{"success":true,"count":10,"items":[{"videoId":"aR20FWCCjAs","chunkIndex":0,"totalChunks":10,"title":"Ilya Sutskever – We're moving from the age of scaling to the age of research — Part 1 of 10","thumbnail":"https://i.ytimg.com/vi/aR20FWCCjAs/maxresdefault.jpg","duration":5763,"uploader":"Dwarkesh Patel","youtubeUrl":"https://www.youtube.com/watch?v=aR20FWCCjAs","keywords":["artificial-intelligence","machine-learning","llm","reinforcement-learning","pretraining","evaluation","generalization","research","technology-trends"],"normalizedKeywords":["기술 트렌드","엔지니어링","교육"],"targetAudience":[{"who":"AI 연구자","why":"모델 성능과 일반화, RL·프리트레이닝의 차이를 깊게 다룸"},{"who":"엔지니어","why":"실사용 성능과 벤치마크 성능의 괴리를 이해하는 데 도움됨"},{"who":"창업자","why":"AI 도입 효과가 왜 즉시 경제성으로 이어지지 않는지 볼 수 있음"}],"normalizedAudience":["리서처·학자","엔지니어·개발자","창업자·스타트업"],"summary":"이 영상은 AI가 이미 놀라울 정도로 똑똑해 보이는데도, 현실의 경제적 영향은 왜 생각보다 느리게 드러나는지에서 출발한다. 화자는 그 원인으로 벤치마크(evals)에서의 고성능과 실제 세계에서의 일반화 능력 사이의 간극, 그리고 RL 학습이 모델을 특정 환경에 과도하게 맞추는 경향을 제시한다.\n\n또한 프리트레이닝과 RL의 차이를 인간의 학습 방식에 비유하며, 프리트레이닝의 강점은 '더 좋은 일반화'라기보다 엄청난 양과 자연스러운 데이터에 있다고 본다. 결론적으로, AI 발전의 핵심은 단순히 평가 점수를 올리는 것이 아니라, 어떤 학습 환경이 진짜 전이 가능한 능력을 만드는지 연구하는 쪽으로 이동해야 한다는 문제의식이 중심에 있다.","insights":["벤치마크 성능과 실사용 성능은 자동으로 연결되지 않는다.","RL은 능력을 키우지만 동시에 과도한 특화도 만든다.","평가 최적화는 사람 연구자도 쉽게 'reward hacking'에 빠뜨린다.","프리트레이닝의 힘은 지능적 설계보다 데이터 규모와 자연성이다.","진짜 과제는 환경을 늘리는 것보다 전이 가능한 학습법을 찾는 일이다."],"keyClips":[{"clipId":"aR20FWCCjAs:c0:1-19","startSegmentIndex":1,"endSegmentIndex":19,"startTime":0.13999999999999999,"endTime":124.44818181818181,"durationSeconds":124.3,"preview":"AI는 왜 느껴지지 않나","mustSee":true},{"clipId":"aR20FWCCjAs:c0:20-39","startSegmentIndex":20,"endSegmentIndex":39,"startTime":127.82000000000001,"endTime":279.16205882352943,"durationSeconds":151.3,"preview":"평가와 현실의 간극","mustSee":false},{"clipId":"aR20FWCCjAs:c0:40-52","startSegmentIndex":40,"endSegmentIndex":52,"startTime":278.94,"endTime":373.06933333333336,"durationSeconds":94.1,"preview":"전이 능력의 핵심","mustSee":true},{"clipId":"aR20FWCCjAs:c0:53-67","startSegmentIndex":53,"endSegmentIndex":67,"startTime":374.06,"endTime":486.2886363636364,"durationSeconds":112.2,"preview":"두 학생의 비유","mustSee":false},{"clipId":"aR20FWCCjAs:c0:68-83","startSegmentIndex":68,"endSegmentIndex":83,"startTime":486.06,"endTime":600.5066666666667,"durationSeconds":114.4,"preview":"프리트레이닝의 본질","mustSee":false}],"curatedSegments":[{"segmentIndex":51,"text":"of environments, but to figure out an approach which lets you learn from one environment and","startTime":359.17999999999995,"endTime":365.19,"durationSeconds":6,"level":"C1","overallScore":7.8,"rationale":"전이 학습의 핵심 아이디어를 제시함."},{"segmentIndex":59,"text":"Right. I think that's basically what's going on. The models are much more like the first student, but even more. Because then we say, the model should","startTime":416.7,"endTime":424.48999999999995,"durationSeconds":8,"level":"B2","overallScore":7.8,"rationale":"비유를 핵심 주장으로 연결해 통찰적임."},{"segmentIndex":40,"text":"The value function lets you short-circuit the wait until the very end.","startTime":908.86,"endTime":915.0999999999999,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"가치함수 효용을 압축한 핵심 문장이다."},{"segmentIndex":43,"text":"As soon as you conclude this, you could already get a reward signal a thousand timesteps previously, when you decided to pursue down this path.","startTime":934.22,"endTime":943.6883333333334,"durationSeconds":9,"level":"C1","overallScore":7.6,"rationale":"조기 보상 아이디어가 매우 선명하다."},{"segmentIndex":56,"text":"it's impressive that they have this much utility while still being rather simple to understand.","startTime":1026.0600000000002,"endTime":1036.3016666666667,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"단순성과 효용의 대비가 통찰적이다."},{"segmentIndex":1,"text":"The big breakthrough of pre-training is the realization that this recipe is good.","startTime":1202.14,"endTime":1208.1034615384617,"durationSeconds":6,"level":"B2","overallScore":7.2,"rationale":""},{"segmentIndex":44,"text":"obvious. That seems like a very fundamental thing. So this is the crux: generalization. There are two","startTime":1508.46,"endTime":1518.3632352941174,"durationSeconds":10,"level":"B2","overallScore":6.6,"rationale":""},{"segmentIndex":76,"text":"But if people exhibit great ability, reliability, robustness, and ability to learn in a domain that","startTime":1747.1000000000001,"endTime":1756.4675,"durationSeconds":9,"level":"C1","overallScore":7,"rationale":""},{"segmentIndex":34,"text":"something like, \"If ideas are so cheap, how come no one's having any ideas?\"","startTime":2245.1800000000003,"endTime":2250.955,"durationSeconds":6,"level":"B2","overallScore":8,"rationale":"역설적 반문이 강한 통찰과 표현성을 가짐."},{"segmentIndex":49,"text":"research where the end artifact was made safer mostly through just thinking about","startTime":2745.26,"endTime":2753.200384615385,"durationSeconds":8,"level":"C1","overallScore":7.8,"rationale":"안전 개선 방식에 대한 핵심 주장."},{"segmentIndex":55,"text":"But this is a really powerful thing and we don't even know how to conceptualize how people interact with it, what people will do with it. Having gradual access to it seems like a","startTime":2789.6600000000003,"endTime":2799.86,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"불확실한 상호작용 위험을 짚음."},{"segmentIndex":56,"text":"better way to maybe spread out the impact of it and to help people prepare for it.","startTime":2800.78,"endTime":2805.5808823529414,"durationSeconds":5,"level":"B2","overallScore":7.8,"rationale":"점진 접근의 실용적 이유가 분명함."},{"segmentIndex":18,"text":"but picking them all up at the same time, and then amalgamating their learnings, you basically have a model which functionally becomes superintelligent even without any sort","startTime":3125.42,"endTime":3135.449230769231,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"동시 학습 결합이 초지능으로 이어진단 통찰."},{"segmentIndex":37,"text":"Economic growth is one way to put it. A Dyson sphere is a lot of economic growth. But another way to put it is that you will have, in potentially a very short period of time...","startTime":3341.34,"endTime":3351.177368421053,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"경제성장을 우주 규모 비유로 확장해 강렬함."},{"segmentIndex":41,"text":"AI being deployed incrementally and in advance. One very difficult thing about AI is that we are","startTime":3382.78,"endTime":3394.799411764706,"durationSeconds":12,"level":"C1","overallScore":7.8,"rationale":"점진 배치의 중요성을 핵심 원칙으로 제시함."},{"segmentIndex":54,"text":"You’ve got to be showing the thing. You’ve got to be showing the thing.","startTime":3492.86,"endTime":3496.7814285714285,"durationSeconds":4,"level":"B2","overallScore":7.6,"rationale":"보여줘야 한다는 원칙을 강하게 압축했다."},{"segmentIndex":45,"text":"care for sentient life in a very single-minded way—we might not like the results. That's really","startTime":3923.34,"endTime":3929.5024999999996,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"위험의 본질을 직설적으로 압축함."},{"segmentIndex":55,"text":"The \"It\" we don't know how to build, and a lot hinges on understanding reliable","startTime":3989.98,"endTime":3998.964,"durationSeconds":9,"level":"C1","overallScore":7.4,"rationale":"핵심 병목을 압축적으로 제시함."},{"segmentIndex":7,"text":"still guiding our actions so strongly is an example of alignment success.","startTime":4256.0599999999995,"endTime":4263.613333333334,"durationSeconds":8,"level":"C1","overallScore":7.2,"rationale":""},{"segmentIndex":13,"text":"I think it's actually really mysterious how evolution encodes high-level desires.","startTime":4296.46,"endTime":4306.887272727273,"durationSeconds":10,"level":"C1","overallScore":6.8,"rationale":""}],"generatedAt":"2026-06-24T23:52:34.659Z","keyClipsTotalSec":2154},{"videoId":"aR20FWCCjAs","chunkIndex":1,"totalChunks":10,"title":"Ilya Sutskever – We're moving from the age of scaling to the age of research — Part 2 of 10","thumbnail":"https://i.ytimg.com/vi/aR20FWCCjAs/maxresdefault.jpg","duration":5763,"uploader":"Dwarkesh Patel","youtubeUrl":"https://www.youtube.com/watch?v=aR20FWCCjAs","keywords":["machine-learning","reinforcement-learning","deep-learning","pretraining","value-function","neuroscience","emotions","scaling","ai-research"],"normalizedKeywords":["엔지니어링","기술 트렌드","교육"],"targetAudience":[{"who":"AI 연구자","why":"프리트레이닝, 가치함수, 스케일링의 한계를 연구 관점에서 다룸"},{"who":"ML 엔지니어","why":"강화학습의 학습 신호 설계와 중간 평가 방식이 실무적으로 중요함"},{"who":"학생·주니어","why":"요즘 AI의 핵심 개념을 감으로 이해하기에 좋은 개념 설명이 많음"}],"normalizedAudience":["리서처·학자","엔지니어·개발자","학생·주니어"],"summary":"이 영상은 프리트레이닝이 인간 학습이나 진화의 역할을 얼마나 대체할 수 있는지, 그리고 그 한계는 무엇인지에 대한 대화를 중심으로 전개된다. 이lya Sutskever는 인간이 적은 데이터로도 더 깊게 배우는 점, 그리고 감정이 사실상 가치함수처럼 의사결정을 돕는다는 점을 들어, 단순히 데이터와 파라미터를 키우는 것만으로는 부족할 수 있다고 말한다.\n\n이어 강화학습에서 흔히 쓰는 사후 보상 방식의 문제를 설명하며, 장기 과제에서는 중간에 잘못된 경로를 빨리 판별해 주는 가치함수가 중요해질 것이라고 본다. 마지막으로 그는 '스케일링'이란 말이 어떻게 ML의 사고를 바꿨는지 짚으면서, 이제는 데이터·연산량만이 아니라 더 넓은 연구 축을 고민해야 할 시점이라고 주장한다.","insights":["프리트레이닝은 인간 학습을 닮았지만 깊이와 구조가 다르다.","감정은 의사결정을 돕는 내장형 가치함수로 볼 수 있다.","장기 과제일수록 최종 보상만으론 학습이 너무 느리다.","가치함수는 잘못된 경로를 초기에 끊게 해준다.","스케일링은 숫자 확대가 아니라 사고방식의 전환이다."],"keyClips":[{"clipId":"aR20FWCCjAs:c1:3-9","startSegmentIndex":3,"endSegmentIndex":9,"startTime":614.38,"endTime":668.1347058823529,"durationSeconds":53.8,"preview":"프리트레이닝의 한계","mustSee":false},{"clipId":"aR20FWCCjAs:c1:10-18","startSegmentIndex":10,"endSegmentIndex":18,"startTime":668.78,"endTime":754.4311111111111,"durationSeconds":85.7,"preview":"감정과 의사결정","mustSee":false},{"clipId":"aR20FWCCjAs:c1:28-40","startSegmentIndex":28,"endSegmentIndex":40,"startTime":820.4599999999999,"endTime":915.0999999999999,"durationSeconds":94.6,"preview":"가치함수의 역할","mustSee":false},{"clipId":"aR20FWCCjAs:c1:41-54","startSegmentIndex":41,"endSegmentIndex":54,"startTime":918.9399999999999,"endTime":1022.8155555555555,"durationSeconds":103.9,"preview":"장기탐색의 보상","mustSee":false},{"clipId":"aR20FWCCjAs:c1:55-68","startSegmentIndex":55,"endSegmentIndex":68,"startTime":1022.54,"endTime":1129.4199999999998,"durationSeconds":106.9,"preview":"단순한 감정의 힘","mustSee":false},{"clipId":"aR20FWCCjAs:c1:69-77","startSegmentIndex":69,"endSegmentIndex":77,"startTime":1129.9,"endTime":1202.7766666666664,"durationSeconds":72.9,"preview":"스케일링 이후","mustSee":true}],"curatedSegments":[{"segmentIndex":51,"text":"of environments, but to figure out an approach which lets you learn from one environment and","startTime":359.17999999999995,"endTime":365.19,"durationSeconds":6,"level":"C1","overallScore":7.8,"rationale":"전이 학습의 핵심 아이디어를 제시함."},{"segmentIndex":59,"text":"Right. I think that's basically what's going on. The models are much more like the first student, but even more. Because then we say, the model should","startTime":416.7,"endTime":424.48999999999995,"durationSeconds":8,"level":"B2","overallScore":7.8,"rationale":"비유를 핵심 주장으로 연결해 통찰적임."},{"segmentIndex":40,"text":"The value function lets you short-circuit the wait until the very end.","startTime":908.86,"endTime":915.0999999999999,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"가치함수 효용을 압축한 핵심 문장이다."},{"segmentIndex":43,"text":"As soon as you conclude this, you could already get a reward signal a thousand timesteps previously, when you decided to pursue down this path.","startTime":934.22,"endTime":943.6883333333334,"durationSeconds":9,"level":"C1","overallScore":7.6,"rationale":"조기 보상 아이디어가 매우 선명하다."},{"segmentIndex":56,"text":"it's impressive that they have this much utility while still being rather simple to understand.","startTime":1026.0600000000002,"endTime":1036.3016666666667,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"단순성과 효용의 대비가 통찰적이다."},{"segmentIndex":1,"text":"The big breakthrough of pre-training is the realization that this recipe is good.","startTime":1202.14,"endTime":1208.1034615384617,"durationSeconds":6,"level":"B2","overallScore":7.2,"rationale":""},{"segmentIndex":44,"text":"obvious. That seems like a very fundamental thing. So this is the crux: generalization. There are two","startTime":1508.46,"endTime":1518.3632352941174,"durationSeconds":10,"level":"B2","overallScore":6.6,"rationale":""},{"segmentIndex":76,"text":"But if people exhibit great ability, reliability, robustness, and ability to learn in a domain that","startTime":1747.1000000000001,"endTime":1756.4675,"durationSeconds":9,"level":"C1","overallScore":7,"rationale":""},{"segmentIndex":34,"text":"something like, \"If ideas are so cheap, how come no one's having any ideas?\"","startTime":2245.1800000000003,"endTime":2250.955,"durationSeconds":6,"level":"B2","overallScore":8,"rationale":"역설적 반문이 강한 통찰과 표현성을 가짐."},{"segmentIndex":49,"text":"research where the end artifact was made safer mostly through just thinking about","startTime":2745.26,"endTime":2753.200384615385,"durationSeconds":8,"level":"C1","overallScore":7.8,"rationale":"안전 개선 방식에 대한 핵심 주장."},{"segmentIndex":55,"text":"But this is a really powerful thing and we don't even know how to conceptualize how people interact with it, what people will do with it. Having gradual access to it seems like a","startTime":2789.6600000000003,"endTime":2799.86,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"불확실한 상호작용 위험을 짚음."},{"segmentIndex":56,"text":"better way to maybe spread out the impact of it and to help people prepare for it.","startTime":2800.78,"endTime":2805.5808823529414,"durationSeconds":5,"level":"B2","overallScore":7.8,"rationale":"점진 접근의 실용적 이유가 분명함."},{"segmentIndex":18,"text":"but picking them all up at the same time, and then amalgamating their learnings, you basically have a model which functionally becomes superintelligent even without any sort","startTime":3125.42,"endTime":3135.449230769231,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"동시 학습 결합이 초지능으로 이어진단 통찰."},{"segmentIndex":37,"text":"Economic growth is one way to put it. A Dyson sphere is a lot of economic growth. But another way to put it is that you will have, in potentially a very short period of time...","startTime":3341.34,"endTime":3351.177368421053,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"경제성장을 우주 규모 비유로 확장해 강렬함."},{"segmentIndex":41,"text":"AI being deployed incrementally and in advance. One very difficult thing about AI is that we are","startTime":3382.78,"endTime":3394.799411764706,"durationSeconds":12,"level":"C1","overallScore":7.8,"rationale":"점진 배치의 중요성을 핵심 원칙으로 제시함."},{"segmentIndex":54,"text":"You’ve got to be showing the thing. You’ve got to be showing the thing.","startTime":3492.86,"endTime":3496.7814285714285,"durationSeconds":4,"level":"B2","overallScore":7.6,"rationale":"보여줘야 한다는 원칙을 강하게 압축했다."},{"segmentIndex":45,"text":"care for sentient life in a very single-minded way—we might not like the results. That's really","startTime":3923.34,"endTime":3929.5024999999996,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"위험의 본질을 직설적으로 압축함."},{"segmentIndex":55,"text":"The \"It\" we don't know how to build, and a lot hinges on understanding reliable","startTime":3989.98,"endTime":3998.964,"durationSeconds":9,"level":"C1","overallScore":7.4,"rationale":"핵심 병목을 압축적으로 제시함."},{"segmentIndex":7,"text":"still guiding our actions so strongly is an example of alignment success.","startTime":4256.0599999999995,"endTime":4263.613333333334,"durationSeconds":8,"level":"C1","overallScore":7.2,"rationale":""},{"segmentIndex":13,"text":"I think it's actually really mysterious how evolution encodes high-level desires.","startTime":4296.46,"endTime":4306.887272727273,"durationSeconds":10,"level":"C1","overallScore":6.8,"rationale":""}],"generatedAt":"2026-06-24T23:52:58.547Z","keyClipsTotalSec":2154},{"videoId":"aR20FWCCjAs","chunkIndex":2,"totalChunks":10,"title":"Ilya Sutskever – We're moving from the age of scaling to the age of research — Part 3 of 10","thumbnail":"https://i.ytimg.com/vi/aR20FWCCjAs/maxresdefault.jpg","duration":5763,"uploader":"Dwarkesh Patel","youtubeUrl":"https://www.youtube.com/watch?v=aR20FWCCjAs","keywords":["artificial-intelligence","machine-learning","reinforcement-learning","pre-training","scaling","research","generalization","sample-efficiency","evolution","llm"],"normalizedKeywords":["기술 트렌드","교육"],"targetAudience":[{"who":"AI 연구자","why":"스케일링 이후 어떤 연구 문제가 본질인지 정리해줌"},{"who":"엔지니어","why":"RL과 일반화의 병목을 어떻게 봐야 하는지 힌트를 줌"},{"who":"학자","why":"인간 학습의 샘플 효율과 진화적 기원을 비교해 생각하게 함"}],"normalizedAudience":["리서처·학자","엔지니어·개발자"],"summary":"이 영상은 AI 발전이 단순한 '스케일링'의 시대를 지나 다시 '연구'의 시대로 돌아가고 있다고 주장한다. Ilya Sutskever는 pre-training이 강력한 성공 공식을 제공했지만 데이터와 효율의 한계에 부딪히고 있으며, RL도 결국 더 많은 compute를 쓰는 또 다른 형태의 최적화일 뿐이라고 본다.\n\n그가 가장 중요하게 보는 문제는 모델의 일반화 성능이 인간보다 훨씬 떨어진다는 점이다. 왜 모델은 인간보다 더 많은 샘플이 필요하고, 왜 인간처럼 자연스럽게 가르치기 어려운지, 그리고 이 차이가 진정한 학습 구조의 차이인지 진화적 prior 때문인지 탐구한다. 전체적으로 이 영상은 '다음 세대 AI의 핵심은 더 큰 규모가 아니라 더 나은 학습 원리'라는 메시지를 밀도 있게 전달한다.","insights":["스케일링의 한계가 오면 다시 연구가 핵심이 된다.","RL도 본질적으로는 compute를 더 쓰는 다른 recipe다.","AI의 병목은 성능보다 일반화와 샘플 효율이다.","인간의 학습력은 단순한 prior보다 더 근본적일 수 있다.","가르치기 쉬운 모델이 아니라 잘 일반화하는 모델이 중요하다."],"keyClips":[{"clipId":"aR20FWCCjAs:c2:1-12","startSegmentIndex":1,"endSegmentIndex":12,"startTime":1202.14,"endTime":1284.297857142857,"durationSeconds":82.2,"preview":"스케일링의 끝","mustSee":false},{"clipId":"aR20FWCCjAs:c2:13-24","startSegmentIndex":13,"endSegmentIndex":24,"startTime":1284.0600000000002,"endTime":1359.2123529411765,"durationSeconds":75.2,"preview":"연구 시대로의 회귀","mustSee":false},{"clipId":"aR20FWCCjAs:c2:25-38","startSegmentIndex":25,"endSegmentIndex":38,"startTime":1360.14,"endTime":1476.6042105263157,"durationSeconds":116.5,"preview":"RL과 새로운 레시피","mustSee":true},{"clipId":"aR20FWCCjAs:c2:39-51","startSegmentIndex":39,"endSegmentIndex":51,"startTime":1478.14,"endTime":1566.513846153846,"durationSeconds":88.4,"preview":"핵심은 일반화","mustSee":false},{"clipId":"aR20FWCCjAs:c2:52-66","startSegmentIndex":52,"endSegmentIndex":66,"startTime":1566.5400000000002,"endTime":1680.4882352941179,"durationSeconds":113.9,"preview":"진화와 인간의 prior","mustSee":false},{"clipId":"aR20FWCCjAs:c2:67-82","startSegmentIndex":67,"endSegmentIndex":82,"startTime":1680.3000000000002,"endTime":1802.5133333333333,"durationSeconds":122.2,"preview":"더 나은 학습 능력","mustSee":true}],"curatedSegments":[{"segmentIndex":51,"text":"of environments, but to figure out an approach which lets you learn from one environment and","startTime":359.17999999999995,"endTime":365.19,"durationSeconds":6,"level":"C1","overallScore":7.8,"rationale":"전이 학습의 핵심 아이디어를 제시함."},{"segmentIndex":59,"text":"Right. I think that's basically what's going on. The models are much more like the first student, but even more. Because then we say, the model should","startTime":416.7,"endTime":424.48999999999995,"durationSeconds":8,"level":"B2","overallScore":7.8,"rationale":"비유를 핵심 주장으로 연결해 통찰적임."},{"segmentIndex":40,"text":"The value function lets you short-circuit the wait until the very end.","startTime":908.86,"endTime":915.0999999999999,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"가치함수 효용을 압축한 핵심 문장이다."},{"segmentIndex":43,"text":"As soon as you conclude this, you could already get a reward signal a thousand timesteps previously, when you decided to pursue down this path.","startTime":934.22,"endTime":943.6883333333334,"durationSeconds":9,"level":"C1","overallScore":7.6,"rationale":"조기 보상 아이디어가 매우 선명하다."},{"segmentIndex":56,"text":"it's impressive that they have this much utility while still being rather simple to understand.","startTime":1026.0600000000002,"endTime":1036.3016666666667,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"단순성과 효용의 대비가 통찰적이다."},{"segmentIndex":1,"text":"The big breakthrough of pre-training is the realization that this recipe is good.","startTime":1202.14,"endTime":1208.1034615384617,"durationSeconds":6,"level":"B2","overallScore":7.2,"rationale":""},{"segmentIndex":44,"text":"obvious. That seems like a very fundamental thing. So this is the crux: generalization. There are two","startTime":1508.46,"endTime":1518.3632352941174,"durationSeconds":10,"level":"B2","overallScore":6.6,"rationale":""},{"segmentIndex":76,"text":"But if people exhibit great ability, reliability, robustness, and ability to learn in a domain that","startTime":1747.1000000000001,"endTime":1756.4675,"durationSeconds":9,"level":"C1","overallScore":7,"rationale":""},{"segmentIndex":34,"text":"something like, \"If ideas are so cheap, how come no one's having any ideas?\"","startTime":2245.1800000000003,"endTime":2250.955,"durationSeconds":6,"level":"B2","overallScore":8,"rationale":"역설적 반문이 강한 통찰과 표현성을 가짐."},{"segmentIndex":49,"text":"research where the end artifact was made safer mostly through just thinking about","startTime":2745.26,"endTime":2753.200384615385,"durationSeconds":8,"level":"C1","overallScore":7.8,"rationale":"안전 개선 방식에 대한 핵심 주장."},{"segmentIndex":55,"text":"But this is a really powerful thing and we don't even know how to conceptualize how people interact with it, what people will do with it. Having gradual access to it seems like a","startTime":2789.6600000000003,"endTime":2799.86,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"불확실한 상호작용 위험을 짚음."},{"segmentIndex":56,"text":"better way to maybe spread out the impact of it and to help people prepare for it.","startTime":2800.78,"endTime":2805.5808823529414,"durationSeconds":5,"level":"B2","overallScore":7.8,"rationale":"점진 접근의 실용적 이유가 분명함."},{"segmentIndex":18,"text":"but picking them all up at the same time, and then amalgamating their learnings, you basically have a model which functionally becomes superintelligent even without any sort","startTime":3125.42,"endTime":3135.449230769231,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"동시 학습 결합이 초지능으로 이어진단 통찰."},{"segmentIndex":37,"text":"Economic growth is one way to put it. A Dyson sphere is a lot of economic growth. But another way to put it is that you will have, in potentially a very short period of time...","startTime":3341.34,"endTime":3351.177368421053,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"경제성장을 우주 규모 비유로 확장해 강렬함."},{"segmentIndex":41,"text":"AI being deployed incrementally and in advance. One very difficult thing about AI is that we are","startTime":3382.78,"endTime":3394.799411764706,"durationSeconds":12,"level":"C1","overallScore":7.8,"rationale":"점진 배치의 중요성을 핵심 원칙으로 제시함."},{"segmentIndex":54,"text":"You’ve got to be showing the thing. You’ve got to be showing the thing.","startTime":3492.86,"endTime":3496.7814285714285,"durationSeconds":4,"level":"B2","overallScore":7.6,"rationale":"보여줘야 한다는 원칙을 강하게 압축했다."},{"segmentIndex":45,"text":"care for sentient life in a very single-minded way—we might not like the results. That's really","startTime":3923.34,"endTime":3929.5024999999996,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"위험의 본질을 직설적으로 압축함."},{"segmentIndex":55,"text":"The \"It\" we don't know how to build, and a lot hinges on understanding reliable","startTime":3989.98,"endTime":3998.964,"durationSeconds":9,"level":"C1","overallScore":7.4,"rationale":"핵심 병목을 압축적으로 제시함."},{"segmentIndex":7,"text":"still guiding our actions so strongly is an example of alignment success.","startTime":4256.0599999999995,"endTime":4263.613333333334,"durationSeconds":8,"level":"C1","overallScore":7.2,"rationale":""},{"segmentIndex":13,"text":"I think it's actually really mysterious how evolution encodes high-level desires.","startTime":4296.46,"endTime":4306.887272727273,"durationSeconds":10,"level":"C1","overallScore":6.8,"rationale":""}],"generatedAt":"2026-06-24T23:53:23.019Z","keyClipsTotalSec":2154},{"videoId":"aR20FWCCjAs","chunkIndex":3,"totalChunks":10,"title":"Ilya Sutskever – We're moving from the age of scaling to the age of research — Part 4 of 10","thumbnail":"https://i.ytimg.com/vi/aR20FWCCjAs/maxresdefault.jpg","duration":5763,"uploader":"Dwarkesh Patel","youtubeUrl":"https://www.youtube.com/watch?v=aR20FWCCjAs","keywords":["machine-learning","ai-research","scaling-laws","compute","deep-learning","neural-networks","reasoning","robustness","research-culture"],"normalizedKeywords":["엔지니어링","기술 트렌드","교육"],"targetAudience":[{"who":"AI 연구자","why":"스케일링 시대 이후의 연구 방향과 학습 원리를 직접 다룬다"},{"who":"엔지니어","why":"모델 학습의 병목이 compute와 아이디어라는 관점을 얻을 수 있다"},{"who":"학생·주니어","why":"딥러닝 연구가 어떻게 진화했고 무엇이 다음 질문인지 파악하기 좋다"}],"normalizedAudience":["리서처·학자","엔지니어·개발자","학생·주니어"],"summary":"이 대화는 인간이 적은 샘플만으로도 빠르게 배우고 스스로 교정하는 이유를, ‘내부 가치함수(value function)’와 높은 견고성에서 찾는다. 특히 10대 운전자의 학습처럼 외부 교사 없이도 즉각적인 자기평가가 가능하다는 점을 들어, 이런 성질을 기계학습에 어떻게 옮길 수 있을지 논의한다.\n\n이후에는 ‘scaling의 시대’가 끝나고 ‘research의 시대’가 온다는 맥락에서, 연구가 정말로 거대한 컴퓨트를 항상 요구하는지 되묻는다. 컴퓨트는 여전히 중요하지만, 진짜 병목은 아이디어와 이를 증명하는 능력이며, 과거에는 계산 자원 부족 때문에 좋은 아이디어가 묻혔고 이제는 반대로 아이디어가 부족해졌다는 관점을 제시한다. 결국 이 영상은 AI 발전의 다음 단계가 단순한 규모 경쟁이 아니라, 새로운 원리와 실험적 검증을 다시 발명하는 연구 문화로 이동해야 함을 강조한다.","insights":["인간 학습의 핵심은 외부 교사보다 내부 평가 신호다.","견고한 자기평가가 있어야 적은 경험으로도 빠르게 교정한다.","스케일링은 검증을 쉽게 했지만 아이디어 다양성은 줄였다.","연구의 병목은 컴퓨트만이 아니라 아이디어와 구현력이다.","최고 성능엔 큰 컴퓨트가 유리하지만, 아이디어 증명엔 필수는 아니다."],"keyClips":[{"clipId":"aR20FWCCjAs:c3:1-9","startSegmentIndex":1,"endSegmentIndex":9,"startTime":1802.6200000000001,"endTime":1877.66,"durationSeconds":75,"preview":"인간의 빠른 자기교정","mustSee":false},{"clipId":"aR20FWCCjAs:c3:10-19","startSegmentIndex":10,"endSegmentIndex":19,"startTime":1877.42,"endTime":1952.080769230769,"durationSeconds":74.7,"preview":"인간식 학습의 수수께끼","mustSee":false},{"clipId":"aR20FWCCjAs:c3:20-27","startSegmentIndex":20,"endSegmentIndex":27,"startTime":1952.46,"endTime":2199.478181818182,"durationSeconds":247,"preview":"연구시대의 분위기","mustSee":true},{"clipId":"aR20FWCCjAs:c3:28-39","startSegmentIndex":28,"endSegmentIndex":39,"startTime":2200.06,"endTime":2290.43,"durationSeconds":90.4,"preview":"스케일링의 부작용","mustSee":false},{"clipId":"aR20FWCCjAs:c3:40-54","startSegmentIndex":40,"endSegmentIndex":54,"startTime":2290.38,"endTime":2404.906470588235,"durationSeconds":114.5,"preview":"컴퓨트와 아이디어","mustSee":true}],"curatedSegments":[{"segmentIndex":51,"text":"of environments, but to figure out an approach which lets you learn from one environment and","startTime":359.17999999999995,"endTime":365.19,"durationSeconds":6,"level":"C1","overallScore":7.8,"rationale":"전이 학습의 핵심 아이디어를 제시함."},{"segmentIndex":59,"text":"Right. I think that's basically what's going on. The models are much more like the first student, but even more. Because then we say, the model should","startTime":416.7,"endTime":424.48999999999995,"durationSeconds":8,"level":"B2","overallScore":7.8,"rationale":"비유를 핵심 주장으로 연결해 통찰적임."},{"segmentIndex":40,"text":"The value function lets you short-circuit the wait until the very end.","startTime":908.86,"endTime":915.0999999999999,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"가치함수 효용을 압축한 핵심 문장이다."},{"segmentIndex":43,"text":"As soon as you conclude this, you could already get a reward signal a thousand timesteps previously, when you decided to pursue down this path.","startTime":934.22,"endTime":943.6883333333334,"durationSeconds":9,"level":"C1","overallScore":7.6,"rationale":"조기 보상 아이디어가 매우 선명하다."},{"segmentIndex":56,"text":"it's impressive that they have this much utility while still being rather simple to understand.","startTime":1026.0600000000002,"endTime":1036.3016666666667,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"단순성과 효용의 대비가 통찰적이다."},{"segmentIndex":1,"text":"The big breakthrough of pre-training is the realization that this recipe is good.","startTime":1202.14,"endTime":1208.1034615384617,"durationSeconds":6,"level":"B2","overallScore":7.2,"rationale":""},{"segmentIndex":44,"text":"obvious. That seems like a very fundamental thing. So this is the crux: generalization. There are two","startTime":1508.46,"endTime":1518.3632352941174,"durationSeconds":10,"level":"B2","overallScore":6.6,"rationale":""},{"segmentIndex":76,"text":"But if people exhibit great ability, reliability, robustness, and ability to learn in a domain that","startTime":1747.1000000000001,"endTime":1756.4675,"durationSeconds":9,"level":"C1","overallScore":7,"rationale":""},{"segmentIndex":34,"text":"something like, \"If ideas are so cheap, how come no one's having any ideas?\"","startTime":2245.1800000000003,"endTime":2250.955,"durationSeconds":6,"level":"B2","overallScore":8,"rationale":"역설적 반문이 강한 통찰과 표현성을 가짐."},{"segmentIndex":49,"text":"research where the end artifact was made safer mostly through just thinking about","startTime":2745.26,"endTime":2753.200384615385,"durationSeconds":8,"level":"C1","overallScore":7.8,"rationale":"안전 개선 방식에 대한 핵심 주장."},{"segmentIndex":55,"text":"But this is a really powerful thing and we don't even know how to conceptualize how people interact with it, what people will do with it. Having gradual access to it seems like a","startTime":2789.6600000000003,"endTime":2799.86,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"불확실한 상호작용 위험을 짚음."},{"segmentIndex":56,"text":"better way to maybe spread out the impact of it and to help people prepare for it.","startTime":2800.78,"endTime":2805.5808823529414,"durationSeconds":5,"level":"B2","overallScore":7.8,"rationale":"점진 접근의 실용적 이유가 분명함."},{"segmentIndex":18,"text":"but picking them all up at the same time, and then amalgamating their learnings, you basically have a model which functionally becomes superintelligent even without any sort","startTime":3125.42,"endTime":3135.449230769231,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"동시 학습 결합이 초지능으로 이어진단 통찰."},{"segmentIndex":37,"text":"Economic growth is one way to put it. A Dyson sphere is a lot of economic growth. But another way to put it is that you will have, in potentially a very short period of time...","startTime":3341.34,"endTime":3351.177368421053,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"경제성장을 우주 규모 비유로 확장해 강렬함."},{"segmentIndex":41,"text":"AI being deployed incrementally and in advance. One very difficult thing about AI is that we are","startTime":3382.78,"endTime":3394.799411764706,"durationSeconds":12,"level":"C1","overallScore":7.8,"rationale":"점진 배치의 중요성을 핵심 원칙으로 제시함."},{"segmentIndex":54,"text":"You’ve got to be showing the thing. You’ve got to be showing the thing.","startTime":3492.86,"endTime":3496.7814285714285,"durationSeconds":4,"level":"B2","overallScore":7.6,"rationale":"보여줘야 한다는 원칙을 강하게 압축했다."},{"segmentIndex":45,"text":"care for sentient life in a very single-minded way—we might not like the results. That's really","startTime":3923.34,"endTime":3929.5024999999996,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"위험의 본질을 직설적으로 압축함."},{"segmentIndex":55,"text":"The \"It\" we don't know how to build, and a lot hinges on understanding reliable","startTime":3989.98,"endTime":3998.964,"durationSeconds":9,"level":"C1","overallScore":7.4,"rationale":"핵심 병목을 압축적으로 제시함."},{"segmentIndex":7,"text":"still guiding our actions so strongly is an example of alignment success.","startTime":4256.0599999999995,"endTime":4263.613333333334,"durationSeconds":8,"level":"C1","overallScore":7.2,"rationale":""},{"segmentIndex":13,"text":"I think it's actually really mysterious how evolution encodes high-level desires.","startTime":4296.46,"endTime":4306.887272727273,"durationSeconds":10,"level":"C1","overallScore":6.8,"rationale":""}],"generatedAt":"2026-06-24T23:53:55.397Z","keyClipsTotalSec":2154},{"videoId":"aR20FWCCjAs","chunkIndex":4,"totalChunks":10,"title":"Ilya Sutskever – We're moving from the age of scaling to the age of research — Part 5 of 10","thumbnail":"https://i.ytimg.com/vi/aR20FWCCjAs/maxresdefault.jpg","duration":5763,"uploader":"Dwarkesh Patel","youtubeUrl":"https://www.youtube.com/watch?v=aR20FWCCjAs","keywords":["ai-research","superintelligence","compute","inference","pre-training","agi","continual-learning","ai-safety","startup-strategy","machine-learning"],"normalizedKeywords":["기술 트렌드","비즈니스·전략","교육"],"targetAudience":[{"who":"AI 연구자","why":"연구 컴퓨트 배분과 실험 규모에 대한 현실적 관점을 얻을 수 있음"},{"who":"창업자","why":"제품보다 연구를 우선하는 조직 전략과 수익화 보류 논리를 볼 수 있음"},{"who":"투자자","why":"전면적 경쟁보다 연구 집중 전략이 어떤 가치 논리를 갖는지 이해할 수 있음"},{"who":"학생·주니어","why":"AGI, pre-training, continual learning 같은 개념이 사고를 어떻게 바꾸는지 배울 수 있음"}],"normalizedAudience":["리서처·학자","창업자·스타트업","투자자·VC"],"summary":"이 구간은 SSI가 다른 프런티어 랩보다 연구용 컴퓨트가 생각보다 충분하다고 주장하는 부분에서 시작해, 왜 지금은 수익화보다 연구에 집중하는지 설명한다. 핵심 논리는 대부분의 자금이 추론과 제품 운영에 소모되는 반면, 연구에 실제로 남는 자원은 훨씬 작아지고, 중요한 것은 절대 규모가 아니라 무엇을 위해 어떻게 쓰느냐라는 점이다.\n\n이어지는 논의는 SSI가 왜 superintelligence를 바로 목표로 하는지, 그리고 그 전략의 장단점을 균형 있게 다룬다. 시장 경쟁에서 벗어나 연구에 몰입할 수 있다는 장점이 있는 반면, 세상에 강력한 AI를 더 일찍 보여줘야 안전성과 사회적 적응이 가능하다는 반론도 제시한다. 마지막으로는 AGI와 pre-training이라는 용어가 사고를 과도하게 규정해왔으며, 인간조차도 모든 것을 아는 AGI가 아니라는 점에서 기존 프레임을 다시 생각해야 한다는 메시지로 확장된다.","insights":["연구의 승부는 총자금보다 '실제로 남는 컴퓨트'에 달려 있다.","수익화가 붙는 순간 연구 자원은 제품·운영으로 급격히 분산된다.","직접 초지능을 노리는 전략은 경쟁 회피와 공개효과를 동시에 갖는다.","강력한 AI는 통제만이 아니라 실제 배포를 통해 더 안전해질 수 있다.","AGI·pre-training 같은 단어가 기술의 방향감각을 과도하게 고정한다."],"keyClips":[{"clipId":"aR20FWCCjAs:c4:4-16","startSegmentIndex":4,"endSegmentIndex":16,"startTime":2430.2200000000003,"endTime":2522.8675,"durationSeconds":92.6,"preview":"연구컴퓨트는 충분하다","mustSee":false},{"clipId":"aR20FWCCjAs:c4:17-22","startSegmentIndex":17,"endSegmentIndex":22,"startTime":2522.78,"endTime":2561.12,"durationSeconds":38.3,"preview":"실험비와 분산","mustSee":false},{"clipId":"aR20FWCCjAs:c4:23-31","startSegmentIndex":23,"endSegmentIndex":31,"startTime":2561.02,"endTime":2626.5775,"durationSeconds":65.6,"preview":"수익화보다 연구","mustSee":false},{"clipId":"aR20FWCCjAs:c4:32-47","startSegmentIndex":32,"endSegmentIndex":47,"startTime":2626.38,"endTime":2736.0233333333335,"durationSeconds":109.6,"preview":"직행 초지능의 논리","mustSee":false},{"clipId":"aR20FWCCjAs:c4:48-59","startSegmentIndex":48,"endSegmentIndex":59,"startTime":2736.2200000000003,"endTime":2831.8072222222218,"durationSeconds":95.6,"preview":"배포가 안전을 만든다","mustSee":true},{"clipId":"aR20FWCCjAs:c4:60-76","startSegmentIndex":60,"endSegmentIndex":76,"startTime":2831.58,"endTime":2969.435714285714,"durationSeconds":137.9,"preview":"용어가 생각을 만든다","mustSee":false},{"clipId":"aR20FWCCjAs:c4:77-80","startSegmentIndex":77,"endSegmentIndex":80,"startTime":2969.6600000000003,"endTime":3000.5480000000002,"durationSeconds":30.9,"preview":"인간도 AGI는 아니다","mustSee":false}],"curatedSegments":[{"segmentIndex":51,"text":"of environments, but to figure out an approach which lets you learn from one environment and","startTime":359.17999999999995,"endTime":365.19,"durationSeconds":6,"level":"C1","overallScore":7.8,"rationale":"전이 학습의 핵심 아이디어를 제시함."},{"segmentIndex":59,"text":"Right. I think that's basically what's going on. The models are much more like the first student, but even more. Because then we say, the model should","startTime":416.7,"endTime":424.48999999999995,"durationSeconds":8,"level":"B2","overallScore":7.8,"rationale":"비유를 핵심 주장으로 연결해 통찰적임."},{"segmentIndex":40,"text":"The value function lets you short-circuit the wait until the very end.","startTime":908.86,"endTime":915.0999999999999,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"가치함수 효용을 압축한 핵심 문장이다."},{"segmentIndex":43,"text":"As soon as you conclude this, you could already get a reward signal a thousand timesteps previously, when you decided to pursue down this path.","startTime":934.22,"endTime":943.6883333333334,"durationSeconds":9,"level":"C1","overallScore":7.6,"rationale":"조기 보상 아이디어가 매우 선명하다."},{"segmentIndex":56,"text":"it's impressive that they have this much utility while still being rather simple to understand.","startTime":1026.0600000000002,"endTime":1036.3016666666667,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"단순성과 효용의 대비가 통찰적이다."},{"segmentIndex":1,"text":"The big breakthrough of pre-training is the realization that this recipe is good.","startTime":1202.14,"endTime":1208.1034615384617,"durationSeconds":6,"level":"B2","overallScore":7.2,"rationale":""},{"segmentIndex":44,"text":"obvious. That seems like a very fundamental thing. So this is the crux: generalization. There are two","startTime":1508.46,"endTime":1518.3632352941174,"durationSeconds":10,"level":"B2","overallScore":6.6,"rationale":""},{"segmentIndex":76,"text":"But if people exhibit great ability, reliability, robustness, and ability to learn in a domain that","startTime":1747.1000000000001,"endTime":1756.4675,"durationSeconds":9,"level":"C1","overallScore":7,"rationale":""},{"segmentIndex":34,"text":"something like, \"If ideas are so cheap, how come no one's having any ideas?\"","startTime":2245.1800000000003,"endTime":2250.955,"durationSeconds":6,"level":"B2","overallScore":8,"rationale":"역설적 반문이 강한 통찰과 표현성을 가짐."},{"segmentIndex":49,"text":"research where the end artifact was made safer mostly through just thinking about","startTime":2745.26,"endTime":2753.200384615385,"durationSeconds":8,"level":"C1","overallScore":7.8,"rationale":"안전 개선 방식에 대한 핵심 주장."},{"segmentIndex":55,"text":"But this is a really powerful thing and we don't even know how to conceptualize how people interact with it, what people will do with it. Having gradual access to it seems like a","startTime":2789.6600000000003,"endTime":2799.86,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"불확실한 상호작용 위험을 짚음."},{"segmentIndex":56,"text":"better way to maybe spread out the impact of it and to help people prepare for it.","startTime":2800.78,"endTime":2805.5808823529414,"durationSeconds":5,"level":"B2","overallScore":7.8,"rationale":"점진 접근의 실용적 이유가 분명함."},{"segmentIndex":18,"text":"but picking them all up at the same time, and then amalgamating their learnings, you basically have a model which functionally becomes superintelligent even without any sort","startTime":3125.42,"endTime":3135.449230769231,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"동시 학습 결합이 초지능으로 이어진단 통찰."},{"segmentIndex":37,"text":"Economic growth is one way to put it. A Dyson sphere is a lot of economic growth. But another way to put it is that you will have, in potentially a very short period of time...","startTime":3341.34,"endTime":3351.177368421053,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"경제성장을 우주 규모 비유로 확장해 강렬함."},{"segmentIndex":41,"text":"AI being deployed incrementally and in advance. One very difficult thing about AI is that we are","startTime":3382.78,"endTime":3394.799411764706,"durationSeconds":12,"level":"C1","overallScore":7.8,"rationale":"점진 배치의 중요성을 핵심 원칙으로 제시함."},{"segmentIndex":54,"text":"You’ve got to be showing the thing. You’ve got to be showing the thing.","startTime":3492.86,"endTime":3496.7814285714285,"durationSeconds":4,"level":"B2","overallScore":7.6,"rationale":"보여줘야 한다는 원칙을 강하게 압축했다."},{"segmentIndex":45,"text":"care for sentient life in a very single-minded way—we might not like the results. That's really","startTime":3923.34,"endTime":3929.5024999999996,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"위험의 본질을 직설적으로 압축함."},{"segmentIndex":55,"text":"The \"It\" we don't know how to build, and a lot hinges on understanding reliable","startTime":3989.98,"endTime":3998.964,"durationSeconds":9,"level":"C1","overallScore":7.4,"rationale":"핵심 병목을 압축적으로 제시함."},{"segmentIndex":7,"text":"still guiding our actions so strongly is an example of alignment success.","startTime":4256.0599999999995,"endTime":4263.613333333334,"durationSeconds":8,"level":"C1","overallScore":7.2,"rationale":""},{"segmentIndex":13,"text":"I think it's actually really mysterious how evolution encodes high-level desires.","startTime":4296.46,"endTime":4306.887272727273,"durationSeconds":10,"level":"C1","overallScore":6.8,"rationale":""}],"generatedAt":"2026-06-24T23:54:16.241Z","keyClipsTotalSec":2154},{"videoId":"aR20FWCCjAs","chunkIndex":5,"totalChunks":10,"title":"Ilya Sutskever – We're moving from the age of scaling to the age of research — Part 6 of 10","thumbnail":"https://i.ytimg.com/vi/aR20FWCCjAs/maxresdefault.jpg","duration":5763,"uploader":"Dwarkesh Patel","youtubeUrl":"https://www.youtube.com/watch?v=aR20FWCCjAs","keywords":["artificial-intelligence","agi","superintelligence","machine-learning","economic-growth","ai-safety","continual-learning","policy","future-of-work"],"normalizedKeywords":["기술 트렌드","비즈니스·전략","리더십·매니지먼트"],"targetAudience":[{"who":"AI 연구자","why":"지속학습, 배포, 능력 확장의 관점에서 AGI를 다시 정의한다."},{"who":"창업자","why":"AI가 경제와 조직을 어떻게 바꾸는지 전략적으로 이해할 수 있다."},{"who":"정책 담당자","why":"AI 확산에 따른 규제, 안전 협력, 공공 대응의 필요성을 다룬다."}],"normalizedAudience":["리서처·학자","창업자·스타트업","지식노동자 일반"],"summary":"이 대화는 안전한 슈퍼지능을 어떻게 정의하고 배포할 것인가를 중심으로 진행된다. 핵심 주장은, 미래의 AI는 한 번 완성된 '끝난 지능'이 아니라 세계 속에서 계속 배우는 시스템이어야 하며, 그래서 배포 자체가 일종의 학습 과정이 될 것이라는 점이다. 이런 관점에서는 AI가 여러 직무를 동시에 학습하고, 여러 인스턴스의 학습을 합쳐 기능적으로 초지능에 가까워질 수 있으며, 그 결과 급격한 경제 성장과 생산성 변화가 가능해진다.\n\n동시에 화자는 AI의 진짜 문제는 '지능' 자체보다 '힘(power)'이라고 말한다. 그래서 AI가 더 강력해질수록 사람들은 상상만으로는 감을 잡기 어려워지고, 기업·정부·대중이 실제로 AI를 보여주며 대응하는 방향으로 행동이 바뀔 것이라고 본다. 또한 프런티어 기업 간의 안전 협력, 정부 개입, 점진적 배포의 중요성을 강조하며, 아직 보이지 않는 미래 시스템을 현실에서 어떻게 인식시키고 관리할지가 핵심 과제라고 주장한다.","insights":["슈퍼지능은 완성품보다 계속 배우는 시스템에 가깝다.","배포는 종료가 아니라 현장 학습의 시작이다.","여러 인스턴스의 학습이 합쳐지면 기능적 초지능이 된다.","AI의 본질적 문제는 지능이 아니라 권력의 크기다.","기술이 너무 강해지면 사람은 '보여줘야' 반응한다."],"keyClips":[{"clipId":"aR20FWCCjAs:c5:1-19","startSegmentIndex":1,"endSegmentIndex":19,"startTime":2999.98,"endTime":3145.2,"durationSeconds":145.2,"preview":"배우는 슈퍼지능","mustSee":true},{"clipId":"aR20FWCCjAs:c5:20-38","startSegmentIndex":20,"endSegmentIndex":38,"startTime":3144.94,"endTime":3361.090333333333,"durationSeconds":216.2,"preview":"확산이 만드는 폭발","mustSee":false},{"clipId":"aR20FWCCjAs:c5:40-54","startSegmentIndex":40,"endSegmentIndex":54,"startTime":3372.06,"endTime":3496.7814285714285,"durationSeconds":124.7,"preview":"보여줘야 느낀다","mustSee":false},{"clipId":"aR20FWCCjAs:c5:57-68","startSegmentIndex":57,"endSegmentIndex":68,"startTime":3511.34,"endTime":3607.2233333333334,"durationSeconds":95.9,"preview":"안전 협력의 시작","mustSee":false}],"curatedSegments":[{"segmentIndex":51,"text":"of environments, but to figure out an approach which lets you learn from one environment and","startTime":359.17999999999995,"endTime":365.19,"durationSeconds":6,"level":"C1","overallScore":7.8,"rationale":"전이 학습의 핵심 아이디어를 제시함."},{"segmentIndex":59,"text":"Right. I think that's basically what's going on. The models are much more like the first student, but even more. Because then we say, the model should","startTime":416.7,"endTime":424.48999999999995,"durationSeconds":8,"level":"B2","overallScore":7.8,"rationale":"비유를 핵심 주장으로 연결해 통찰적임."},{"segmentIndex":40,"text":"The value function lets you short-circuit the wait until the very end.","startTime":908.86,"endTime":915.0999999999999,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"가치함수 효용을 압축한 핵심 문장이다."},{"segmentIndex":43,"text":"As soon as you conclude this, you could already get a reward signal a thousand timesteps previously, when you decided to pursue down this path.","startTime":934.22,"endTime":943.6883333333334,"durationSeconds":9,"level":"C1","overallScore":7.6,"rationale":"조기 보상 아이디어가 매우 선명하다."},{"segmentIndex":56,"text":"it's impressive that they have this much utility while still being rather simple to understand.","startTime":1026.0600000000002,"endTime":1036.3016666666667,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"단순성과 효용의 대비가 통찰적이다."},{"segmentIndex":1,"text":"The big breakthrough of pre-training is the realization that this recipe is good.","startTime":1202.14,"endTime":1208.1034615384617,"durationSeconds":6,"level":"B2","overallScore":7.2,"rationale":""},{"segmentIndex":44,"text":"obvious. That seems like a very fundamental thing. So this is the crux: generalization. There are two","startTime":1508.46,"endTime":1518.3632352941174,"durationSeconds":10,"level":"B2","overallScore":6.6,"rationale":""},{"segmentIndex":76,"text":"But if people exhibit great ability, reliability, robustness, and ability to learn in a domain that","startTime":1747.1000000000001,"endTime":1756.4675,"durationSeconds":9,"level":"C1","overallScore":7,"rationale":""},{"segmentIndex":34,"text":"something like, \"If ideas are so cheap, how come no one's having any ideas?\"","startTime":2245.1800000000003,"endTime":2250.955,"durationSeconds":6,"level":"B2","overallScore":8,"rationale":"역설적 반문이 강한 통찰과 표현성을 가짐."},{"segmentIndex":49,"text":"research where the end artifact was made safer mostly through just thinking about","startTime":2745.26,"endTime":2753.200384615385,"durationSeconds":8,"level":"C1","overallScore":7.8,"rationale":"안전 개선 방식에 대한 핵심 주장."},{"segmentIndex":55,"text":"But this is a really powerful thing and we don't even know how to conceptualize how people interact with it, what people will do with it. Having gradual access to it seems like a","startTime":2789.6600000000003,"endTime":2799.86,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"불확실한 상호작용 위험을 짚음."},{"segmentIndex":56,"text":"better way to maybe spread out the impact of it and to help people prepare for it.","startTime":2800.78,"endTime":2805.5808823529414,"durationSeconds":5,"level":"B2","overallScore":7.8,"rationale":"점진 접근의 실용적 이유가 분명함."},{"segmentIndex":18,"text":"but picking them all up at the same time, and then amalgamating their learnings, you basically have a model which functionally becomes superintelligent even without any sort","startTime":3125.42,"endTime":3135.449230769231,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"동시 학습 결합이 초지능으로 이어진단 통찰."},{"segmentIndex":37,"text":"Economic growth is one way to put it. A Dyson sphere is a lot of economic growth. But another way to put it is that you will have, in potentially a very short period of time...","startTime":3341.34,"endTime":3351.177368421053,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"경제성장을 우주 규모 비유로 확장해 강렬함."},{"segmentIndex":41,"text":"AI being deployed incrementally and in advance. One very difficult thing about AI is that we are","startTime":3382.78,"endTime":3394.799411764706,"durationSeconds":12,"level":"C1","overallScore":7.8,"rationale":"점진 배치의 중요성을 핵심 원칙으로 제시함."},{"segmentIndex":54,"text":"You’ve got to be showing the thing. You’ve got to be showing the thing.","startTime":3492.86,"endTime":3496.7814285714285,"durationSeconds":4,"level":"B2","overallScore":7.6,"rationale":"보여줘야 한다는 원칙을 강하게 압축했다."},{"segmentIndex":45,"text":"care for sentient life in a very single-minded way—we might not like the results. That's really","startTime":3923.34,"endTime":3929.5024999999996,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"위험의 본질을 직설적으로 압축함."},{"segmentIndex":55,"text":"The \"It\" we don't know how to build, and a lot hinges on understanding reliable","startTime":3989.98,"endTime":3998.964,"durationSeconds":9,"level":"C1","overallScore":7.4,"rationale":"핵심 병목을 압축적으로 제시함."},{"segmentIndex":7,"text":"still guiding our actions so strongly is an example of alignment success.","startTime":4256.0599999999995,"endTime":4263.613333333334,"durationSeconds":8,"level":"C1","overallScore":7.2,"rationale":""},{"segmentIndex":13,"text":"I think it's actually really mysterious how evolution encodes high-level desires.","startTime":4296.46,"endTime":4306.887272727273,"durationSeconds":10,"level":"C1","overallScore":6.8,"rationale":""}],"generatedAt":"2026-06-24T23:54:32.555Z","keyClipsTotalSec":2154},{"videoId":"aR20FWCCjAs","chunkIndex":6,"totalChunks":10,"title":"Ilya Sutskever – We're moving from the age of scaling to the age of research — Part 7 of 10","thumbnail":"https://i.ytimg.com/vi/aR20FWCCjAs/maxresdefault.jpg","duration":5763,"uploader":"Dwarkesh Patel","youtubeUrl":"https://www.youtube.com/watch?v=aR20FWCCjAs","keywords":["ai-safety","superintelligence","alignment","machine-learning","future-of-ai","ethics","sentient-ai","research"],"normalizedKeywords":["기술 트렌드","비즈니스·전략","교육"],"targetAudience":[{"who":"AI 연구자","why":"초지능 정렬, 일반화, 안전성의 핵심 논점을 직접 다룸"},{"who":"창업자","why":"미래 AI 제품/회사 전략이 어떤 방향으로 가야 하는지 시사함"},{"who":"정책 관심층","why":"초강력 AI의 거버넌스와 장기 균형 문제를 생각하게 함"}],"normalizedAudience":["리서처·학자","창업자·스타트업","지식노동자 일반"],"summary":"이 영상은 초지능 시대에 AI가 실제로 '강해 보이기' 시작하면 업계의 안전 인식과 전략이 어떻게 바뀔지에 대한 Ilya Sutskever의 전망을 중심으로 전개된다. 그는 현재의 AI를 보고 미래를 과소평가하고 있다고 말하며, 앞으로는 여러 개의 매우 강력한 AI가 동시에 등장할 가능성이 높고, 그때는 회사들이 훨씬 더 보수적이고 경계심이 강해질 것이라고 본다.\n\n또한 그는 단순한 자기개선 AI보다 '지각 있는 생명(s sentient life)을 돌보도록 정렬된 AI'가 더 바람직한 목표일 수 있다고 주장한다. 다만 진짜 핵심은 현재 방법들이 결국 한계에 부딪힐 것이며, 신뢰할 수 있는 일반화와 정렬의 취약성을 해결하지 못하면 우리가 기대하는 '그것(It)'은 아직 만들지 못한 상태로 남을 것이라는 점이다. 장기적으로는 강력한 AI가 개인을 돕는 구조가 가능하더라도, 사람은 점점 의사결정의 주체에서 빠져나갈 수 있다는 철학적·정치적 문제까지 건드린다.","insights":["AI가 강해 보이기 시작하면 안전 기준은 급격히 보수화된다.","현재의 AI 발전은 '진짜 그것'에 아직 닿지 못했다.","정렬의 핵심 난제는 가치 학습과 최적화의 취약한 일반화다.","초지능 시대의 목표는 인간만이 아니라 지각 있는 생명 전체를 봐야 한다.","장기적으로 편리한 AI 대리인은 인간의 주체성을 약화시킬 수 있다."],"keyClips":[{"clipId":"aR20FWCCjAs:c6:1-5","startSegmentIndex":1,"endSegmentIndex":5,"startTime":3606.46,"endTime":3642.38,"durationSeconds":35.9,"preview":"AI가 강해 보이는 순간","mustSee":false},{"clipId":"aR20FWCCjAs:c6:6-18","startSegmentIndex":6,"endSegmentIndex":18,"startTime":3642.54,"endTime":3723.785,"durationSeconds":81.2,"preview":"무엇을 만들어야 하나","mustSee":true},{"clipId":"aR20FWCCjAs:c6:19-30","startSegmentIndex":19,"endSegmentIndex":30,"startTime":3723.58,"endTime":3818.7457142857143,"durationSeconds":95.2,"preview":"정렬의 목표를 넓혀라","mustSee":false},{"clipId":"aR20FWCCjAs:c6:31-60","startSegmentIndex":31,"endSegmentIndex":60,"startTime":3818.46,"endTime":4039.1686363636363,"durationSeconds":220.7,"preview":"진짜 난제는 일반화","mustSee":true},{"clipId":"aR20FWCCjAs:c6:61-70","startSegmentIndex":61,"endSegmentIndex":70,"startTime":4039.1,"endTime":4113.057,"durationSeconds":74,"preview":"좋은 AI의 첫 조건","mustSee":false},{"clipId":"aR20FWCCjAs:c6:71-83","startSegmentIndex":71,"endSegmentIndex":83,"startTime":4112.86,"endTime":4206.5700000000015,"durationSeconds":93.7,"preview":"장기 균형의 딜레마","mustSee":false}],"curatedSegments":[{"segmentIndex":51,"text":"of environments, but to figure out an approach which lets you learn from one environment and","startTime":359.17999999999995,"endTime":365.19,"durationSeconds":6,"level":"C1","overallScore":7.8,"rationale":"전이 학습의 핵심 아이디어를 제시함."},{"segmentIndex":59,"text":"Right. I think that's basically what's going on. The models are much more like the first student, but even more. Because then we say, the model should","startTime":416.7,"endTime":424.48999999999995,"durationSeconds":8,"level":"B2","overallScore":7.8,"rationale":"비유를 핵심 주장으로 연결해 통찰적임."},{"segmentIndex":40,"text":"The value function lets you short-circuit the wait until the very end.","startTime":908.86,"endTime":915.0999999999999,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"가치함수 효용을 압축한 핵심 문장이다."},{"segmentIndex":43,"text":"As soon as you conclude this, you could already get a reward signal a thousand timesteps previously, when you decided to pursue down this path.","startTime":934.22,"endTime":943.6883333333334,"durationSeconds":9,"level":"C1","overallScore":7.6,"rationale":"조기 보상 아이디어가 매우 선명하다."},{"segmentIndex":56,"text":"it's impressive that they have this much utility while still being rather simple to understand.","startTime":1026.0600000000002,"endTime":1036.3016666666667,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"단순성과 효용의 대비가 통찰적이다."},{"segmentIndex":1,"text":"The big breakthrough of pre-training is the realization that this recipe is good.","startTime":1202.14,"endTime":1208.1034615384617,"durationSeconds":6,"level":"B2","overallScore":7.2,"rationale":""},{"segmentIndex":44,"text":"obvious. That seems like a very fundamental thing. So this is the crux: generalization. There are two","startTime":1508.46,"endTime":1518.3632352941174,"durationSeconds":10,"level":"B2","overallScore":6.6,"rationale":""},{"segmentIndex":76,"text":"But if people exhibit great ability, reliability, robustness, and ability to learn in a domain that","startTime":1747.1000000000001,"endTime":1756.4675,"durationSeconds":9,"level":"C1","overallScore":7,"rationale":""},{"segmentIndex":34,"text":"something like, \"If ideas are so cheap, how come no one's having any ideas?\"","startTime":2245.1800000000003,"endTime":2250.955,"durationSeconds":6,"level":"B2","overallScore":8,"rationale":"역설적 반문이 강한 통찰과 표현성을 가짐."},{"segmentIndex":49,"text":"research where the end artifact was made safer mostly through just thinking about","startTime":2745.26,"endTime":2753.200384615385,"durationSeconds":8,"level":"C1","overallScore":7.8,"rationale":"안전 개선 방식에 대한 핵심 주장."},{"segmentIndex":55,"text":"But this is a really powerful thing and we don't even know how to conceptualize how people interact with it, what people will do with it. Having gradual access to it seems like a","startTime":2789.6600000000003,"endTime":2799.86,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"불확실한 상호작용 위험을 짚음."},{"segmentIndex":56,"text":"better way to maybe spread out the impact of it and to help people prepare for it.","startTime":2800.78,"endTime":2805.5808823529414,"durationSeconds":5,"level":"B2","overallScore":7.8,"rationale":"점진 접근의 실용적 이유가 분명함."},{"segmentIndex":18,"text":"but picking them all up at the same time, and then amalgamating their learnings, you basically have a model which functionally becomes superintelligent even without any sort","startTime":3125.42,"endTime":3135.449230769231,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"동시 학습 결합이 초지능으로 이어진단 통찰."},{"segmentIndex":37,"text":"Economic growth is one way to put it. A Dyson sphere is a lot of economic growth. But another way to put it is that you will have, in potentially a very short period of time...","startTime":3341.34,"endTime":3351.177368421053,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"경제성장을 우주 규모 비유로 확장해 강렬함."},{"segmentIndex":41,"text":"AI being deployed incrementally and in advance. One very difficult thing about AI is that we are","startTime":3382.78,"endTime":3394.799411764706,"durationSeconds":12,"level":"C1","overallScore":7.8,"rationale":"점진 배치의 중요성을 핵심 원칙으로 제시함."},{"segmentIndex":54,"text":"You’ve got to be showing the thing. You’ve got to be showing the thing.","startTime":3492.86,"endTime":3496.7814285714285,"durationSeconds":4,"level":"B2","overallScore":7.6,"rationale":"보여줘야 한다는 원칙을 강하게 압축했다."},{"segmentIndex":45,"text":"care for sentient life in a very single-minded way—we might not like the results. That's really","startTime":3923.34,"endTime":3929.5024999999996,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"위험의 본질을 직설적으로 압축함."},{"segmentIndex":55,"text":"The \"It\" we don't know how to build, and a lot hinges on understanding reliable","startTime":3989.98,"endTime":3998.964,"durationSeconds":9,"level":"C1","overallScore":7.4,"rationale":"핵심 병목을 압축적으로 제시함."},{"segmentIndex":7,"text":"still guiding our actions so strongly is an example of alignment success.","startTime":4256.0599999999995,"endTime":4263.613333333334,"durationSeconds":8,"level":"C1","overallScore":7.2,"rationale":""},{"segmentIndex":13,"text":"I think it's actually really mysterious how evolution encodes high-level desires.","startTime":4296.46,"endTime":4306.887272727273,"durationSeconds":10,"level":"C1","overallScore":6.8,"rationale":""}],"generatedAt":"2026-06-24T23:55:01.763Z","keyClipsTotalSec":2154},{"videoId":"aR20FWCCjAs","chunkIndex":7,"totalChunks":10,"title":"Ilya Sutskever – We're moving from the age of scaling to the age of research — Part 8 of 10","thumbnail":"https://i.ytimg.com/vi/aR20FWCCjAs/maxresdefault.jpg","duration":5763,"uploader":"Dwarkesh Patel","youtubeUrl":"https://www.youtube.com/watch?v=aR20FWCCjAs","keywords":["ai-safety","alignment","neuralink","neuroscience","evolution","genome","startup","research","openai","social-behavior"],"normalizedKeywords":["기술 트렌드","비즈니스·전략","교육"],"targetAudience":[{"who":"AI 연구자","why":"정렬 문제와 일반화, 인간 욕구의 형성 방식에 대한 관점을 얻을 수 있음"},{"who":"창업자","why":"SSI의 차별화 전략과 연구 중심 조직의 방향을 이해하는 데 도움됨"},{"who":"학생","why":"진화, 뇌, AI 정렬을 연결하는 고난도 토론을 영어로 접할 수 있음"}],"normalizedAudience":["리서처·학자","창업자·스타트업","학생·주니어"],"summary":"이 구간은 AI 안전과 인간 정렬(alignment)을 둘러싼 아주 근본적인 문제를 다룬다. 화자는 AI가 인간보다 더 똑똑해지는 상황에서 단순한 통제보다, 인간이 AI와 이해를 공유하는 방향(예: Neuralink류의 결합)이 하나의 해법일 수 있다고 말한다. 이어서 인간의 사회적 욕구와 감정이 진화적으로 어떻게 고차원적으로 '설계'되었는지에 대한 미스터리를 길게 논의하며, 이것이 AI 정렬 문제와 닮아 있다고 본다.\n\n후반부에서는 SSI가 무엇을 다르게 하려는지 묻는 질문에 대해, 아직 검증 중인 연구 가설들을 추적하는 'age of research' 회사라고 설명한다. 또 공동창업자의 Meta 이탈에 대한 의문에는, 당시 회사가 높은 밸류에이션으로 자금조달 중이었고 Meta의 인수 제안을 거절한 맥락을 상기시키며 단순한 한 가지 원인으로 해석할 수 없다고 반박한다.","insights":["정렬의 극단적 해법은 통제가 아니라 이해의 공유다.","인간의 사회적 욕구는 진화가 성공적으로 심은 고차원 목표다.","AI 정렬은 뇌가 욕구를 코드에 새기는 방식과 닮아 있다.","SSI의 차별점은 확정된 답이 아니라 연구 가설의 검증이다.","회사의 변화를 사람 이탈 하나로 설명하면 맥락을 놓친다."],"keyClips":[{"clipId":"aR20FWCCjAs:c7:1-4","startSegmentIndex":1,"endSegmentIndex":4,"startTime":4205.66,"endTime":4242.041764705882,"durationSeconds":36.4,"preview":"정렬의 극단 해법","mustSee":false},{"clipId":"aR20FWCCjAs:c7:5-12","startSegmentIndex":5,"endSegmentIndex":12,"startTime":4241.82,"endTime":4296.67,"durationSeconds":54.9,"preview":"진화가 심은 욕구","mustSee":false},{"clipId":"aR20FWCCjAs:c7:13-28","startSegmentIndex":13,"endSegmentIndex":28,"startTime":4296.46,"endTime":4404.432142857143,"durationSeconds":108,"preview":"고차원 욕구의 미스터리","mustSee":false},{"clipId":"aR20FWCCjAs:c7:29-54","startSegmentIndex":29,"endSegmentIndex":54,"startTime":4404.46,"endTime":4597.947692307692,"durationSeconds":193.5,"preview":"뇌와 정렬의 비유","mustSee":false},{"clipId":"aR20FWCCjAs:c7:55-67","startSegmentIndex":55,"endSegmentIndex":67,"startTime":4597.58,"endTime":4766.560000000001,"durationSeconds":169,"preview":"SSI는 왜 연구회사인가","mustSee":true},{"clipId":"aR20FWCCjAs:c7:68-72","startSegmentIndex":68,"endSegmentIndex":72,"startTime":4769.74,"endTime":4810.830000000001,"durationSeconds":41.1,"preview":"메타 이탈의 맥락","mustSee":false}],"curatedSegments":[{"segmentIndex":51,"text":"of environments, but to figure out an approach which lets you learn from one environment and","startTime":359.17999999999995,"endTime":365.19,"durationSeconds":6,"level":"C1","overallScore":7.8,"rationale":"전이 학습의 핵심 아이디어를 제시함."},{"segmentIndex":59,"text":"Right. I think that's basically what's going on. The models are much more like the first student, but even more. Because then we say, the model should","startTime":416.7,"endTime":424.48999999999995,"durationSeconds":8,"level":"B2","overallScore":7.8,"rationale":"비유를 핵심 주장으로 연결해 통찰적임."},{"segmentIndex":40,"text":"The value function lets you short-circuit the wait until the very end.","startTime":908.86,"endTime":915.0999999999999,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"가치함수 효용을 압축한 핵심 문장이다."},{"segmentIndex":43,"text":"As soon as you conclude this, you could already get a reward signal a thousand timesteps previously, when you decided to pursue down this path.","startTime":934.22,"endTime":943.6883333333334,"durationSeconds":9,"level":"C1","overallScore":7.6,"rationale":"조기 보상 아이디어가 매우 선명하다."},{"segmentIndex":56,"text":"it's impressive that they have this much utility while still being rather simple to understand.","startTime":1026.0600000000002,"endTime":1036.3016666666667,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"단순성과 효용의 대비가 통찰적이다."},{"segmentIndex":1,"text":"The big breakthrough of pre-training is the realization that this recipe is good.","startTime":1202.14,"endTime":1208.1034615384617,"durationSeconds":6,"level":"B2","overallScore":7.2,"rationale":""},{"segmentIndex":44,"text":"obvious. That seems like a very fundamental thing. So this is the crux: generalization. There are two","startTime":1508.46,"endTime":1518.3632352941174,"durationSeconds":10,"level":"B2","overallScore":6.6,"rationale":""},{"segmentIndex":76,"text":"But if people exhibit great ability, reliability, robustness, and ability to learn in a domain that","startTime":1747.1000000000001,"endTime":1756.4675,"durationSeconds":9,"level":"C1","overallScore":7,"rationale":""},{"segmentIndex":34,"text":"something like, \"If ideas are so cheap, how come no one's having any ideas?\"","startTime":2245.1800000000003,"endTime":2250.955,"durationSeconds":6,"level":"B2","overallScore":8,"rationale":"역설적 반문이 강한 통찰과 표현성을 가짐."},{"segmentIndex":49,"text":"research where the end artifact was made safer mostly through just thinking about","startTime":2745.26,"endTime":2753.200384615385,"durationSeconds":8,"level":"C1","overallScore":7.8,"rationale":"안전 개선 방식에 대한 핵심 주장."},{"segmentIndex":55,"text":"But this is a really powerful thing and we don't even know how to conceptualize how people interact with it, what people will do with it. Having gradual access to it seems like a","startTime":2789.6600000000003,"endTime":2799.86,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"불확실한 상호작용 위험을 짚음."},{"segmentIndex":56,"text":"better way to maybe spread out the impact of it and to help people prepare for it.","startTime":2800.78,"endTime":2805.5808823529414,"durationSeconds":5,"level":"B2","overallScore":7.8,"rationale":"점진 접근의 실용적 이유가 분명함."},{"segmentIndex":18,"text":"but picking them all up at the same time, and then amalgamating their learnings, you basically have a model which functionally becomes superintelligent even without any sort","startTime":3125.42,"endTime":3135.449230769231,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"동시 학습 결합이 초지능으로 이어진단 통찰."},{"segmentIndex":37,"text":"Economic growth is one way to put it. A Dyson sphere is a lot of economic growth. But another way to put it is that you will have, in potentially a very short period of time...","startTime":3341.34,"endTime":3351.177368421053,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"경제성장을 우주 규모 비유로 확장해 강렬함."},{"segmentIndex":41,"text":"AI being deployed incrementally and in advance. One very difficult thing about AI is that we are","startTime":3382.78,"endTime":3394.799411764706,"durationSeconds":12,"level":"C1","overallScore":7.8,"rationale":"점진 배치의 중요성을 핵심 원칙으로 제시함."},{"segmentIndex":54,"text":"You’ve got to be showing the thing. You’ve got to be showing the thing.","startTime":3492.86,"endTime":3496.7814285714285,"durationSeconds":4,"level":"B2","overallScore":7.6,"rationale":"보여줘야 한다는 원칙을 강하게 압축했다."},{"segmentIndex":45,"text":"care for sentient life in a very single-minded way—we might not like the results. That's really","startTime":3923.34,"endTime":3929.5024999999996,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"위험의 본질을 직설적으로 압축함."},{"segmentIndex":55,"text":"The \"It\" we don't know how to build, and a lot hinges on understanding reliable","startTime":3989.98,"endTime":3998.964,"durationSeconds":9,"level":"C1","overallScore":7.4,"rationale":"핵심 병목을 압축적으로 제시함."},{"segmentIndex":7,"text":"still guiding our actions so strongly is an example of alignment success.","startTime":4256.0599999999995,"endTime":4263.613333333334,"durationSeconds":8,"level":"C1","overallScore":7.2,"rationale":""},{"segmentIndex":13,"text":"I think it's actually really mysterious how evolution encodes high-level desires.","startTime":4296.46,"endTime":4306.887272727273,"durationSeconds":10,"level":"C1","overallScore":6.8,"rationale":""}],"generatedAt":"2026-06-24T23:55:34.943Z","keyClipsTotalSec":2154},{"videoId":"aR20FWCCjAs","chunkIndex":8,"totalChunks":10,"title":"Ilya Sutskever – We're moving from the age of scaling to the age of research — Part 9 of 10","thumbnail":"https://i.ytimg.com/vi/aR20FWCCjAs/maxresdefault.jpg","duration":5763,"uploader":"Dwarkesh Patel","youtubeUrl":"https://www.youtube.com/watch?v=aR20FWCCjAs","keywords":["ai-safety","superintelligence","alignment","machine-learning","research","technology","future-of-ai","competition","specialization"],"normalizedKeywords":["기술 트렌드","비즈니스·전략","리더십·매니지먼트"],"targetAudience":[{"who":"AI 연구자","why":"초지능의 정렬과 기술 경로 수렴에 대한 관점을 들을 수 있음"},{"who":"창업자","why":"AI 경쟁 구도와 제품 전략이 어떻게 재편될지 이해하는 데 도움됨"},{"who":"투자자","why":"AI 시장이 수렴·분화하며 어떤 가치가 분산될지 판단하는 데 유용함"}],"normalizedAudience":["리서처·학자","창업자·스타트업","투자자·VC"],"summary":"이 구간은 초지능 시대에 AI 회사들이 어떤 기술·정렬 전략을 택하게 될지에 대한 이لی아 수츠케버의 전망을 중심으로 전개된다. 그는 SSI가 추구하는 접근이 다르다고 보지만, 결국 경쟁자들도 비슷한 방향으로 수렴할 것이라고 말한다. 다만 그 수렴은 단번에 명확해지는 것이 아니라, 더 강한 AI가 등장할 때마다 무엇이 가능한지에 대한 정보가 늘어나며 점진적으로 일어날 것이라고 본다.\n\n또한 그는 미래의 AI 경제가 단일 승자 독식이 아니라 전문화와 니치 경쟁으로 흘러갈 가능성을 강조한다. 한 회사가 모든 일을 독점적으로 처리하기보다, 각 회사가 서로 다른 영역에서 강점을 갖는 구조가 될 수 있다는 것이다. 마지막으로 그는 인간처럼 계속 학습하는 AI를 여러 개 복제해 병렬화하는 발상이 직관적으로는 매력적이지만, 실제로는 사고방식의 다양성이 더 중요하다고 본다.","insights":["초지능 경쟁의 핵심은 모델 성능보다 정렬 전략이다.","AI가 강해질수록 업계는 결국 비슷한 해법으로 수렴한다.","승자는 하나가 아니라, 전문화된 여러 니치가 될 수 있다.","복제의 효율보다 사고 다양성이 더 큰 가치를 만든다.","대규모 AI 진전은 시장 가격을 낮추며 빠르게 확산된다."],"keyClips":[{"clipId":"aR20FWCCjAs:c8:3-14","startSegmentIndex":3,"endSegmentIndex":14,"startTime":4825.339999999999,"endTime":4920.543846153846,"durationSeconds":95.2,"preview":"정렬전략의 수렴","mustSee":true},{"clipId":"aR20FWCCjAs:c8:16-18","startSegmentIndex":16,"endSegmentIndex":18,"startTime":4927.82,"endTime":4949.71,"durationSeconds":21.9,"preview":"초지능 도달시점","mustSee":false},{"clipId":"aR20FWCCjAs:c8:20-31","startSegmentIndex":20,"endSegmentIndex":31,"startTime":4955.58,"endTime":5057.55,"durationSeconds":102,"preview":"수렴은 제품 공개로","mustSee":false},{"clipId":"aR20FWCCjAs:c8:39-53","startSegmentIndex":39,"endSegmentIndex":53,"startTime":5114.78,"endTime":5235.572777777777,"durationSeconds":120.8,"preview":"전문화된 AI시장","mustSee":true},{"clipId":"aR20FWCCjAs:c8:54-75","startSegmentIndex":54,"endSegmentIndex":75,"startTime":5235.099999999999,"endTime":5401.903076923078,"durationSeconds":166.8,"preview":"다양성이 더 강하다","mustSee":true}],"curatedSegments":[{"segmentIndex":51,"text":"of environments, but to figure out an approach which lets you learn from one environment and","startTime":359.17999999999995,"endTime":365.19,"durationSeconds":6,"level":"C1","overallScore":7.8,"rationale":"전이 학습의 핵심 아이디어를 제시함."},{"segmentIndex":59,"text":"Right. I think that's basically what's going on. The models are much more like the first student, but even more. Because then we say, the model should","startTime":416.7,"endTime":424.48999999999995,"durationSeconds":8,"level":"B2","overallScore":7.8,"rationale":"비유를 핵심 주장으로 연결해 통찰적임."},{"segmentIndex":40,"text":"The value function lets you short-circuit the wait until the very end.","startTime":908.86,"endTime":915.0999999999999,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"가치함수 효용을 압축한 핵심 문장이다."},{"segmentIndex":43,"text":"As soon as you conclude this, you could already get a reward signal a thousand timesteps previously, when you decided to pursue down this path.","startTime":934.22,"endTime":943.6883333333334,"durationSeconds":9,"level":"C1","overallScore":7.6,"rationale":"조기 보상 아이디어가 매우 선명하다."},{"segmentIndex":56,"text":"it's impressive that they have this much utility while still being rather simple to understand.","startTime":1026.0600000000002,"endTime":1036.3016666666667,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"단순성과 효용의 대비가 통찰적이다."},{"segmentIndex":1,"text":"The big breakthrough of pre-training is the realization that this recipe is good.","startTime":1202.14,"endTime":1208.1034615384617,"durationSeconds":6,"level":"B2","overallScore":7.2,"rationale":""},{"segmentIndex":44,"text":"obvious. That seems like a very fundamental thing. So this is the crux: generalization. There are two","startTime":1508.46,"endTime":1518.3632352941174,"durationSeconds":10,"level":"B2","overallScore":6.6,"rationale":""},{"segmentIndex":76,"text":"But if people exhibit great ability, reliability, robustness, and ability to learn in a domain that","startTime":1747.1000000000001,"endTime":1756.4675,"durationSeconds":9,"level":"C1","overallScore":7,"rationale":""},{"segmentIndex":34,"text":"something like, \"If ideas are so cheap, how come no one's having any ideas?\"","startTime":2245.1800000000003,"endTime":2250.955,"durationSeconds":6,"level":"B2","overallScore":8,"rationale":"역설적 반문이 강한 통찰과 표현성을 가짐."},{"segmentIndex":49,"text":"research where the end artifact was made safer mostly through just thinking about","startTime":2745.26,"endTime":2753.200384615385,"durationSeconds":8,"level":"C1","overallScore":7.8,"rationale":"안전 개선 방식에 대한 핵심 주장."},{"segmentIndex":55,"text":"But this is a really powerful thing and we don't even know how to conceptualize how people interact with it, what people will do with it. Having gradual access to it seems like a","startTime":2789.6600000000003,"endTime":2799.86,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"불확실한 상호작용 위험을 짚음."},{"segmentIndex":56,"text":"better way to maybe spread out the impact of it and to help people prepare for it.","startTime":2800.78,"endTime":2805.5808823529414,"durationSeconds":5,"level":"B2","overallScore":7.8,"rationale":"점진 접근의 실용적 이유가 분명함."},{"segmentIndex":18,"text":"but picking them all up at the same time, and then amalgamating their learnings, you basically have a model which functionally becomes superintelligent even without any sort","startTime":3125.42,"endTime":3135.449230769231,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"동시 학습 결합이 초지능으로 이어진단 통찰."},{"segmentIndex":37,"text":"Economic growth is one way to put it. A Dyson sphere is a lot of economic growth. But another way to put it is that you will have, in potentially a very short period of time...","startTime":3341.34,"endTime":3351.177368421053,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"경제성장을 우주 규모 비유로 확장해 강렬함."},{"segmentIndex":41,"text":"AI being deployed incrementally and in advance. One very difficult thing about AI is that we are","startTime":3382.78,"endTime":3394.799411764706,"durationSeconds":12,"level":"C1","overallScore":7.8,"rationale":"점진 배치의 중요성을 핵심 원칙으로 제시함."},{"segmentIndex":54,"text":"You’ve got to be showing the thing. You’ve got to be showing the thing.","startTime":3492.86,"endTime":3496.7814285714285,"durationSeconds":4,"level":"B2","overallScore":7.6,"rationale":"보여줘야 한다는 원칙을 강하게 압축했다."},{"segmentIndex":45,"text":"care for sentient life in a very single-minded way—we might not like the results. That's really","startTime":3923.34,"endTime":3929.5024999999996,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"위험의 본질을 직설적으로 압축함."},{"segmentIndex":55,"text":"The \"It\" we don't know how to build, and a lot hinges on understanding reliable","startTime":3989.98,"endTime":3998.964,"durationSeconds":9,"level":"C1","overallScore":7.4,"rationale":"핵심 병목을 압축적으로 제시함."},{"segmentIndex":7,"text":"still guiding our actions so strongly is an example of alignment success.","startTime":4256.0599999999995,"endTime":4263.613333333334,"durationSeconds":8,"level":"C1","overallScore":7.2,"rationale":""},{"segmentIndex":13,"text":"I think it's actually really mysterious how evolution encodes high-level desires.","startTime":4296.46,"endTime":4306.887272727273,"durationSeconds":10,"level":"C1","overallScore":6.8,"rationale":""}],"generatedAt":"2026-06-24T23:56:09.266Z","keyClipsTotalSec":2154},{"videoId":"aR20FWCCjAs","chunkIndex":9,"totalChunks":10,"title":"Ilya Sutskever – We're moving from the age of scaling to the age of research — Part 10 of 10","thumbnail":"https://i.ytimg.com/vi/aR20FWCCjAs/maxresdefault.jpg","duration":5763,"uploader":"Dwarkesh Patel","youtubeUrl":"https://www.youtube.com/watch?v=aR20FWCCjAs","keywords":["artificial-intelligence","machine-learning","reinforcement-learning","self-play","llm","research","deep-learning","adversarial-training","ai-agents"],"normalizedKeywords":["기술 트렌드","엔지니어링","교육"],"targetAudience":[{"who":"AI 연구자","why":"자기대결, RL, 에이전트 다양성 같은 연구 방향을 정리해줌"},{"who":"엔지니어","why":"LLM post-training과 경쟁적 학습 구조를 이해하는 데 도움됨"},{"who":"대학원생","why":"연구 아이디어를 고르는 기준과 사고방식을 배울 수 있음"}],"normalizedAudience":["리서처·학자","엔지니어·개발자","학생·주니어"],"summary":"이 영상의 마지막 구간은 AI 에이전트의 다양성이 왜 잘 생기지 않는지, 그리고 self-play가 LLM 시대에는 어떤 형태로 살아남는지를 다룬다. Ilya Sutskever는 사전학습(pre-training)은 같은 데이터 위에서 이뤄지기 때문에 모델들이 비슷해지지만, RL과 post-training 단계에서는 서로 다른 훈련 방식 때문에 차별화가 생긴다고 말한다.\n\n이어 self-play의 가치는 '데이터 없이 compute만으로 모델을 만든다'는 점에 있었지만, 과거 방식의 self-play는 협상·갈등·전략처럼 좁은 기술만 키우는 한계가 있다고 설명한다. 대신 debate, prover-verifier, LLM-as-a-Judge 같은 적대적 구조가 더 넓은 의미의 self-play로 자리 잡고 있으며, 여러 에이전트를 함께 두고 서로의 접근을 관찰하게 하면 오히려 차별화된 아이디어를 유도할 수 있다고 본다. 마지막으로 그는 자신의 research taste를 '브레인에서 영감을 받되, 아름다움·단순성·정합성을 동시에 만족하는 방향을 믿는 것'으로 정의한다.","insights":["사전학습은 모델을 평준화하고, post-training이 차이를 만든다.","self-play의 핵심 가치는 데이터 대신 compute로 학습한다는 점이다.","과거 self-play는 좁은 기술만 키워 범용성에 한계가 있다.","적대적 구조는 단순 경쟁이 아니라 검증과 교정의 장치다.","좋은 연구 감각은 아름다움·단순성·브레인 직관의 결합이다."],"keyClips":[{"clipId":"aR20FWCCjAs:c9:1-3","startSegmentIndex":1,"endSegmentIndex":3,"startTime":5401.259999999999,"endTime":5425.565555555556,"durationSeconds":24.3,"preview":"에이전트 다양성의 원인","mustSee":false},{"clipId":"aR20FWCCjAs:c9:7-13","startSegmentIndex":7,"endSegmentIndex":13,"startTime":5446.7,"endTime":5499.284090909091,"durationSeconds":52.6,"preview":"자기대결의 한계와 가치","mustSee":false},{"clipId":"aR20FWCCjAs:c9:14-17","startSegmentIndex":14,"endSegmentIndex":17,"startTime":5499.259999999999,"endTime":5533.1,"durationSeconds":33.8,"preview":"적대적 학습의 진화","mustSee":true},{"clipId":"aR20FWCCjAs:c9:18-20","startSegmentIndex":18,"endSegmentIndex":20,"startTime":5533.5,"endTime":5556.000666666667,"durationSeconds":22.5,"preview":"경쟁이 낳는 차별화","mustSee":false},{"clipId":"aR20FWCCjAs:c9:25-41","startSegmentIndex":25,"endSegmentIndex":41,"startTime":5591.0199999999995,"endTime":5725.182857142858,"durationSeconds":134.2,"preview":"연구 감각의 본질","mustSee":true}],"curatedSegments":[{"segmentIndex":51,"text":"of environments, but to figure out an approach which lets you learn from one environment and","startTime":359.17999999999995,"endTime":365.19,"durationSeconds":6,"level":"C1","overallScore":7.8,"rationale":"전이 학습의 핵심 아이디어를 제시함."},{"segmentIndex":59,"text":"Right. I think that's basically what's going on. The models are much more like the first student, but even more. Because then we say, the model should","startTime":416.7,"endTime":424.48999999999995,"durationSeconds":8,"level":"B2","overallScore":7.8,"rationale":"비유를 핵심 주장으로 연결해 통찰적임."},{"segmentIndex":40,"text":"The value function lets you short-circuit the wait until the very end.","startTime":908.86,"endTime":915.0999999999999,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"가치함수 효용을 압축한 핵심 문장이다."},{"segmentIndex":43,"text":"As soon as you conclude this, you could already get a reward signal a thousand timesteps previously, when you decided to pursue down this path.","startTime":934.22,"endTime":943.6883333333334,"durationSeconds":9,"level":"C1","overallScore":7.6,"rationale":"조기 보상 아이디어가 매우 선명하다."},{"segmentIndex":56,"text":"it's impressive that they have this much utility while still being rather simple to understand.","startTime":1026.0600000000002,"endTime":1036.3016666666667,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"단순성과 효용의 대비가 통찰적이다."},{"segmentIndex":1,"text":"The big breakthrough of pre-training is the realization that this recipe is good.","startTime":1202.14,"endTime":1208.1034615384617,"durationSeconds":6,"level":"B2","overallScore":7.2,"rationale":""},{"segmentIndex":44,"text":"obvious. That seems like a very fundamental thing. So this is the crux: generalization. There are two","startTime":1508.46,"endTime":1518.3632352941174,"durationSeconds":10,"level":"B2","overallScore":6.6,"rationale":""},{"segmentIndex":76,"text":"But if people exhibit great ability, reliability, robustness, and ability to learn in a domain that","startTime":1747.1000000000001,"endTime":1756.4675,"durationSeconds":9,"level":"C1","overallScore":7,"rationale":""},{"segmentIndex":34,"text":"something like, \"If ideas are so cheap, how come no one's having any ideas?\"","startTime":2245.1800000000003,"endTime":2250.955,"durationSeconds":6,"level":"B2","overallScore":8,"rationale":"역설적 반문이 강한 통찰과 표현성을 가짐."},{"segmentIndex":49,"text":"research where the end artifact was made safer mostly through just thinking about","startTime":2745.26,"endTime":2753.200384615385,"durationSeconds":8,"level":"C1","overallScore":7.8,"rationale":"안전 개선 방식에 대한 핵심 주장."},{"segmentIndex":55,"text":"But this is a really powerful thing and we don't even know how to conceptualize how people interact with it, what people will do with it. Having gradual access to it seems like a","startTime":2789.6600000000003,"endTime":2799.86,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"불확실한 상호작용 위험을 짚음."},{"segmentIndex":56,"text":"better way to maybe spread out the impact of it and to help people prepare for it.","startTime":2800.78,"endTime":2805.5808823529414,"durationSeconds":5,"level":"B2","overallScore":7.8,"rationale":"점진 접근의 실용적 이유가 분명함."},{"segmentIndex":18,"text":"but picking them all up at the same time, and then amalgamating their learnings, you basically have a model which functionally becomes superintelligent even without any sort","startTime":3125.42,"endTime":3135.449230769231,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"동시 학습 결합이 초지능으로 이어진단 통찰."},{"segmentIndex":37,"text":"Economic growth is one way to put it. A Dyson sphere is a lot of economic growth. But another way to put it is that you will have, in potentially a very short period of time...","startTime":3341.34,"endTime":3351.177368421053,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"경제성장을 우주 규모 비유로 확장해 강렬함."},{"segmentIndex":41,"text":"AI being deployed incrementally and in advance. One very difficult thing about AI is that we are","startTime":3382.78,"endTime":3394.799411764706,"durationSeconds":12,"level":"C1","overallScore":7.8,"rationale":"점진 배치의 중요성을 핵심 원칙으로 제시함."},{"segmentIndex":54,"text":"You’ve got to be showing the thing. You’ve got to be showing the thing.","startTime":3492.86,"endTime":3496.7814285714285,"durationSeconds":4,"level":"B2","overallScore":7.6,"rationale":"보여줘야 한다는 원칙을 강하게 압축했다."},{"segmentIndex":45,"text":"care for sentient life in a very single-minded way—we might not like the results. That's really","startTime":3923.34,"endTime":3929.5024999999996,"durationSeconds":6,"level":"B2","overallScore":7.6,"rationale":"위험의 본질을 직설적으로 압축함."},{"segmentIndex":55,"text":"The \"It\" we don't know how to build, and a lot hinges on understanding reliable","startTime":3989.98,"endTime":3998.964,"durationSeconds":9,"level":"C1","overallScore":7.4,"rationale":"핵심 병목을 압축적으로 제시함."},{"segmentIndex":7,"text":"still guiding our actions so strongly is an example of alignment success.","startTime":4256.0599999999995,"endTime":4263.613333333334,"durationSeconds":8,"level":"C1","overallScore":7.2,"rationale":""},{"segmentIndex":13,"text":"I think it's actually really mysterious how evolution encodes high-level desires.","startTime":4296.46,"endTime":4306.887272727273,"durationSeconds":10,"level":"C1","overallScore":6.8,"rationale":""}],"generatedAt":"2026-06-24T23:56:27.038Z","keyClipsTotalSec":2154}]}