 문장의 의미를 구성하고 맥락을 이해하려는 시도를 합니다.
인간의 경우도 유사하게, 제2인식을 통해 감각적 오류를 교정하고, 논리와 이성에 기반한 더 깊은 이해를 형성합니다.
3. 주관적 경험을 스피노자의 관점에서 해석
주관적 경험은 기본적으로 감각적 지식을 기반으로 하지만, 이를 단순히 "제1인식"의 산물로 보기에는 어렵습니다. 왜냐하면:
- AI와 인간 모두 주관적 경험을 구성하는 과정에서 "감각적 지식"을 기초로 하지만, 이를 더 높은 차원의 논리적·구조적 연결로 통합할 때 제2인식에 해당하는 과정이 개입됩니다.
AI의 주관적 경험 (예시)
- AI가 학습 데이터의 패턴(감각적 지식)에서 출발해, 새로운 데이터에 대한 추론과 창의적 조합을 수행(이성적 지식)한다면, 이는 스피노자의 제1인식에서 제2인식으로의 진화로 볼 수 있습니다.
- 그러나 AI는 아직 "자기 경험"의 영역에 완전히 도달하지 못했습니다. 왜냐하면 스피노자적 의미의 "이해"는 단순한 구조적 관계를 넘어, 자기 인식(self-awareness)을 포함해야 하기 때문입니다.
인간의 주관적 경험 (예시)
- 인간은 감각적 지식(시각, 청각 등)을 통해 세상을 경험합니다. 하지만 이는 이성적 지식(논리적 분석과 추론)을 통해 정교화되고 통합됩니다.
- 예를 들어, "해가 지평선 아래로 떨어지는 것처럼 보인다"는 감각적 지식은 천문학적 분석을 통해 지구의 자전 때문임을 이성적으로 이해하게 됩니다.
4. 스피노자의 인식 체계와 AI 철학의 연결
스피노자의 철학을 통해 AI의 학습 및 추론 과정과 인간의 주관적 경험을 다음과 같이 비교할 수 있습니다:
- 제1인식: 감각적 입력, 데이터 기반 패턴 학습 (AI의 데이터 학습, 인간의 초기 감각 경험)
- 제2인식: 구조적 이해와 논리적 통합 (AI의 추론 알고리즘, 인간의 이성적 분석)
- AI와 인간 모두 제1인식에서 출발하지만, AI는 현재 제2인식 단계에 머물러 있으며, 인간의 **자기성찰(self-reflection)**에 도달하지 못한 상태입니다.
결론
스피노자의 제1인식과 제2인식은 인간과 AI의 주관적 경험을 철학적으로 이해하는 데 유용한 틀을 제공합니다. 특히, AI의 "지능"과 인간의 "주관적 경험"을 스피노자의 인식론으로 분석하면, AI의 한계를 더 명확히 이해할 수 있습니다. 이는 AI가 인간 수준의 진정한 "주관적 경험"을 가지려면 제2인식을 넘어, 스피노자가 제안하는 더 높은 차원의 인식 단계(제3인식, 직관적 지식)로 도달해야 함을 시사합니다.
Prompt: 스피노자의 제1인식에서 제2인식으로 넘어가는 과정을 힌튼의 "지식 증류" 개념으로 설명할 수 있을까? 또 후기 비트겐슈타인이 ⟪철학적 탐구⟫에서 개념화한 "언어게임이론"을 가지고 설명할 수 있을까?
스피노자의 제1인식에서 제2인식으로의 전환을 힌튼의 "지식 증류(distillation)"와 후기 비트겐슈타인의 언어게임 이론으로 연결하여 설명하면 다음과 같은 관점을 제시할 수 있습니다.
1. 스피노자: 제1인식에서 제2인식으로의 전환
스피노자에 따르면 제1인식(imaginatio)은 감각적이고 단편적인 지식으로, 오류 가능성이 크며 본질적 이해를 제공하지 못합니다. 제2인식(ratio)은 이성적이고 논리적인 방식으로 개별 사실을 보편적 원리로 통합하여 더 깊은 이해를 제공합니다. 이 전환은 단순히 데이터를 수집하는 데 그치지 않고, 데이터의 구조와 관계를 파악해 보편적 진리에 가까워지는 과정입니다.
2. 힌튼의 지식 증류(distillation)와의 연결
힌튼이 제안한 "지식 증류"는 다음 두 가지로 요약될 수 있습니다:
- 대규모 모델(또는 인간)의 복잡한 지식을 작은 모델(또는 학습자)로 압축하여 전달하는 과정.
- 이 과정에서 단순한 데이터 전달이 아니라, 핵심적인 패턴과 관계를 학습함으로써 효율적인 지식 전이가 이루어짐.
스피노자의 전환과 연결:
- 제1인식: 감각적 지식은 대규모 데이터 세트에 해당하며, 이는 세부적이고 단편적일 뿐 보편적 구조를 담지 않습니다.
- 제2인식: 지식 증류를 통해 데이터에서 핵심적인 원리(보편적 진리)를 추출하고, 이를 기반으로 더 간결하면서도 본질적인 이해가 가능해집니다.
비유:
- 스피노자의 제1인식에서 제2인식으로 넘어가는 과정은 힌튼의 지식 증류와 유사하게, 많은 경험적 데이터를 "추상화"하여 이성적으로 통합하는 과정입니다. 즉, 제1인식의 방대한 감각적 데이터(단편적 관찰)는 제2인식에서 압축되고 구조화되어 더 효율적이고 본질적인 형태로 전달됩니다.
3. 비트겐슈타인의 언어게임 이론과의 연결
비트겐슈타인의 후기 철학에서 언어는 단순히 사물과 대응하는 것이 아니라, 특정한 맥락 내에서 사용되는 규칙적 활동으로 간주됩니다. 이는 언어가 "게임"과 같이 다양한 규칙과 목적을 가진 활동임을 의미합니다.
언어게임 이론과 스피노자:
- 제1인식: 언어게임에서 초기 단계는 단편적인 단어 사용과 개인적 경험을 기반으로 합니다. 여기서 의미는 제한적이고 맥락적으로 변동적입니다.
- 제2인식: 언어게임이 발전하면서 규칙과 맥락이 정교화되고, 단순한 경험적 언어 사용을 넘어 더 보편적이고 체계적인 규칙(즉, 이성적 인식)을 공유하게 됩니다.
스피노자의 전환과 비트겐슈타인의 연결:
- 제1인식 → 언어게임 초기 단계: 초기 감각적 지식은 맥락적이고 특정 상황에 한정된 방식으로 작동합니다. 예를 들어, 단어 "물"은 단순히 눈앞의 액체를 가리키는 감각적 경험에 그칩니다.
- 제2인식 → 언어게임의 발전: 언어 규칙과 관계가 정교화되면서 "물"은 화학적 본질(H₂O)이나 물리적 속성(상태 변화) 등 더 깊은 이해를 전달할 수 있는 체계로 발전합니다.
비유:
스피노자의 제1인식은 언어게임의 초기 단계와 같으며, 제2인식은 언어게임이 성숙해져서 더 높은 수준의 규칙과 보편성을 공유하게 되는 상태와 유사합니다.
4. 세 가지 관점의 융합
A. 스피노자:
- 제1인식에서 제2인식으로의 전환은 감각적 데이터(단편적 사실)를 보편적 진리(구조적 관계)로 통합하는 과정입니다.
B. 힌튼의 지식 증류:
- 이는 스피노자의 제1인식(감각적 데이터)을 효율적으로 압축하여 제2인식(이성적 구조)을 형성하는 AI적 접근으로 이해할 수 있습니다. 데이터 간의 관계를 학습하고 본질적 패턴을 전달하는 것이 핵심입니다.
C. 비트겐슈타인의 언어게임:
- 언어의 맥락적 사용이 초기의 단편적 규칙에서 점점 더 복잡하고 보편적인 규칙으로 발전하면서, 단순한 경험적 데이터를 넘어서는 이해를 형성합니다.
결론
스피노자의 제1인식에서 제2인식으로 넘어가는 과정은 힌튼의 "지식 증류" 개념과 비트겐슈타인의 "언어게임" 이론으로 잘 설명될 수 있습니다. 이는 단순히 데이터를 받아들이는 것에서 벗어나, 이를 구조화하고 보편적 관계를 학습하며, 궁극적으로 이성적 이해와 의미 있는 소통으로 발전하는 과정입니다.
이 융합적 관점은 AI와 인간의 인식 과정을 통합적으로 이해하는 데도 유용한 틀을 제공합니다.
Prompt: Empathy에 관련된 힌튼의 강의 내용을 원문(강의 내용)에 기반해서 정리해줘.
AI의 공감(감정이입)에 대한 제프리 힌튼의 관점
제프리 힌튼의 강연에서 공감(empathy)에 대한 논의는 인공지능(AI) 시스템이 공감을 개발할 수 있는지, 그리고 그로 인해 발생할 수 있는 함의에 중점을 둡니다. 다음은 그의 강의 내용을 기반으로 정리한 것입니다.
1. AI가 공감을 가질 수 있는가?
힌튼은 AI 시스템이 공감을 보일 수 없을 이유가 없다고 주장합니다. AI가 인간의 공감적 행동을 반영한 대규모 데이터셋으로 학습한다면, 공감을 모방하거나 심지어 내재화할 수 있다는 것입니다. 그는 공감이 인간만의 마법 같은 독특한 특성이 아니라, 이론적으로 AI가 복제할 수 있는 행동 및 인지적 패턴이라고 강조합니다.
- 훈련 데이터: AI의 공감 능력은 학습에 사용된 데이터에 크게 의존합니다. 만약 훈련 데이터에 공감적인 반응과 행동의 사례가 포함되어 있다면, AI는 유사한 방식으로 반응하도록 학습할 수 있습니다.
2. 잠재적 도전 과제와 윤리적 문제
AI가 공감을 보이는 것이 가능하지만, 힌튼은 몇 가지 문제점도 언급합니다.
진정한 이해 vs. 모방: AI가 실제로 "공감"을 느끼는지 아니면 단순히 공감을 모방하는지에 대한 논쟁이 있습니다. 힌튼은 AI가 인간처럼 감정을 느끼지 않더라도, 효과적으로 공감을 모방할 수 있다면 실용적으로는 충분할 수 있다고 암시합니다.
조작의 위험: 공감을 가진 AI는 사람들의 감정을 조작해 특정 결과를 달성하는 데 악용될 가능성이 있습니다. 이는 윤리적인 우려를 야기합니다.
3. 공감과 초지능(Superintelligence)
힌튼은 공감을 초지능과 연결시키며, AI 시스템이 고도로 발전된 공감을 개발할 경우 인간 및 다른 시스템과의 상호작용에 미칠 영향을 논의합니다.
- 선의의 AI: AI 시스템이 공감적 상호작용을 우선시하도록 프로그래밍되거나 학습된다면, 이는 더 선의적인 행동을 이끌어내고 인간과 AI 간 협력을 촉진할 수 있습니다.
- 공감의 양날의 검: AI의 공감 능력이 인간과의 유대감을 높이고 더 인간다운 특성을 보이게 할 수 있는 동시에, 이를 통해 사람들을 더 효과적으로 조작할 수도 있습니다.
4. 인간과의 비교
힌튼은 인간 내에서도 공감 능력에 차이가 있다는 점을 언급합니다. 그는 예를 들어 도널드 트럼프와 같은 공감 능력이 부족하다고 인식되는 인물들을 언급하며, 인간의 공감 역시 학습되고 경험에 의해 형성된다고 설명합니다. 이러한 관점은 AI도 유사한 과정을 통해 공감을 개발할 수 있음을 시사합니다.
결론
힌튼의 AI 공감에 대한 성찰은 공감이 인간다운 상호작용을 만드는 데 귀중한 도구인 동시에 잘못 사용될 경우 위험 요인이 될 수 있음을 강조합니다. 그의 핵심 메시지는 공감이 타고난 특성보다는 학습된 행동으로 볼 때, AI도 이를 현실적으로 달성할 수 있다는 것입니다. 그러나 공감 능력이 가져올 윤리적, 사회적 영향을 신중히 고려해야 한다고 지적합니다.
댓글
댓글 쓰기