본문 바로가기

교차검증 프로젝트: CROSS+CHECK/과학의 오해 [Theory of Everything] + ORIGINS

AI,인공지능에 대한 오해 (1부: Science Fiction과 Fact 구분하기)

1부: 공상과학과 과학의 구분

 

0. 들어가는 말:
혁신? 거품? 아니면 종말의 전조인가?  

 
GPT는 Generative Pre-trained Transformer 의 약자입니다. 

Generative는 'ㅇㅇ을 생성generate 가능한' 이란 뜻과 언어학 용어이기도 합니다.
Pre-trained는 "사전 훈련된"이란 뜻,
Transformer는자연언어 처리에 사용되는 신경망(neural network)의 아키텍쳐 종류 입니다. 

 
처음 ChatGPT(챗GPT)라는 단어를 들은 것은 유튜브 영상에서 조던 피터슨의 강연 속에서였습니다.
따로 구독하고 있는 채널은 아니었는데 조던 피터슨 강의를 들어서인지 알고리듬의 추천에 있었나봅니다. 

챗GPT에 대한 거기에 대한 여러 의견들 찾아보았습니다.
 
흥미로운 발견은 AI에 대한 유튜브 영상들을 검색해봤을 때, 각광받는 신기술은 평가라기보단 홍보에 가까운 내용이 많았고,  대부분이 "챗GPT로 돈 버는 법" 같이 이윤추구적이거나,  SF(공상과학)적이거나 디스토피아적인 내용이 대부분이었습니다.

반대로 AI에 대한 비평을 검색했을 땐, 비교할 수 없이 낮은 조회수를 보이는 영상들이 대부분이었습니다.
(주제상, 유튜브 알고리듬에게 미운 털이 박혔나봅니다.) 

그렇기 때문에 이 인공지능에 대해 둘러쌓인 환상(혹은 거품)과 오해를 주제로 측면에서 접근해보려 합니다.
 
이미 우리 생활 속에 깊숙이 자리 잡은 AI, 인공지능이라는 것에 대해 균형잡힌 시선을 잡는 것에 도움이 되는 글이 되길 바라며 이번 글을 작성합니다. (각 분야의 전문가들과의 인터뷰 내용을 인용 후, 마지막에 제 생각을 넣는 식으로 진행하려 합니다.)
 

(1) 개인적인 감상

저도 여러 실험을 해보았습니다. 하나의 문장 혹은 문단이 툭 튀어나는 것이 아니라, 문장이 생성되는 과정이 흥미로웠고, 그게 사람들에게 경이로움을 느끼게 하는 인터페이스라고 생각됐습니다. 
 
플리글롯(polyglot)으로서 당연히 사용할 줄 아는 외국어들을 사용하여 번역 퀄리티도 테스트 해봤죠. 
일단 기본은 한다는 인상을 줬습니다만, 번역을 시키면 속도가 많이 느려지는 현상이 있더군요. 
엔드유저, 일반 사용자로서의 첫 인상은 "작문이 너무 싫은 공대생이 만들어낸 훌륭한 작문도구...?"였습니다. 
 

또 다른 한국 유튜브 영상에서 보게된 인공지능 회사의 솔트룩스(SaltLux) 대표가 설명하는 챗GPT의 단점과 한계에 대해서도 들어본 후 직접 실험도 해보았습니다.

요약: 논리와 상식의 부족. 할머니의 성별이 여성이다-라는 결론을 도출하지 못합니다. 그게 단순히 미국 측 데이터를 기반으로 학습된 녀석(?)이라 정치적 올바름(policital correctness)를 신경써서 그런 거려나.. 추측만 해봤습니다.

 

 
 
AI를 마주하는 시선에는 여러가지가 있을 것 같습니다.
 
사람처럼 글을 써내는 인공지능을 보며 SF영화나 공상과학소설에서처럼 인공지능이 인간을 지배하는 세상이 오는 걸 두려워하는 사람, 혹은 좀 더 현실적으로 AI로 인해 일자리를 잃을까 두려워하는 사람, 또는 유발 하라리와 같이 인공지능이 인간의 한계를 극복하는데 사용되어 인간이 AI의 노예가 되는 암울한 미래를 그리는 사람도 있습니다. 

개인적으로는 친구와 챗GPT에 대해 이야기하며 인공지능의 시대에 아이들을 어떻게 가르쳐야하나- 이런 얘기도 나눠봤구요. 
 
어디까지가 팩트에 기반한 예측이고 어디까지가 허상이자 망상일까요? 
 
AI에 대한 오해에 대해 알아보려 합니다. 
 

(2) 인용자료와 관련하여 

비전공자의 얕은 감상보다는 역시 여러 분야의 전문가들의 의견을 참고해야 읽을 가치가 있는 글이 될 거라 생각합니다.
좀 더 깊이 있는 비평을 나눠보고자 약 2주 간의 시간을 들여 5개의 인터뷰 영상을 보고 번역했습니다.

이번엔 특별히 비평의 대상인  챗GPT(ChatGPT)에게 현재의 인공지능에 대한 비평에 대해 변론을 할 수 있는 기회도 줘봤습니다. 
 
우선 이번 글에 인용된 저자들을 먼저 간략히 소개합니다. 
 

(A) TEAM HUMAN : 인간계

요약>

인터뷰 인용대상: 노암 촘스키, 개리 마커스, 로버트 J. 마크스, 존 레녹스

 
※ 제가 이번 주제에 대해 조사하게 되면서 알게 된 두 사람에 대해서는 소개 자료를 아래에 추가합니다. 

(1) Robert J. Marks [로버트 J. 마크스 ] - 구글 스콜라 링크 

컴퓨터 과학계에서 잔뼈가 굵은 교수님. 
1968년에 설립된 IEEE[아이트리플이], 미국 전기공학회(Institute of Electrical and Electronical Engineers)의 초대 회장을 역임했을 정도이고, 미국 베일러(Baylor) 대학의 특훈교수(distinguished professor)이니 기본 개념 설명은 이 분께 맡기려 합니다 ㅎㅎ 
 
(2) Gary Marcus [개리 마커스] Linked In 프로필

어디에 나올 때마다 "최연소 명예교수" 라는 타이틀이 언급되는데.. 컴퓨터 과학계 수재인가봅니다. 8세부터 프로그래밍하고 MIT들어가서 연구하다가 뉴욕대학교(NYU)에서 교수하다가 AI 로봇회사 창업하고, 나중에 그 회사가 우버(uber)에 인수되고 하네요.  현재 인공지능에 대한 비평의 목소리 중 가장 드러나는 분인데, 8살 때 처음 프로그래밍을 접한 시점부터 인공지능에 매료되어 MIT를 거쳐 젊은 나이에 (은퇴한) 명예교수가 되었죠. 
 
국내에 번역된 책은 클루지 (Kluge)와 2029 기계가 멈추는 날 (Rebooting AI: Building Artificial Intelligence We Can Trust), 마음이 태어나는 곳 ( the birth of mind) 등이 있습니다. 

 

(B) TEAM MACHINE: OPEN AI

 
챗GPT와 깊은(?) 대화를 나눠본 결과, 챗GPT는 솔직합니다. 자신의 한계에 대해서 거짓말 하는 대신 꽤나 상세하게 설명하고 있는 것으로 보이기에 비평가들의 의견에 대해 챗GPT에게도 변론할 기회를 줘봤는데 꽤 도움이 되는 답변을 생성해냈습니다. 


1. AI 인공지능에 대한 오해 풀기 

전문분야 외의 영역에 대해 알아볼 때 중요한 건 기초지식과 용어의 정의라고 생각합니다. 
기초개념을 모르면 전문분야 설명을 아무리 들어도 오해만 늘어날 수 있으니깐요. 
 
이 부분에 대한 이해 없이 신기술에 대해서는 활용법으로 넘어가는 게 대세이지만, 혹시라도 이런 기초 내용이 필요하신 분이 있을 거라 생각되어 각 인터뷰 내용을 나눕니다. 
 
기초적인 오해를 푸는 것에 도움이 될 Q&A (Q&R) 내용을 시작하겠습니다. 


 

Q. AI/인공지능은 무엇인가요?  어디서 볼 수 있나요? 
What is artifical intelligence? and where do we find it ?

1) 로버트 마크스(Robert J. Marks) 교수의 답변: 

Robert J. Marks (2022) 출처: Discovery Institue

인공지능은 사전에 따라 의미가 달라질 수 있습니다. 
만약 학계에 가면 학파에 따라 AI가 뭔지 다투는 걸 볼 수 있죠. 
AI와 유사어로 취급되는 것들은 Computational Intelligence [컴퓨테이셔널 인텔리전스] (계산 가능한 지능), Machine Intelligence [머신 인텔리전스] (기계 지능) 등이 있고, 각 분야에서는 이 구분에 큰 의미를 두고 있기 때문에 하위범주(Subcategory)를 나누곤 합니다. 
 
우리가 얘기하는 주제를 고려했을 때, 언론 미디어에서 사용하는 의미로 취급하면 좋을 것 같습니다. 
만약 컴퓨터를 통해 할 수 있는 어떤 대단한 것, 컴퓨터가 무언가 대단해 해보이는 걸 해냈을 때 경이롭게 느끼는 그런 것이요. (인공지능을) 그렇게 표현했을 때 우리가 인공지능이 할 수 있는 것에 대해 얘기할 때, 컴퓨터가 할 수 있는 것에 대해 얘기할 수 있고, 만약에 컴퓨터가 할 수 없는 것은 AI도 할 수 없는 거죠. 
 

[영어 인터뷰 원문 접음처리]

더보기

Artifical Intelligence depends on your dictionary.
 If you go into Academia, go into different discipline, you'll find skirmishes as what AI means. 
There's kind of synonyms which are used for AI such as Computational Intelligence, Machine Intelligence, and to those in the field this means a lot, they do subcategories and things of that sort. 

I think, for purposes for our discussion. We should probably treat it just like the media treats it, which in any GEE WHIZ things that can be done by a computer, so if a computer does something and it's just miraculous, and you go Gee-Whiz, then that's probably AI. And I like it for purpose of the discussion. Because what it does is it brings AI to the level of the computer, so when we talk about the capabilities of AI, we can talk about the capabilities of computers, and if a computer can't do something, then AI doesn't have the capabilities to do either. 

2) 존 레녹스(John C. Lennox) 교수의 답변 :

존 레녹스 교수

인공지능은 컴퓨터 분야에서 아주 급속히 발전하고 있는 분야 입니다. 
그건 Narrow AI(or weak AI/ '좁은 인공지능 혹은 특수AI)와 AGI, Artificial General Intelligence두 종류로 나눌 수 있죠. 이 두 개를 구분하는 것은 아주 중요합니다. 
 

(a) Narrow AI (Weak AI) 

먼저 Narrow AI (Weak AI) 를 봅시다.
이건 우리 모두에게 이미 익숙한 겁니다. 실제로 작동하는 것이구요.
보편적으로 AI 시스템은 굉장히 강력한 컴퓨터와 방대한 데이터베이스, 데이터베이스 내의 특정 패턴을 인지하는 알고리듬으로 구성되어 있습니다. 
 
예를 들면, 코로나 대유형의 시점에서 이 바이러스는 사람들의 폐에 영향을 미치죠. 
100만 개의 감염된 사람들의 폐 엑스레이 사진으로 구성된 데이터베이스를 상상해봅시다. 
각 사진들은 세계 저명한 의학자들을 통해 라벨링/판독/구분되구요. 
 
그리고 저의 폐 X-ray를 찍습니다. 만약에 제가 코로나19에 걸렸다면, 알고리듬이 제 폐 엑스레이 사진과 데이터베이스 속의 백만 개 엑스레이와 비교하고, 굉장히 빨리 진단이 나오겠죠. 
요즘엔 이런 진단이 제 동네 병원보다 정확할 수 있습니다. 
이 기술은 코로나19의 백신을 개발한다거나 하는 여러가지 훌륭한 일에 사용됩니다. 
이런 AI에는 긍정적인 면이 있죠. 

[영어 인터뷰 원문 접음처리]

더보기

Artificial Intelligence very rapidly developing field of computation. It splits into two types, there's Narrow AI, and there's AGI-Artifical General Intelligence. It is very important to keep them seperately. 
So let's look at narrow A.I. which wer are all familiar. it's the stuff that actually works. The typical AI system consists of a very powerful computer with a large database and an algorithm that recognizes particular patterns within that data base. 

Take a simple example. In the midst of Covid pandemic , at the moment, and it affects people's lungs, so let's imagine our database consists of one million x-ays, of diseased lungs. They are labeld with their respective diseases by the top medical people in the world. 
That's our database. 
Now a photograph is taken of my lungs. If I catch, COVID 19, then the algorithm compares photograph that's taken my lungs and with the million in the database. and comes up very rappidly with the dignosis. 
now thesedays, that dignosis is likely to be much more acurate than I would get, from my local hospital. 
and systems like this are being employed to find vaccines for covid19, and many other wonderful things. 
so there's a very positive side. to this kind of Ai. 

 
두번째로 언급해야 할 것은, 이런 시스템은 특성상, 한 가지 (일)만 할 수 있습니다.
주로 인간의 지능이 필요한 것들이죠. 
 
그래서 인공지능(人工知能/Artificial Intelligence)이라고 불리는 겁니다. 
이건 진짜 지능이 아닙니다. 
 
컴퓨터는 생각하지 않습니다. 의식이 없죠. 
인풋(input)된 지성, 의사들, 컴퓨터 과학자들과 알고리듬을 개발한 사람들이 있죠. 
 
나온 진단을 봤을 때, 마치 인간의 지성을 통해 만들어진 것처럼 보이지만 그렇지 않아요.
이건 정교한 기계를 통해 만들어 진 겁니다. 이걸 알고 있는 것은 아주 중요합니다. 
 
 

[영어 인터뷰 원문 접음처리]

더보기

[원문] The second thing to mention abut is, that characteristically, such a system, do one single thing. that normally requires human intelligence. 
that's why it's called artifical intelligence. 

it's not real intelligence. 
the machinery, the computer doesn't think. it's not conscious. 
there's intelligent input. the doctors, and the computer scientists and the people that developed the algorithms.

what comes out look as if, the dignosis, looks as if, it's been made by human intelligence. but it hasn't. it's been made by this sophisticated piece of equipment. it's very important to grasp that. 

 
또 패턴 인식의 요즘 아주 자주 사용되는 다른 예를 들면, 안면인식 기술이 아주 고도로 발전되었죠. 
이게 어떻게 유용하게 쓰일 수 있을 지 볼 수 있습니다. 
경찰은 군중 속의 테러리스트나 범죄자를 찾아내서 사회에 위험인 사람들을 잡을 수 있습니다. 

출처: CNBC


하지만 부정적인 면도 있죠. 
이 안면인식기술이 특정 국가에서는 개인 뿐만 아니라 특정 그룹의 사생활을 침해하고, 통제하고 억압하는데 사용되고 있죠.
 
이 예를 든 이유는 Narrow AI (특수AI)는 거대한 도덕적 질문을 발생시키고 있다는 걸 말하기 위해서 입니다. 
 

더보기

(2) Now another example is, we'll stick with pattern recognition, which is enormously popular. facial recognition technology have developed. to a very advanced state. and you can see they are very useful. for police forces to pick out a terrorist or criminal in a crowd. and be used therefore to capture people who are a danger to society. but there's downside. because facial recognition technology is being used at the moment in certain parts of the world, to invade the privacy, not only of individuals but of whole people groups. and actually control them and suppress them.
now, i mentioned that example to say that very rapidly, ai, narrow ai raises huge ethical questions. 

 
이건 실제로 작동하는 것들입니다.
"자율주행"차에도 AI시스템이 탑재되어있죠.

출처: Tesla hacker unlocks Autopilot ‘Augmented Vision’ and it’s awesome

 
하지만 어떤 도덕적 의사결정을 설계해야 합니다. 
 
만약 차의 센서가 어느 노인이 길을 건너고 있는 걸 "보고", 만약 그 노인을 피하면 버스를 기다리기 위해 줄 서 있는 아이들을 치게 할 것이란 걸 계산했다면, 이 차는 어떤 선택을 할까요? 
 
그런 도덕성을 프로그램에 넣어야 합니다. 
물론, 그런 도덕적 규범은 프로그래밍하는 사람들의 도덕이 되겠죠. 
 
그게 큰 의문들을 낳죠. 자율무기 (autonomous weapon)의 경우엔 더 그렇구요. 

더보기

(3) Now, remember this is the stuff that's actually working. 
self-driving cars, autonomous vehicles. AI system built in there, but you have to build into some kind of ethical decision making. 

If he car sensors pick up an old lady crossing the road, and it can see, so to speak, if it misses her, it'll hit a queue of children waiting in line for a bus, what does it decide to do?

so you've got to program that ethically. 
and of course, the ethical norms. will be built in, are the ethics of the people doing the programming.

that raises huge questions even more difficult is the matter of autonomous weapon. so on and so forth. 

 
마지막 예로는 우리 모두가 사용하는 스마트폰입니다. 
이건 우리를 추적하고 있죠. 계속. 
다수가 실감하지 못하고 있는 것 중 하나가, 우리가 아마존에서 뭔가를 사면, 그 후에 "이 책을 산 사람들은 이런 책들도 샀습니다" 라는 팝업을 보게 되죠. (광고들은) 우리가 다른 걸 사도록 부추깁니다. 
네, 그럴 수 있죠. 
 
하지만 우리가 실감하지 못하는 것은 우리가 이 시스템에 엄청난 정보를 제공하고 있고, 대기업은 제3자에게 우리 동의 없이 이걸 팔고 있죠. 이건 아주 큰 문제가 되고 있어요. 
 
쇼샤나 주보프(Shoshana Zuboff)가 여기에 전문가이죠. 

사진출처: TheGuardian.com

그녀가 쓴 책 감시 자본주의 시대 (The Age of Surveillance Capitalism) 에서 이걸 다루고 있죠. 
우린 그런 의미에서 감시 받고 있고, 이건 경제적으로 굉장히 수익성이 있는 건데, 우리는 수수료도 없이 여기에 기여하고 있죠. 그것도 자발적으로요. 
이건 언젠가 저희들에게 한 방 먹일 겁니다. 우리는 이제서야 점차 이 문제에 대해 보고 있죠. 
 

(원문접음)

더보기

Now the final example I would give you is is the fact that we're all involved in ai that is, any of us who own a smartphone, it's tracking us. all the time. 
what many of us don't realize is. that for example, when we make a purchase on Amazon, few days later, we'll get pop-ups saying people bought this book  were in that book. they're trying to encourage you to buy something else, okay. we put up with that. but what we don't often realize is we're feeding a vast amount of information into that system. and large companies are selling that onto third parties without our permission.and this has become a real problem. and it is really is called by Susanna Zubov who's an expert on it.
and has written a book talking about it called Survelliance Capitalism. so we have survelliance in that sense.  which is economically vastly profitable. we're contributing without our commission. and we're doing this voluntarily. 
and this is eventually going to his us really hard, and we begin to see that there are problems with it. 

 
제가 말하고 싶은 건 AI는 저희 삶 속에 남아 있을 거라는 겁니다. 아직 그렇지 않다면 점점 더 급속도로 그렇게 되겠죠. 연간 1조 달러의 사업입니다. 그리고 좋은 영향도 있지만 모든 방향에서 여러 문제를 야기하고 있습니다. ( 이하 생략) 


AGI (Artificial General Intelligence) 인공 일반/보편 지능  

 
반대로 AGI(Artificial General Intelligence)는 기본적으로 말하자면 '슈퍼 인텔리전스', Super Intelligence,초(超)지능,입니다.  
이건 두 부분이 있습니다. 
첫번째로는 현 인류의 지능을 개선하는 것에 대한 아이디어 입니다. 
사이보그 기술 혹은 약물을 통한 바이오 엔지니어링(bio-engineering)과 같이 인간을 기반으로 한 '초(超)지능'을 만들어 내려는 생각이죠. 
 
이 분야에 대해 가장 잘 알려진 사람 중 하나는 과학자가 아닌 유발 하라리 (Yuval Noah Harari)입니다. 

Official Website - Yuval Noah Harari

Prof. Yuval Noah Harari is a historian, philosopher and best-selling author of 'Sapiens' and 'Homo Deus'. Discover his ideas, writing and lectures.

www.ynharari.com

 
그의 책 <호모데우스 Homo Deus> (직역: 신인神人)에서 그는 이런 트랜스휴머니스트의 아젠다를 달성하려 하죠. 

호모데우스 (김영사)

 
트랜스휴머니즘 (Transhumanism)에는 두 가지 부분이 있습니다. 
 
첫 번째는 인간의 죽음은 기술로 해결될 것이라는 것이죠. 
인공지능과 관련된 기술의 주된 노력은 인간의 행복을 증진시키기 위해서이고, 우리 인류가 '신', 소문자g의 god, 그리스 신화 속의 신과 같은 존재가 되는 거죠. '초 지능(super intelligence)'를 통해서요. 
 
또 다른 부분은 인간의 지능을 유기체로부터 독립된 실리콘(칩)으로 옮겨서 복제하려는 시도입니다.
그게 어떤  사람들이 자기 뇌를 실리콘(칩)에 업로드해서 평생 살아갈 거라는 생각을 하게 되는 맥락입니다. 
또 어떤 사람들은 그런 날이 오길 바라며 사망시점에 자기 뇌를 냉동하려는 시도를 하는 것에 대해 알고 있습니다. 언젠가 자기 뇌의 내용물/콘텐츠가 실리콘에 담길 거라는 희망이 깔려있죠. 

더보기

AGI. ON the other hand, basically represented by the quest for super intelligence. 

that has two parts. firstly, the idea of enhancing current human intelligence. By, bio-eginnering, by cyborg technologies, by drugs. and all kinds of things to create super-intelligence-based starting with human being. 

that's the agenda of one of the most known people on this topic is not a scientist, Yuval Nora harari, in his book, Homo Deus, the man who is god, and what he is trying to do. is to achieve tranhumanist agenda. which really is two prompts(?): 
First, the problem of human death will be solved by technology, fairly soon. 

and therefore, the main effort of artifical intelligence and related techonlogoy. will be to enhance human happiness. so we turn human beings into gods-small g, greek style gods, having super intelligence. 

Now, the other, research, is on trying to replicate the human intelligence onto silicone, to make it independant of organic materials. and it's in that kind of context that you get notions of people hoping that their brains will one day be, capable of being uploaded unto silicon and so they live forever. 

and we're all aware of people who go to great expense to have their brain frozen on the death, so they hope one day they'll be woken up, and their brain contents will be put on the silicon. 

자, 이 부분은 굉장한 추측에 불가합니다. 과학보다는 공상과학이 훨씬 더 많이 가미 되어있죠. 
하지만 마틴 리스 경 (lord Martin Rees)와 같은 일부 과학자들은 몇 백년 안에 우리가 인류의 속성을 바꿀 거라고 생각합니다.

영국 황실 천문학자, 존재 위협 연구소 공동창립자 Existential Risk

 
그리고 그 신인류들은 지금의 우리에 대해 모를 거라는 거죠.
그건 그저 추측을 낳을 뿐만 아니라 큰 걱정거리가 되죠. 일론 머스크(Elon Musk)가 이걸 걱정합니다. 
또 어떤 이들은 이런 걸 개발하는 것이 악마를 소환하는 거라고 말하기도 하죠.
 
그래서 AGI는 주로 공상과학/SF/사이언스 픽션의 내용입니다. 
그들은 특이점(Singularity), 빅뱅의 기원이 되는 그 특이점이 아닌 - 인간이 기술과 융합되는 특이점, 즉 기술이 (인류를) 근본적으로 인류를 장악하게 되는 시점이 온다고 말합니다. (이하 생략)
 

더보기

Now, all of that is highly-speculative. There's much more science-fiction than science. However, there are scientists, like leading scientists, like our own astronomer royal, lord, Reece, who thinks that we are going to change the nature of human being. so that within the few hundred years. that being exist would have no knowledge of us. really, at all, no idea, what we are like, and that has created not only speculation, but great concern. Elon Musk is worried about it. Somebody else said, in developing this kind of stuff, we're summoning the demon. So AGI is largely the stuff of science-fiction. 

They tell us there's a Singularity coming. That's Ray Kurzweil who is a guru in this field -- this is not the singularity of origin of Big Bang, 
The singularity is when human beings become merged with technology, and technology essentially take over. I see a few problem of this. but you asked me what it is. and that is what it is. 

3) Gary Marcus 의 답변: 

 

AI(인공지능)에 대해 잘못된 정보가 많이 있습니다.

또 전 AI가 잘못된 정보를 생성해내는 것에 대해 더 걱정이 됩니다. 
인지 과학에 대한 배경이 있어야 무슨 일이 일어나고 있는지 이해할 수 있을 것 같습니다. 
제가 볼 땐, 굉장한 열풍이 돌고 있는 것 같아요. 
만약 인지(cognition)이 뭔지 모른다면, 이 열풍에 휩쓸리기 쉬울 것 같습니다. 
우리가 어느 부분에서 발전하고 있고, 어느 부분에서 그렇지 못한 지 구분하는 게 중요하다고 생각합니다. 
 
만약 여러분이 그저 AI는 마술같은 거라고 생각한다면 한 영역에서 발전이 있으니 당연히 다른 쪽에서도 마찬가지라고 생각하기 쉬운 거죠. 
 
유명한 실패 사례로는 IBM이 있죠.


"재퍼디(Jepardy!)"에서 우승한 훌륭한 AI가 있으니 우리가 생각할 수 있는 가장 어려운 문제를 해결시켜보자! 
종양학을 해결해보자. 대표 John Kelly가 TV프로그램 찰리 로즈Charlie Rose 에 나왔었고, 제대로된 실적이 없었기 때문에 그 후 26분기 동안 계속 하락했죠.
 
어떻게 보면 이건 인공 지능에 대한 문제를 하나 해결했으니 모든 걸 다 해결할 수 있다고 생각한 자만이었죠. 
현실은 다릅니다. 
인식/인지(cognition/認識)의 모든 부분은 다릅니다. 
 
비지니스에서 해결해야 하는 문제들도 다르구요. 
 
전 아이들이 어떻게 학습하는 지, 인공지능이 어떻게 뭘 학습하는지를 연구한 사람으로서 의미있는 관점을 가지고 있다고 생각합니다. 그리고 전 권력에 대해 사실을 말하는 걸 두려워하지 않고 헛소리를 헛소리라고 부를 수 있는 사람이라고 생각합니다. 그래서 어떤 영역에선 인기가 있고, 어떤 영역에선 인기가 없죠. 아마 그래서 제가 이 자리에 나오게 된 것 같습니다. 

더보기
There's a lot of misinformation about A.I., I am concerned about A.I creating misinformation.
I think that one needs to have background in cognitive science behind to really appreciate what's going on. 
I think there's a lot of hype out there. 
and it's easy, if you don't understand what cognition is, to get caught up on the excitement.
It's important to distinguish between where-we-are in making progress, where-we-are-not making progress. 
If you just think sort of A.I is magic, then when you think there's a progress in one place, you automatically assume, there is other.
One of the famous mistake is IBM thought "Okay, we bit Jeopardy!. We have this amazing  piece of A.I, let's throw this into the hardest problem we can't think of. Oncology(종양학). and it will all be great, John Kelly on Charlie Rose that we're gonna change the world, after that the stock tanks 26 quoters in a row. Because they never delivered goods. 
In some way this was hubris, around thinking that, 
"because I have solved one AI problem, so I can solve them all"
the reality is, every piece of cognition is different. every problem you want to solve in business is different. 

so I'ld like to think that I have some perspective there, as someone who studied both how children learn about things, and how AI learns about things. And I also like to think that I am not afraid to speak truth to powers and call bullshit.  that makes me I am popular in some quarters, not populars in others. Maybe that's why I am here today. 

 
인공지능이란 단어는 (인공적이지 않은) 자연적인 지능을 떠오르게 합니다. 인간이나 꿀벌과 같이요.
그렇게 생각해보면 그 "지능"이란 것에는 많은 것이 포함되어 있습니다. 
그저 하나가 아닙니다. 
 
만약 지능에 대한 하나의 정의만 적용해야 한다면, "새로운 문제를 해결하기 위해 적응할 수 있는 능력"이라고 하겠지만, 그건 큰 도움이 되지 않을 것 같네요.
 
지능에는 여러가지 면이 있습니다. IQ 테스트란 걸 생각하면 한 가지 숫자를 떠올리죠. 
저 사람은 112, 다른 사람은 72.. 마치 그게 저울에서의 수치인 것 처럼요. 
 
하지만 현실을 보면 지능에 여러 종류가 있습니다. 
언어지능, 수학지능, 미적지능(심미지각), 또 저에겐 없는 훌륭한 농구선수가 지니고 있을 지능 (Basketball IQ)... 등이 있죠. 

Howard Gardner의 이론이 말하는 8가지 종류의 지능 - 공간, 음악, 언어, 논리수학, 관계, 내면적, 자연패턴(출처:CNBC)

 
우리는 그 중 일부를 기계에 설계할 줄 아는 것이 있죠. 어떤 경우에는 인간이 하는 것을 모방하고, 어떤 경우엔 그렇지 않죠. 
 
좋은 예는 "체스를 잘 두는 것" 입니다. 
체스를 잘 두는 기계(프로그램)은 이미 있죠. 

체스 로봇과 체스를 두다가 로봇에 의해 손가락이 부러진 소년에 대한 기사

지능의 또 다른 특징으론 대화가 있죠. 일관성 있고 조리있는 대화 능력이요. 
하지만 그걸 해낼 수 있는 기계는 없습니다. 
그걸 가짜로 흉내내는 기계는 있죠. (we have machines that fake it.)
하지만 금방 그들이 흉내내고 있다는 걸 알아차릴 수 있습니다.
그들이 어디에 있는 지 물어보거나, 오늘이 무슨 요일인지 .. 
잠시동안 천진한 사람이나 멍청한 사람을 속일 수 있을 지 모르지겠지만, 전문가는 속일 수 없습니다. 
 
제 생각에 인간 지능의 가장 큰 특징은 아주 어린 아이도 새로운 것들을 아주 빨리 배울 수 있다는 겁니다. 
여덟살 아이에게 새로운 게임을 하자고 묻고, 규칙을 알려주면 3분만 지나도 아이들은 그 게임을 할 줄 알게 되죠. 세계 최고 수준은 아니겠지만 설명을 들으면 보드겍임이든 카드게임이든 운동과 같이 몸으로 하는 놀이이든 규칙을 배울 수 있습니다. 그런 걸 할 수 있는 기계는 없습니다. 
 
또 다른 특징은 유연성이 있죠. 여러 특징 중에 하나에 불과하죠. 
 
현재 기계가 어떤 일을 할 수 있는 지 알고 싶을 때, 세부사항/디테일에 대해 알아봐야 합니다. 
그냥 "이젠 기계들이 스마트해져서, 뭔가 할 수 있게 됐다" 라고 말할 수 없습니다. 어떤 분야를 얘기하는 지 봐야하죠. 아직 가정에서 쓸 수 있는 (제대로 된) 로봇은 없습니다. (이하 생략)

더보기

So, A.I. means Artificial Intelligence, so that makes you think of natural intelligence, like what we humans have, what honey bees have, and already, once you think about that way.  you realize there's a lot of thing that are involved in intelligence. It's not one thing. If I had one definition, it might be 'the ability to adapt when we solve new problems" or something. but that would be staking the deck. There's lots of facets on intelligence. it's not one dimensional variable.  if you think about IQ test, innitially makes you think one number. This guy's IQ is 112, this other guy is 72 and whatever, as if it's one number on a scale. 
But, the reality is, we have different kind of intelligence. We have verbal intelligence, mathmatical intelligence, aesthetic intelligence, whatever intelligence great basketball player has - i don't have that. 
we do know how to build some of that intelligence into machines, sometimes we do it by copying like humans do. sometimes, we don't. 

Good example where I would say intelligence is playing chess well.  we have machines that can play chess very well.And other things that are hallmark of intlligence is having conversation, coherent converation.  you and I are doing that.  And no machine that do that right now. 
We have machines that fake it. Very quickly you'll realize they are faking it - where they are, what day it is, ...  so for a little while, they can fool a stupid person, or a naive person. But they can't fool an expert. So they're not good at whatever is that goes into conversation. 
I think the biggest hallmark of human intelligence, is even a child can learn a lot of new things very quickly. So you take a 8-year-old child, and you say, I bought a new game at the store, you want to play? what are the rules? threeminutes later, they're playing that game.  
maybe not a a world class level, they pick up the rules by verbal instruction. how to play this board came or card game. or how to something physical. some new sport or whatever. we don't have machines that can do that. 
machines we have takes A LOT of trainning to do whatever it is they do. often mesaured in mega bytes and giga bytes. sometimes taking months of training. what have you. There's kind of flexibility is a hallmark, A hallmark, not the only hallmark, that machines don't have. 
When you wanna know can a machine do this thing now, you have to dive into details, you can't just say "they're smart now. They can do stuff. " 
Depends on what you're talking about. 
They're not smart enough to be domestic robots now.

웹서밋에서 노암 촘스키와 인터뷰에 참여하는 개리 마커스

2. A.I의 미래에 대한 예측과 현재 A.I의 한계

 
Q. 컴퓨터 처리 기술의 급속한 발전이 이런 기술이 (인류를) 장악/침공할 것이라는 생각에 기여하나요?
Does the rapid growth of technology in computer processing contribute to these ideas of technology taking over? 
 
[존 레녹스 교수 답변] : 

 
계속 증가하는 컴퓨터 능력에 대해 18,24개월마다 2배로 성장한 '무어의 법칙'이 있죠. 하지만 거기엔 한계가 있을 겁니다. 하지만 우리가 이 모든 문제를 해결 할 수 있을 거라는 그 생각에 대해서 전 큰 의문이 있습니다. 
 
이 분야의 심오한 문제 중 하나는 '지성intelligence'과 '의식consciousness'의 연결고리 입니다. 
전 AI/인공지능은 artificial, 인공적이라고 말씀드렸습니다. 
 
여기에 대한 선구자로는 조셉 멜리챔프 (McRae Mellichamp) 교수의 논문 "The Artificial in artificial intelligence is real (직역: 인공지능의 "인공"은 진짜다/실재한다)" 이 있습니다. 아주 훌륭한 제목입니다. 풀어서 말하면 인공지능Artificial Intelligence은 정말 인공적(인위적Aritifical)이라는 거죠. 
 
그래서 어떤 이들은 이렇게 주장합니다. 
"실리콘 위에 의식을 개발할 필요는 없고, 인간 지성을 모사/복제해내면 돼. "
다른 말로하면 튜링 테스트(Turing Test)를 통과하면 돼. 그저 다른 사람들에게 혹은 심문하는 시스템에게 사람처럼 보이면 되고, 인간과 동일할 필요는 없다는 거죠. 

출처: Researchgate.net

 
 
여기에 대해 버클리 대학 존 써얼(John Sealre)교수는 어떤 시스템이 지성/지능(intelligence /知能)이 없이 지성/지능이 있는 것처럼 보일 수 있다고 주장합니다.
 

 
유명한 "The Chinese Room Experiment (중국어 방 실험) " 이죠. (혹은 "The Chinese Room Argument")
 

더보기

[원문] Computer power is increasing rapidly and there's Moore's Law that says it double roughly, every 18 months or so. There'll be eventual limit to that. 
But the idea will be able to go so far that we'll be able to solve all these problems. 

But i have very big questions about that. 
one of the deep questions in this field is the connection between intelligence and consciousness. 
You see, I said earlier, that artificial intelligence is aritificial. 

There's wonderful paper written by one of the pioneers, to the topic. 
Dr., professor Melanie Mitchell, who said, this is the title of his paper-i think it's marvellous- The Artificial in aritifial intelligence is real. In other words, it really is artificial. 

 

Therefore some argue, "look, we don't need to develop consciousness.say, based on silicon, all we need to do, is replicate human intelligence. in other words, get something to pass the Turing Test. 
that it appears to other people interrogating the system that it is human but it needn't be identical at all.

 

 And John Searle in Berkley, pointed out that a system, could appear to be, intelligent, without being intelligent at all. 

The his famouse Chinese Room experiment.

 

Q. 존 써얼의 '중국어 방 실험'  : 컴퓨터가 무언가를 이해할 수 있나요?
Can computers understand? 

Robert J. Marks 사진 출처 : mindswithmedved.com

[로버트 마크스 교수 답변] 

컴퓨터들은 자기가 뭘하는 지 이해하지 못합니다.
컴퓨터는 6과 48을 더할 수있지만, 숫자 6이 뭔지 숫자 48이 뭔지 모릅니다. 
 
컴퓨터가 이해력이 없다는 것을 설명하는 훌륭한 예시가 있습니다. 
존 써얼(John Searle)은 수십년 전에 Searle's Chinese Room 의 예를 제시했죠. 

Illustration of John Searle's Chinese Room Argument (출처:towarddatatscience.com)

이 예시 안에서 써얼(Searle)이 중국어로 된 질문과 답변이 담겨 있는 숫자로 표기된 파일 캐비넷이 가득한 폐쇄된 방에 있습니다.
 
써얼은 중국어를 쓸 줄도 읽을 줄도 모릅니다. 하지만 이 방 안에서 누군가 중국어로 된 질문을 넣어주면, 그 질문을 모고 파일 캐비넷에서 답변을 찾죠. 일치되는 결과를 찾으면 그 답변이 적힌 카드를 꺼내서 전달해줍니다. 
 

출처: Learning Machines 101.com

 
밖에서 보면 이 "중국어 방"에 있는 사람은 중국어를 할 수 있는 것처럼 보일 수 있지만, 방 안의 사람은 그저 알고리듬을 따랐을 뿐이죠.

더보기

Computers don't understand what they're doing. They can add the numbers 6 and 48, but they don't understand what the number 6 is, they don't understand what the number 48 is. 

In a great example of the inability of computers to understand, John Searle decades ago gave the example of Searle's Chinese Room, Searle is isolated in a room with a bunch of file cabinets in these file cabinets are filed a number of questions and answers written in Chinese now. 

Now, Searle did not write or read Chinese, but he was in this room with these file cabinets and somebody would come and they would slip a slip of paper in with a question, written in Chinese, Searle would take the question, he would look at it, then he would go to all the file cabinets, until he finally found a match, once he found a match, he would take out the card, he would write down the answer on the card, it would refile the card, go over to the door and slip the answer out the door, now. 

Now, externally, it looks like Searle in this Chinese Room understands Chinese, he doesn't. He's simply following an algorithm. 

 
IBM의 왓슨(Watson)이 Jeopardy! 제퍼디 세계챔피언을 이겼을 때도 마찬가지죠. 


IBM의 왓슨은 거대한 "중국어 방"이었습니다. IBM 왓슨은 인터넷의 위키피디아에 접속 가능했고, 질문에 대한 답변을 찾을 수 있었습니다. 하지만 그게 '이해력'을 가졌던 건 아니죠. 왓슨은 자기가 왜 그렇게 반응했는 지 모릅니다. 컴퓨터는 "이해력"이 없습니다. 

더보기

[원문] where we had IBM Watson beat the world champion in Jeopardy, this was a big Chinese Room. IBM Watson had access to all of Wikipedia everything on the web, and was able to answer the queries to which it the queries made to it, but  it had no understanding, it doesn't understand why it reponded that way. Computers have no understanding. 


개리 마커스(Gary Marcus) 의 표현을 빌리면 이렇습니다. 
 
(구글의 LaMDA에 대한 평가 'Nonsense on Slit' 중) 

이 시스템들이 하는 건, 일관성이나 세상에 이해를 바탕으로 한 것이 아니라
단어를 순서대로 조합하는 것에 불과 합니다.
외국인 스크래블 선수가 영어단어가 무슨 뜻인지 모른 채
포인트를 얻기 위한 도구로 사용하는 것과 마찬가지죠. 

(출처: garymarcus.substack.com)
더보기

What these systems do, no more and no less, is to put together sequences of words, but without any coherent understanding of the world behind them, like foreign language Scrabble players who use English words as point-scoring tools, without any clue about what that mean.

 


Q. 인간만 가능하고 인공지능AI에게는 불가능한 게 어떤 게 있나요? 

What things can humans to that AI can't ? 

 
[로버트 마크스 교수 답변]: 

Dr. Robert J. Marks, II 좀 더 젊었을 때 사진

 
인간이 하는 것들 중 컴퓨터들이 할 수 없고, 절대 할 수 없는 것들이 있습니다. 
그 중에 가장 분명한 것들은 사랑, 연민과 공감 등을 포함하죠. 
그 중 분명하지 않은 것들 중에는 창작성, 이해, 지각력이 있습니다.  
이것들은 컴퓨터를 통해 복제불가능한 인간의 속성입니다. 
 
세 개 중 가장 쉬운 첫번째에 대해 얘기해봅시다. 
제 생각엔 그건 Qualia(감각질)를 포함한 지각력Sentience인데, 무언가를 할 때 우리 감각을 통해 경험하는 것들 입니다. 

감각질

감각의 질 또는 속성으로 감각질은 시각에서만이 아니라 미각,후각,청각,피부감각 등 모든 감각에서 찾아 볼 수 있다. 빨간색에는 진홍빛도 있고 연분홍도 있으며, 꼭두서니빛도 있다. 또, 밝은

terms.naver.com

 
우리가 잘 익은 복숭아를 한 입 베어 먹는다고 상상해봅시다.

출처: istock.com

 
그걸 깨물어 먹을 때, 과즙이 터져나오고, 혀를 통해 그걸 느끼고, 그 향을 느끼고, 씹습니다. 과즙을 맛보며 씹어 삼키고 다음 베어먹을 한 입을 기대하게 되죠. 
 
그걸 태생부터 후각과 미각이 없는 사람에게 설명하고 이해시키거나 소통해보려는 노력을 한다고 상상해봅시다.
 
우리는 그 사람에게 혀에서 느끼는 화학물들을 설명한다거나, 미뢰의 생물학을 설명할 수 있고, 씹는동작의 운동역학을 설명할 수 있겠지만, 그 사람에게 우리의 경험을 똑같이 경험할 수 있게 해주는 것은 불가능 합니다. 만약에 우리가 후각, 미각이 부재한 사람에게 이걸 설명하는 게 불가능하다면, 어떻게 그 경험을 복제할 수 있는 인공지능 컴퓨터 프로그램을 쓸 수 있을까요?
 
그건 그냥 불가능합니다. 
 

[영어 인터뷰 원문 접음처리]

더보기

There are somethings that humans do, that computers can't do, nor will they ever be able to do them. There obvious ones that includes love compassion and empathy. 
Some of the not-so-obvious ones that computers will never be able to generate an artificial intelligence will never be able to generate, are creativity, understanding, and sentience. 
These are attributes of humans that will not an cannot be replicated by a computer. 
well, let's take first of all, the easiest of the three. 
Which I believe is sentience, which includes, qualia - Quallia is the experience that you have from your senses when you do something. 
Imagine, taking a bite of a peach, it's nice ripe Peach, and you bite into it, you feel the just spurt in your mouth, and you feel it on your tongeu, you feel the aroma, of the peach, you chew it. and get juice and you swallow and anticipate the next bite. 
Now, try to understand, or try to comprehend, communicating that to a person who has had no sense of taste, and no sense of smell, since brith. 
Now, you could explain to that person, you could explain the chemicals that were on the tongue, you could explain the biology of taste buds. You could explain the Newtonian mechanics of biting into the peach, but duplicating that experience in that person would not be possible. 

Now, it can't explain it to a person void of the senses of smell and taste. How are you going to write a computer program to duplicate that experience, in artificial intelligence. It's simply not possible. 

 
 
 

Q. 왜 컴퓨터는 창의적이 될 수 없나요?
Why Can't Computers be creative?

 
[로버트 마크스 교수 답변] : 

사진출처: robertmarks.org

컴퓨터는 결코 창의적(creative)이 될 수 없습니다. 
이 부분에 대해서는 로저 펜로즈 교수 (Sir Roger Penrose) 이 쓴 위대한 책 '황제의 새 마음' (원제: Emperor's New Mind)이 절 납득시켰죠. 
 

컴퓨터가 창의적(creative)이 될 수 있는지에 대해 이야기하기 전에 우리는 창의성(creativity)가 무엇인지 정의 해야 합니다. 
창의성Creativity를 가장 잘 정의한 것은 셀머 브링스욜드(Selmer Bringsjord) 교수의 러브레이스 테스트(Lovelace Test)입니다.
러브레이스 테스트에 따르면 컴퓨터는 컴퓨터가 프로그래머의 의도와 설명 이상의 것을 했을 때 창의적이라고 할 수 있다고 합니다.  그렇게 간단한 거죠. 
 
아직까지 컴퓨터 프로그램이 그렇게 했다는 건 들어보지 못했습니다. 
 

[가독성 증가를 위한 영어 인터뷰 원문 접음처리]

더보기

Computers will never be creative.

This was something which was noticed by a number of people including, the genius Sir Roger Penrose, who wrote a great book, which convinced me, which was called The Emperor's New Mind. 

Before we talk about whether computer can be creative. We have to define, exactly what we mean by creativity.  Creativity is best defined by Selmer Bringsjord's Lovelace test. 

According to Lovelace Test, computers are going to be creative if the computer program does something beyond the intent  or the explanation of the computer programmer. It's that simple. There is no instance of which I am aware that a computer, a computer program has been creative. 

 
자, 예시를 들어보죠. 

이세돌이 알파고(Alphago)에게 바둑을 졌을 때, 경기 중에 알파고가 예상 외의 수를 뒀습니다. 
그 때 많은 사람들이 아무도 예상하지 못한 수를 뒀다며 놀랐었죠. 
 
이게 창의성이었을까요?
아니죠. 
알파고는 바둑을 두도록 훈련되었습니다. 
알파고는 바둑을 둔 거죠. 
만약 누군가 "알파고에게 바둑의 규칙을 설명해줘" 라고 물었는데 설명할 수 있도록 프로그램 되어 있지 않았는데도 그 규칙을 설명할 수 있다고 하면, 그건 창의적이라고 말할 수 있는 거죠. 
 
그게 프로그래머의 의도를 초월한다는 의미 입니다. 
혹은 알파고에게 포커도 둘 수 있어? 라고 물었을 때, 대답할 수 있다면 그건 약간의 창의성이라고 할 수 있겠죠. 
 
우리는 컴퓨터로부터 이런 창의성이 있다는 증거를 발견하지 못했습니다. 
 

[가독성 증가를 위한 영어 인터뷰 원문 접음처리]

더보기

Now, we can go to as an exmplae, when Lisa Dell was beat by Alphago, in the game of Go. There was one point in the game, where alphago made the surprising move. And everybody says, oh my goodness. It made a move which nobody anticipated. 

Was creativity? No, Alphago was trained to play Go. That was exactly what it was doing. 

If somebody came to Alphago and said, can you explain me the rules of Go. and it was able to explain it, without being programmed to do so. 
Then, that would be creative.

 

That would be beyond the intent of the original programmers, or if you went to the Alphago, and said, do you play Poker? and it would either answer yes or no. That would be modicum of creativity. so, no. 

We see no evidence of creativity in computers. 

 

Q. 음악을 만드는 AI가 있잖아요? (그건 창의적인 게 아닌가요?)
What about the cases where AI is said to creat music?

음악생성 AI 의 예 (SoundRaw, AIVA, SoundFul, Ecrette Music, JukeBox (by Open AI 등)

 
[로버트 마크스 교수 답변]: 
 

Robert J. Marks 교수 사진 (출처: wikipedia.org)

AI가 창의적인 활동을 한다고 종종 주장됩니다. 작곡을 하는 AI를 만들었다거나. 
전형적인 시나리오에서 AI가 만든 음악이 어디서 나왔는지 생각해봅시다. 
 
누군가 음악을 만드는 AI를 만들고 싶다고 하면, 특정 장르를 사용하겠죠. 바하 라든지요. 
그리고 인공지능을 (머신러닝을 통해) 바하의 여러가지 음악 작품로 훈련시키겠죠.
그런 훈련이 끝난 후, 음악을 생성해해겠죠. 그 음악이 누구의 음악과 유사할까요? 
바하입니다. 
 
그렇게 훈련된 인공지능이 스트라빈스키의 음악을 생성해내지 않을 거고, 재즈를 생성하지 않을꺼고, 
제가 제일 좋아하는 작곡가 찰스 아이브스의 곡을 만들어내지 못할 겁니다. 
인공지능은 사일로(데이터베이스) 안에서만 작동합니다. 
'상자 밖에서 생각할 수 없습니다.''상자 밖'에서 생각하는 건 창의력에 기본이죠. 
 

[영어 인터뷰 원문 접음처리]

더보기

AI is often purported to perfrom, creative acts. 
Such as writing music and you'll hear for example, somebody saying 
I wrote artificial intelligence that creates music. 
Let's peel that back and look at where that music came from. 
Here's a very typical scenario.

One wants to write artificial inteligence that composes music. 
So one will go to a specific genre, such as Bach. 
And it will train the artificial intelligence using all of the voluminous works of Bach. 

And then once it's trained it will generate music. 
Guess what that music souds like, Bach. 

That trained artificial intelligence will never generate the work of Stravinsky. 
It'll never generate Jazz, It'll never generate one of my favorite composers Charles Ives. 
Artificial Intelligence is only able to work within the silo of what it was trained to do. 
It cannot think out side of the box. 

Thinking outside of the box mandatory for creativity. 

 


Q. 멋진 그림을 생성해내는 AI가 있잖아요? 그건 창의적인 거 아닌가요?

AI can generate amazing painting doesn't that show creativity? 

자연언어 설명을 통해 그림/이미지를 생성한 AI - DALL.E [달이] *화가 달리의 이름으로 한 말 장난같음

 
[로버트 J. 마크스 교수 답변]: 
 
에드먼드 벨라미 (Edmond de Belamy)라는 그림이 있습니다. 
이건 초상화의 대가들의 그림을 통해 학습된 AI를 통해 생성된 그림이죠.
인공지능을 훈련시키기 위해 사용된 그림 중엔 화가 렘브란트 (Rembrandt)의 그림들이 있다고 언급됐었죠. 
 

출처: Researchgate.net&amp;amp;amp;amp;amp;amp;nbsp; (에드먼드 벨라미 그림을 보고 있는 두 사람)

 
자, 이 건은 어떤가요? 
AI가 만들어낸 그림들은 누구의 그림과 비슷할까요?
초상화의 대가들이겠죠. 
제 추측엔 훈련데이터에는 피카소의 큐비즘을 사용하지 않았을 거고, 잭슨 폴락(Jackson Pollock)의 그림을 쓰지 않았을꺼에요. 

 

 그 인공지능은 초상화의 대가들의 그림을 봤을 뿐이죠. 

(하얀 선 위) 인공지능이 그린 에드먼드 벨라미 (선 아래) 렘브란트의 인물초상화 4점

다시 한 번 말하지만 인공지능은 '상자 밖'에서 생각할 수 없습니다.
인공지능은 상자 안에서 훈련된 것들을 통해 보간할 뿐이죠. 
 
창의성은 '상자 안'의 것들, 현재 상황들을 다 버리고 시작됩니다. 
거기서 새로운 혁신적인 생각들을 해내는 겁니다. 
 
아인슈타인의 예를 들어보죠. 
그는  당대 사람들이 전기장이 진공상태에서 어떻게 이동할 수 있는 지 몰라서 생각해냈던 거 에테르(luminous aether)라는 '신조'를 버렸죠. 그 당시의 에테르에 대한 신조가 '사일로/상자'였던 거죠. 
아인슈타인은 마이클슨 몰리 실험 (The Michelson-Morely Experiment)과 같은 다른 실험들에서 모티브를 얻어 창의적인 상대성이론을 만들어냈죠. 

출처: 브리태니커 사전

[가독성 증가를 위한 영어 인터뷰 원문 접음처리]

더보기

There was painting called Edmund Bellamy. And Edmund Bellamy was a painting that was done by aritifical intelligence trained on the portraits of Masters. Rembrandt was specifically mentioned as one of the places that  they went to get paintings to train this artificial intelligence. So we have a case. Guess what the paintings that were generated by AI looked like?
It looks like the paintings of the Masters. Right? I suspect, I don't know for sure, but I suspect they did not use Picasso's cubism in any of the training. I suspect they didn't use Jackson Pollock's paint splattering paintings in the training. 
They  simply looked at these great masters of the past that did portraits. 
 So, again, artificial intelligence is unable to think outside the box. 

And thinking outside of the box, is a characteristic of creativity. 
It can only take what is inside the box-what it was trained to do, and interpolate.
The idea of creativity, in fact, takes the status quo, inside the box, and it discards it. 
It discards the status quo for some new innovative ideas. you'd only have to go back to Einstein for example.

What did he do? He discarded the dogma of Ether, in space. because they had no idea, 
that electromagnetic waves could travel in a vacuum. He discarded the idea that light was a relative in speed, this was the accepted Dogma at the time. This was inside the silo. 
He discarded that and he took these ideas motivated by a number of experiments, including Michaelson Worley Experiment, and from this generated the wonderfully creative idea of relativity. 

So that is an example of creativity. 
Creativity normally requires some sort of avoidance of things within the silo, that need that are to be done. external to the silo in order to see creativity. 

 

 

Q. 컴퓨터가 의식을 갖게 될 수 있을까요?
Will a computer ever be conscious? 

 
[존 레녹스 교수 답변] 

John Lennox 교수 사진 (출처: socrates in the city)

우리는 아직 의식이 뭔지도 모르는데 여기에 대해 어떻게 답을 할 수 있을까요?
의식이라는 아직 인류에게 커다란 장벽입니다.
제가 얘기해본 사람, 제가 읽어본 책의 저자 중 아무도 의식이 뭔지 모릅니다. 
그렇기 때문에 추측은 무의미 합니다. 
 
로저 펜로즈(Roger Penrose)는 이 주제에 대해 흥미로운 의견을 제시했죠. 
그는 원칙적으로 인간의 생각/마음(mind)의 활동을 컴퓨터에 시뮬레이션할 수 없다고 주장합니다. 
인간의 생각/마음은 컴퓨터가 할 수 없는 걸 할 수 있기 때문이죠. 그것들은 계산/연산가능(computable/컴퓨터로 구현가능)한 것들이 아닙니다. 
 
수학에서의 예를 들면, 괴델(gödel)의 유명한 이론이 있습니다. Arthimetic,계산 혹은 연산이 완전할 수 없다는 거죠. (괴델의 불완전성 정리/Gödel’s Incompleteness Theorems)
이건 아주 복잡한 내용이지만 굉장히 흥미로운 내용입니다. 
 
최근에 노벨상을 수상한 훌륭한 로저 펜로즈와 같은 사람이 그건 (인간의 생각/mind)는 연산가능한 것이 아니라고 하고 있죠. 
 
컴퓨터는 그걸 할 수 없습니다. 
 

[가독성 증가를 위한 영어 인터뷰 원문 접음처리]

더보기

how could you answer something like that when we have no idea what consciousness is. 

Consciousness is the huge barrier, no one that I've ever spoken to or read has any idea what it is. So speculation is irrelevant. I doubt it very much. Because computers appear to certain limits. 

Roger Penrose is very interesting on this topic. He argues that you'll never be able even in principle to simulate, the activity of the human mind on a computer because the human mind can do things that a computer cannot do. they are not computable. and one of them is for example in Mathematics the proof of gödel's famouse theorem that arithmetic is not complete. these are verycomplext things. but it's very intersting to see that. Someone as bright as Penrose with this recent Nobel Prize argues that "no, it isn't computable". So, computers will never be able to do that. 

 

 

Q. 왜 컴퓨터는 사랑, 연민, 공감을 할 수 없나요?
Why Can't Computers Experience Love, Compassion, or Empathy ? 

 
[로버트 마크스 교수 답변] 

전 확실히  AI, 인공지능이 사랑, 연민, 공감 등 인간의 속성의 많은 것들을 경험할 수 없다고 했습니다. 
하지만 그렇다고 AI가 그걸 흉내내지 못할 거란 건 아니죠. 
 
스티븐 스필버그의 영화 중, A.I 란 영화가 있었죠. 

영화 A.I (2001년) 포스터

이 소년 로봇은 사랑을 모방하도록 훈련되었습니다.
그리고 이 작은 소년 로봇이 사랑하도록 활성화(activate)되는 장면이 있습니다. 
 

 
배우의 이름은 오스만트(Harley Joel Osment) 였는데, 갑자기 연민과 사랑에 휩싸이죠. 
이건 사랑의 시뮬레이션입니다. 
흉내를 내는 거죠.
David Glortner(?)는 AI가 흉내내지 못할 건 아주 적을 거라고 말한 바 있죠.  
 

[가독성 증가를 위한 영어 인터뷰 원문 접음처리]

더보기

I did make the claim that aritifical intelligence cannot experience love, compassion, and empathy and other human attributes of that sort. It doesn't mean that they can't mimic it.
There was a great Steven Spielberg Movie, called A.I, where this little boy robot was trained to emulate love and, there's this great scene where the little boy robot is activated to love, and the little actor that played it, I think his name was Osment, he, all of a sudden you just saw him a wave of compassion and love. 
This was a simulation of love. It was something that was being mimicked. 
David Glortner, said there's probably very few things that A.I can't mimic. 

 


 

 

Q.미래의 AI에 대한 자극적인 예측은 어떤 게 있나요?
 What are the heady predictions of Future AI? 

 
[로버트 마크스 교수 답변]: 

Dr. Robert J. Marks, II

 
AI/인공지능이 더 나은 AI/인공지능를 작성하고, 그 AI는 더 나은 AI를 작성하고, 소위 '특이점 Singularity' 에 도달할거라는 주장을 하는 사람들이 있습니다. 인공지능이 인류의 능력을 복제(모방)하는 것에 성공하거나 초월하는 시점을 말하죠. 
 
하지만 그 주장은 AI/인공지능이 창의적일 수 있다는 걸 가정합니다. 
하지만 AI/인공지능은 창의적일 수 있는 능력이 없습니다. 
창의성(creativity)가 무엇인지 정확히 정의할 필요가 있겠지만 AI는 그 역량이 없습니다. 
그래서 그건 실체가 없는 공허한 주장에 불과합니다. 
 
컴퓨터 개발 분야에는 '무어의 법칙 Moore's Law'  (마이크로칩의 밀도가 24개월 마다 2배로 늘어난다는 법칙)이라는 게 있죠. 컴퓨터의 능력이 계속 배가 된다는 거죠. 이 법칙은 한동안 유지되어왔습니다. 
 
하지만 컴퓨터 성능의 기하급수적 증가는 지속가능하지 않습니다. 모든 기하급수적 증가는 미래의 한 시점에서 사그라들었습니다. 
 
어떤 이들은 우리의 컴퓨터 능력(computational power)가 증가하면 지금의 AI가 할 수 없는 것들을 해낼 수 있다고 주장하죠. 더 뛰어난 AI를 만들어낼 수 있는 AI가 인류의 능력치를 초월할 거라는 거죠.
이 역시 가능한 게 아닙니다. 
 

[가독성 증가를 위한 영어 인터뷰 원문 접음처리]

더보기

[원문] (1) We have a number of, often made claims about artificail intelligence for example intelligence is going to write better artificial intelligence, that's going to write better artificial intelligence. and pretty soon, we're going to reach the so-called Singularity, where artificail intelligence, duplicates or exceeds, the capability of human beings. This however assumes that Aritificial Intelligence can be creative. 

And Artificial Intelligence doesn't have the capability of being creative. 

We have to define exactly what we mean by being creative, but it doesn't have the capability. So that problem is, that claim is totally vacuous. It has no substance at all. 

well there are certain laws in the development of computer, one of them is Moore's Law, which says that there's a doubling in the capacity, of computers, every year, every 18 months or whatever. And it's been held for a long time.  
However, the expontential increase in the performance of computers is not sustainable. 
Anytime you have an exponential increase, in anything, it has level out at some point in the future. 

(2) You also have claims that because of this increase in computational power, that we're gonna be able to do things that we couldn't do with AI currently. 
Onf of these is AI writing better AI that writes better AI that is ultimately going to exceed capabilities of human beings. and this is also not something which is possible.

 
 

Q. 레이 커츠와일이 제안한 특이점(singularity)는 가능한가요?
Is Ray Kurzweil's Proposed Singularity Possible? 

[로버트 마크스 교수 답변]: 

Robert J. Marks

 

레이 커츠와일(Ray Kurtweil)은 초지능(Super-intelligence)에 대한 생각을 가지고 있죠. AI가 언젠가 더 나은 소프트웨어를 작성할 수 있는 AI를 작성해서 인간의 능력치를 초월할 거라는 생각이죠. 

커즈와일

어릴 때부터 과학에 관심이 많았던 커즈와일은 1965년 위대한 작곡가들의 스타일을 모방해 작곡할 수 있는 컴퓨터 프로그램으로 국제과학기술경진대회에서 1등상을 수상했다. 그 프로그램

100.daum.net

 
이건 AI가 창의적creative이라는 것을 가정하고 있습니다. 이 프로그램은 AI가 창조한 거죠. 하지만 인공지능은 창의적이지 않습니다. 
 

[가독성 증가를 위한 영어 인터뷰 원문 접음처리]

더보기

Ray Kurzweil has the idea of super-intelligence, that AI at some point will be able to write better software, that in turn, writes better AI that in turn writes better AI, and pretty soon it will exceed the capabilities of human beings. 

Now, this assumes AI is creative. If AI creats a better AI program, this is something that the AI program has created. and Artificial is not creative. 

 

※[역자추가]
레이 커즈와일은 컴퓨터 과학자이자 미래학자인데, MIT출신으로 1982년에는 뮤지션 스티비 원더(Stevie Wonder)와의 친구가 되고 나중에 음악 신디사이저를 개발하는 사업도 했다고 합니다. 그 회사를 우리나라 영창악기가 인수했다는 재밌는 역사도 있네요. (한국에선 같은 Kurtzweil 인데, 사람은 커즈와일, 악기 브랜드의 이름은 커츠와일로 번역되서 사용되고 있네요.) 

 

출처: Kurtzweil.com

 
 
Q. 어떤 사람들은 언젠가 우리가 우리를 컴퓨터로 업로드 할 수 있을 거라고 하는데, 이건 가능한가요?
Some say we will be able to upload ourselves into computers one day, is that possible? 
 
 
[로버트 마크스 교수 답변: 

Robert J. Marks

우리가 우리를 컴퓨터로 업로드해서 일종의 "영생永生"을 얻을 수 있을 거라고 주장는 사람들이 있습니다. (기사링크)

Apparently we’re all going to live forever by 2029 «  Kurzweil

Source: -->Metro — May 6, 2016 | Ashitha Nagesh This is a summary. Read original article in full here. Google’s chief futurist Ray Kurzweil says humans will have eternal life by 2029. Kurzweil has been accurate in his predictions so far. Back in the 19

www.kurzweilai.net

 
하지만 그건 불가능 합니다. 
 
인간의 속성/특성 중에는 연산가능(computable)하지 않은 것들이 있고, 그런 특성들은 업로드 될 수 없습니다.  인간의 연산가능한 부분들만 업로드 할 수 있겠죠. 하지만 우리에겐 연산가능하지 않은 아주 많은 여러가지 다른 특성이 존재합니다. 

 
[가독성 증가를 위한 영어 인터뷰 원문 접음처리]

더보기

There's a claim that we might be able to upload ourselves into a computer, in some sense, gain eternal life, I suppose. 
No, that's not possible. 

There are certain aspects about humans that are not computable. And the aspects of humans that are not computable cannot be uploaded. 
You can only upload the computable parts of human beings. 
and we have a lot of different qualities that are not computable. 

 


3. 챗GPT 에 대해 "딥deep" 하게 Learning알아봅시다

 

01. 챗GPT에게 물어본 인공지능의 한계 

Tell me about limitation of AI 라고 요청했더니 쓱쓱 답변을 적어내려갑니다. 
쉽게 요약하면 방대한 량의 언어 데이터에 대한 통계를 활용한 '자동완성'기능과 비슷한 거라고 하지만 
정말 질문을 이해한 것 같이 답안 텍스트를 생성해내 나가는 것이 재밌습니다. 

번역도 시켜봤습니다. 그냥 번역시키면 재미 없으니 소제목은 진하게(bold)처리하고 추가로 3개 더해달라고 주문했습니다. 
 
번역하다 8에서 끊기는데 (답변의 용량한계가 있습니다.), 이어서 계속 답변해봐- 하니 이어서 써줍니다. 

(이하 챗GPT 작성) 

"인공지능(AI)은 최근 몇 년 동안 놀라운 발전을 이루었지만, 아직 해결해야 할 몇 가지 한계점이 있습니다. AI의 주요 한계점은 다음과 같습니다.

1. 상식의 부재: 인간이 가지고 있는 상식적 추론 능력을 이해하지 못하는 AI는 문맥을 이해해야하는 작업을 수행하는 것이 어렵습니다.

2. 한계된 창의성: 미리 정의된 알고리즘과 데이터 세트에 의존하는 AI 시스템은 창의적인 해결책이나 아이디어를 생성할 수 없습니다.

3. 데이터에 대한 의존도: AI 알고리즘은 학습에 충분한 데이터가 없거나 데이터가 편향되어 있는 경우에도 어려움을 겪을 수 있습니다.

4. 감성 지능(Emotional Intelligence)의 부재: AI 시스템은 감정을 인식하거나 인간의 의사소통의 뉘앙스를 이해하는 능력이 없기 때문에 비꼬거나 반어적인 말 등을 해석하는 것이 어렵습니다.

5. 윤리와 책임성: AI가 더욱 발전함에 따라 윤리적 사용과 책임성에 대한 문제가 더욱 중요해지고 있습니다. AI 시스템의 결정은 실제 세계적인 영향을 미칠 수 있으므로, 이러한 시스템이 윤리적이고 책임성있게 개발 및 사용되도록 보장하는 것이 중요합니다.

6. 한계된 전이 학습: AI 모델은 특정 작업을 해결하는 데 특화되어 있으며, 학습한 지식을 다른 작업으로 전이하는 것이 어렵습니다.

7. 한계된 설명: 많은 AI 알고리즘은 "블랙 박스"로 간주되어 해석하기 어렵기 때문에 알고리즘이 결정을 내리는 방식을 이해하는 것이 어렵습니다.

위와 같이 AI는 많은 한계점이 있지만, 아직 많은 도전적인 과제들이 존재하여, AI를 보다 효과적이고 윤리적이며 책임성있게 만드는 것이 중요합니다.

추가로 3가지 한계점은 다음과 같습니다.

8. 인간적인 결정력의 부재: AI는 인간과 같은 직관력과 판단력을 가지지 못해, 일부 분야에서는 아직까지 인간의 판단을 대체할 수 없습니다.

9. 보안 문제: AI 시스템은 인간의 개입 없이 작동할 수 있기 때문에, 보안 문제가 발생할 수 있습니다. 예를 들어, 해커들은 악의적인 목적으로 AI 시스템을 조작할 수 있습니다.

10. 기술 발전 속도: AI 기술은 매우 빠르게 발전하고 있지만, 이로 인해 기존에 작동하던 시스템과의 호환성 문제가 발생할 수 있습니다. 이러한 호환성 문제는 업계에서 최신 기술을 적용하기 위해 추가적인 비용과 시간을 필요로 합니다.

이러한 한계점들은 AI를 보다 효과적으로 활용하기 위해서는 극복되어야 할 문제들입니다. 이를 극복하기 위해서는 AI 연구에 더 많은 투자와 인프라 구축, 윤리적인 가이드라인의 제공, 그리고 보안 문제와 호환성 문제에 대한 대비책을 마련하는 것이 중요합니다.

 
위 인간학자(?)들이 인공지능의 제약에 대한 내용, 그리고 챗GPT가 핫이슈가 된 후 나온 기사에서도 나왔던 사례들을 떠오르게 하는 적합한 요약인 것 같습니다. 
 
이번 글을 위해 참고한 영상 중, 유럽 최대 기술 컨퍼런스인 Web Summit 2023의 영상이 있습니다. 
제목은 무려 "Debunking the great AI lie" 였는데 번역하면 "AI에 관한 거대한 거짓말을 파헤쳐보다" 정도가 되겠네요. 
 

YouTube 검색결과

 
여기서 현재의 인공지능시스템과 개발방향에 대한 비평을 마주하게 됐습니다. . 
 
제가 에버노트에 transcribe한 노암 촘스키의 비평을 챗GPT에게 요약시켜봤습니다. 
 

 
학술적이지 않은 블로거 스타일로 번역해보라고 시켰더니.. 해냅니다. 
 

일부 문장을 수정해서 아래 기재합니다. (삭선과 괄호로 정정 및 추가) 

안녕하세요! 이번에는 Noam Chomsky의 비판에 대해 이야기해 볼까요?

그는 GPT와 같은 AI 시스템에 대해은 두 종류의 문제점이 있다고 말합니다. 

첫째, 이 시스템이 (너무 약해서) 특정한 작업을 수행하지 못한다는 경우이죠. 
이는 시스템의 성능을 더욱 강화하여 해결할 수 있습니다.

하지만 두 번째 문제는 이 시스템이 너무 강력하다는 것입니다 한 경우입니다.

이는 시스템이 어떤 작업을 수행하면 안 되는 일도 수행한다는 것을 의미합니다.
이런 경우는 수정하기 어렵습니다.

특히, 언어와 인지 과정에 대한 연구를 진행하는 촘스키(Chomsky)는 GPT 시스템이 언어에 대해 일부 한계가 있다는 것을 지적합니다.
이는 시스템이 어떤 언어라도 동일한 성능을 발휘한다는 것을 의미합니다.

예를 들어, 언어의 규칙적인 순서를 무시하는 불가능한 언어를 제시해도 GPT는 이를 잘 처리합니다.
그러나 이것은 사실상 언어가 아니기 때문에 문제가 됩니다.

따라서 Chomsky는 이러한 강력한 AI 시스템에 대한 문제를 해결할 수 있는 방법을 찾기 어렵다는 것을 지적하고 있습니다. 이러한 시스템이 어떤 작업이든 수행할 수 있는 만큼 강력하기 때문에 언어에 대한 연구를 진행하는 데에는 한계가 있다는 것입니다.

이상 Noam Chomsky의 AI 시스템에 대한 비판에 대해 알아보았습니다. 감사합니다!

[인터뷰 원문 접음처리] 

더보기

(error/ muted) .. be efficient, 
one way, is, it's not strong enough. 
fails to do certain things. 

The other way is it's too strong. 
It does what it shuoodn't do. 

Well, my own interest happen to be language and cognition, language, specifically, 

So take GPT. there's, uh, Gary Marcus and other have found lots of ways in which the systems deficient. This system and others doesn't do certain things.
 That can in principle at least, be fixed, you add another trillion parameters, double the number of terrabyes, and maybe do better. 


When a system is too strong, it's unfixable typically.
That's the problem with GPT and the other systems.
So if you give a database the GPT system which happens to be from an impossible language, one that violates the rules of language, they'll do just as well. often better. Because the rules can be simpler. 
For example, one of the fundamental properties of the way language works, there's good reason for it. Is that the rules, the core rules, do not ignore linear order of words. They'll ignore everything that you hear.
they attend only to abstract structures that the mind creates. 

So it's very easy to contruct impossible  languages which use very simple procedures, involving linear order of words. 

Trouble is that's not language. But GPT will do just fine with them. 

So it's kind of as if somebody were to propse, say a revised version of the periodic table which included all the elements all the possible elements, and all the impossible elements, and didn't make any distinction between them, that wouldn't tell us anything about the elements and if a system works just as well for impossible languages as for possible ones, by definition, not telling us anything about language. 

And that's the way these systems work. it generalizes the other systems too.  So the problem is the Deep problem, that concerns me is too much strength. I don't see any conceivable way to remedy that. " 

요약은 정말 탁월합니다. 
근데 이 질문을 할 때, 챗GPT는 오류를 하나 보여줬죠. 
 
제 질문:
다음 질문에 넣을 노암 촘스키에 대한 비평을 요약해줄 수 있어? 
 
챗GPT 답변: 
(갑자기 노암 촘스키의 언어학 관련 비평을 요약함..)


똑똑한 아이인 것 같은데 뭔가 아직 신뢰는 안갑니다. 


1부를 통해 개념정리를 해봤습니다. 
1부 내용만으로도 인터뷰 내용을 다 집어 넣으니 너무 길어졌네요. 노암 촘스키와 개리 마커스의 챗GPT에 대한 비평은 2부로 넘겨야할 것 같습니다. 
긴 글 읽어주셔서 감사합니다.