{"success":true,"count":3,"items":[{"videoId":"96jN2OCOfLs","chunkIndex":0,"totalChunks":3,"title":"Andrej Karpathy: From Vibe Coding to Agentic Engineering — Part 1 of 3","thumbnail":"https://i.ytimg.com/vi/96jN2OCOfLs/maxresdefault.jpg","duration":1789,"uploader":"Sequoia Capital","youtubeUrl":"https://www.youtube.com/watch?v=96jN2OCOfLs","keywords":["ai","llm","agentic-ai","programming","software-paradigm","automation","neural-networks","productivity","technical-trends"],"normalizedKeywords":["기술 트렌드","엔지니어링","프로덕트"],"targetAudience":[{"who":"엔지니어","why":"에이전트와 LLM이 개발 방식을 어떻게 바꾸는지 직접적으로 다룬다."},{"who":"초기 창업자","why":"새로운 AI 패러다임에서 어떤 제품을 만들 수 있는지 힌트를 준다."},{"who":"기획자","why":"코딩보다 더 넓은 정보처리 자동화 관점을 이해하는 데 유용하다."}],"normalizedAudience":["엔지니어·개발자","창업자·스타트업","프로덕트 매니저·기획자"],"summary":"이 영상은 Andrej Karpathy가 말하는 AI 시대의 새로운 프로그래밍 패러다임을 중심으로, LLM과 에이전트가 단순히 '더 좋은 도구'가 아니라 컴퓨터 자체를 다시 정의하고 있다고 주장한다. 그는 2024년 말부터 최신 모델들이 코드 조각을 거의 틀리지 않게 생성하면서 자신이 '프로그램을 짜는 사람'이 아니라 '에이전트에게 텍스트를 건네는 사람'이 되었다는 경험을 공유한다.\n\n이어 software 1.0, 2.0, 3.0의 차이를 통해, 이제는 코드 작성보다 프롬프트·컨텍스트 설계가 더 중요한 프로그래밍이 되었고, 설치 스크립트나 메뉴 이미지 생성 같은 사례도 기존 앱 레이어가 사라지는 방향으로 진화한다고 설명한다. 또한 AI가 구조화된 데이터뿐 아니라 문서·이미지·영상 같은 비정형 정보까지 재구성해 새로운 산출물을 만들 수 있으며, 앞으로는 '무엇이 더 빨라지나'보다 '이전엔 불가능했던 무엇이 가능해지나'를 보는 것이 더 중요하다고 강조한다.","insights":["AI는 코딩 속도 향상보다 프로그래밍 정의 자체를 바꾼다.","이제 중요한 레버는 코드가 아니라 프롬프트와 컨텍스트다.","에이전트가 정답을 찾아가면 인간은 세부 구현보다 목표를 준다.","비정형 데이터를 재조합하는 일은 새 소프트웨어가 된다.","미래의 핵심 질문은 '더 빠른가'가 아니라 '새로 가능한가'다."],"keyClips":[{"clipId":"96jN2OCOfLs:c0:9-18","startSegmentIndex":9,"endSegmentIndex":18,"startTime":60.2,"endTime":145.4,"durationSeconds":85.2,"preview":"에이전트가 바꾼 감각","mustSee":false},{"clipId":"96jN2OCOfLs:c0:23-38","startSegmentIndex":23,"endSegmentIndex":38,"startTime":170.64,"endTime":289.96,"durationSeconds":119.3,"preview":"Software 3.0 정의","mustSee":true},{"clipId":"96jN2OCOfLs:c0:39-49","startSegmentIndex":39,"endSegmentIndex":49,"startTime":289.96,"endTime":382.56,"durationSeconds":92.6,"preview":"앱이 사라지는 순간","mustSee":false},{"clipId":"96jN2OCOfLs:c0:53-58","startSegmentIndex":53,"endSegmentIndex":58,"startTime":406.56,"endTime":455,"durationSeconds":48.4,"preview":"비정형 정보의 재컴파일","mustSee":false},{"clipId":"96jN2OCOfLs:c0:63-74","startSegmentIndex":63,"endSegmentIndex":74,"startTime":496,"endTime":581.2,"durationSeconds":85.2,"preview":"신경망이 호스트가 되면","mustSee":false}],"curatedSegments":[{"segmentIndex":25,"text":"And so that's why I think I'm stressing this dimension of it as we are slightly at the mercy of whatever the labs are doing, whatever they happen to put into the mix and you have to actually explore this thing that they give you that has no manual and it works in certain settings but maybe not in some settings and you have to kind of explore it a little bit and if you're in the circuits that were part of the RL, you fly and if you're in the circuits that are out of the data distribution, you're going to struggle and you have to kind of figure out which circuits you're in your application.","startTime":777.12,"endTime":807.12,"durationSeconds":30,"level":"C2","overallScore":8.8,"rationale":"실무 적용 한계와 탐색 원칙을 압축함."},{"segmentIndex":34,"text":"Um and these are jagged forms of intelligence that are shaped by data and reward functions, but not by intrinsic motivation or fun or curiosity or empowerment, uh things that kind of came about via evolution.","startTime":1425.76,"endTime":1439.6,"durationSeconds":14,"level":"C1","overallScore":8.6,"rationale":"AI 지능의 성격을 깊게 규정함."},{"segmentIndex":67,"text":"It was something along the lines of um you can outsource your thinking, but you can't outsource your understanding.","startTime":1690.24,"endTime":1696.48,"durationSeconds":6,"level":"B2","overallScore":8.6,"rationale":"대조적 통찰과 기억할 표현이 강함."},{"segmentIndex":15,"text":"I think that a lot of people actually I tried to stress this on uh Twitter and or X because I think a lot of people experienced AI uh last year as ChatGPT adjacent thing, uh but you really had to look again, and you had to look as of December uh because things have changed fundamentally and uh especially on this like agentic coherent workflow that really started to actually work.","startTime":103.36,"endTime":125.04,"durationSeconds":22,"level":"C1","overallScore":8,"rationale":"변화 시점과 이유를 통찰적으로 설명."},{"segmentIndex":27,"text":"So software 3.0 is kind of about uh you know, your programming now turns to prompting and what's in the context window is your lever over the interpreter that is the LLM that is kind of like interpreting your context and uh performing computation in the digital information space.","startTime":200.36,"endTime":215.56,"durationSeconds":15,"level":"C2","overallScore":8,"rationale":"새 패러다임을 밀도 높게 설명함."},{"segmentIndex":35,"text":"And the reason this is a lot more powerful is you're working now in the software 3.0 paradigm where you don't have to precisely uh spell out, you know, all the individual details of that setup.","startTime":260.92,"endTime":269.36,"durationSeconds":8,"level":"C1","overallScore":8,"rationale":"왜 강력한지 원리를 분명히 설명함."},{"segmentIndex":36,"text":"The agent has its own intelligence that it packages up and then it kind of like follows the instructions and it looks at your environment, your computer, and it kind of like performs intelligent actions to make things work and debugs things in the loop.","startTime":269.36,"endTime":279.84000000000003,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"에이전트 작동 방식을 구체화함."},{"segmentIndex":44,"text":"And then I saw the software 3.0 version of this, which is which blew my mind, which is literally just take your photo, give it to Gemini, and say use Nano Banana to overlay the things onto the menu.\"","startTime":337.56,"endTime":350,"durationSeconds":12,"level":"C1","overallScore":7.8,"rationale":"기존 앱을 대체하는 전환이 강렬함."},{"segmentIndex":46,"text":"And this blew my mind because actually all of my menu gen is spurious.","startTime":359.52,"endTime":364.24,"durationSeconds":5,"level":"B2","overallScore":7.4,"rationale":"기존 작업의 무의미함을 통찰적으로 밝힘."},{"segmentIndex":47,"text":"It's working in the old paradigm that app shouldn't exist.","startTime":364.24,"endTime":367.44,"durationSeconds":3,"level":"B2","overallScore":7.6,"rationale":"패러다임 전환의 함의를 날카롭게 말함."},{"segmentIndex":49,"text":"It just um your neural network is doing more and more of the work, and your prompt or context is just the image, and the output is an image, and there's no need to have any of the app in between.","startTime":371.72,"endTime":382.56,"durationSeconds":11,"level":"C1","overallScore":7.8,"rationale":"중간 앱이 불필요한 이유를 설명함."},{"segmentIndex":50,"text":"Um so, I think that people have to kind of like reframe, you know, not to work in the existing paradigm of what things existed and just think about it as a speed up of what exists.","startTime":382.56,"endTime":393.2,"durationSeconds":11,"level":"C1","overallScore":8,"rationale":"사고 전환을 직접 요구하는 문장."},{"segmentIndex":51,"text":"It's actually like new things are available now. And going back to your programming question, it's not even I think that's also an example of working in the old mindset because it's not just about programming and programming becoming faster.","startTime":393.2,"endTime":403.44,"durationSeconds":10,"level":"C1","overallScore":8,"rationale":"질문 틀 자체를 재정의하는 통찰."},{"segmentIndex":56,"text":"But now you can just take these documents and uh basically uh recompile them in a different way, and uh reorder them, and create something that is uh new and interesting uh as a reframing of the data.","startTime":429.76,"endTime":440.28,"durationSeconds":11,"level":"C1","overallScore":7.8,"rationale":"새 처리 가능성을 구체적으로 설명함."},{"segmentIndex":58,"text":"Uh and so, I think this is uh something that I keep trying to get back to as to not only what can we do that existed that is faster now, but I think there's new opportunities of just things that couldn't be possible before.","startTime":443.48,"endTime":455,"durationSeconds":12,"level":"C1","overallScore":8,"rationale":"논지의 핵심을 잘 압축한 문장."},{"segmentIndex":60,"text":"If you extrapolate that further, what is the 2026 equivalent for building websites in the '90s, building mobile apps in the 2010s, building SaaS in the last cloud era?","startTime":469.88,"endTime":482.48,"durationSeconds":13,"level":"C1","overallScore":8,"rationale":"미래 비유를 묻는 통찰적 질문."},{"segmentIndex":71,"text":"So, you could imagine something really weird and foreign when where neural nets are doing most of the heavy lifting, they're using tool use as just like, you know, historical appendage for some kinds of like deterministic tasks.","startTime":553.96,"endTime":565.24,"durationSeconds":11,"level":"C1","overallScore":8,"rationale":"핵심 관점이 선명하고 표현도 풍부함."},{"segmentIndex":76,"text":"The fact that AI will automate faster and more easily domains where the output can be verified.","startTime":585.12,"endTime":591,"durationSeconds":6,"level":"B2","overallScore":7.8,"rationale":"검증가능성 원리를 명확히 제시함."},{"segmentIndex":4,"text":"So, they are given a verification rewards. And then because of the way that these models are trained, they end up basically uh, progressing and creating these like jagged entities that really peak in capability in kind of like verifiable domains like math and code and adjacent.","startTime":624.52,"endTime":638.76,"durationSeconds":14,"level":"C2","overallScore":7.8,"rationale":"모델 능력의 들쭉날쭉함 원인 설명."},{"segmentIndex":14,"text":"How is it possible that state-of-the-art Opus 4.7 will simultaneously refactor a 100,000 like >> [laughter] >> code base a line code base or find zero-day vulnerabilities and yet tells me to walk to this car wash?","startTime":700.64,"endTime":714.16,"durationSeconds":14,"level":"C2","overallScore":7.8,"rationale":"능력 대비 오류의 역설을 강하게 제시."}],"generatedAt":"2026-06-25T00:31:50.845Z","keyClipsTotalSec":778},{"videoId":"96jN2OCOfLs","chunkIndex":1,"totalChunks":3,"title":"Andrej Karpathy: From Vibe Coding to Agentic Engineering — Part 2 of 3","thumbnail":"https://i.ytimg.com/vi/96jN2OCOfLs/maxresdefault.jpg","duration":1789,"uploader":"Sequoia Capital","youtubeUrl":"https://www.youtube.com/watch?v=96jN2OCOfLs","keywords":["ai","llm","verifiability","reinforcement-learning","agentic-engineering","vibe-coding","software-engineering","founder-advice","tooling"],"normalizedKeywords":["엔지니어링","비즈니스·전략","기술 트렌드"],"targetAudience":[{"who":"엔지니어","why":"LLM을 실제 개발 도구로 쓰는 방식과 한계를 이해하는 데 유용함"},{"who":"초기 단계 창업자","why":"AI 시대에 어떤 문제를 골라야 하는지 제품/기술 전략 힌트를 줌"},{"who":"채용 담당자","why":"에이전틱 엔지니어를 어떻게 평가하고 뽑을지 관점을 제공함"}],"normalizedAudience":["엔지니어·개발자","창업자·스타트업"],"summary":"이 대화는 LLM과 에이전트가 왜 어떤 영역에서는 놀랍도록 잘하고, 어떤 영역에서는 이상하게 약한지 설명하는 데 집중한다. 핵심 키워드는 'verifiability(검증 가능성)'로, 코드·수학처럼 결과를 쉽게 평가할 수 있는 영역에서는 RL과 데이터 투입을 통해 능력이 빠르게 뾰족해지지만, 그렇지 않은 영역에서는 여전히 jagged intelligence가 남는다는 주장이다. 카르파티는 이 현상이 단순한 모델 발전의 선형 증가가 아니라, 학습 방식과 labs가 무엇을 데이터에 넣는지에 크게 좌우된다고 본다.\n\n후반부에서는 vibe coding과 agentic engineering의 차이를 정의한다. 전자는 누구나 소프트웨어를 만들 수 있게 바닥을 올리는 변화라면, 후자는 품질 기준을 유지한 채 에이전트를 조율해 더 빠르게 일하는 기술이다. 따라서 실무자는 LLM을 '자동화 기계'로 보기보다 도구로 다루며, 큰 프로젝트를 안전하게 완성할 수 있는 능력, 시스템 세팅, 판단과 맛(taste)을 더 중요하게 키워야 한다고 말한다.","insights":["검증 가능한 일은 RL로 빠르게 뾰족해진다.","모델의 기괴한 편차는 학습법과 데이터 선택의 산물이다.","LLM은 만능 자동화기가 아니라, 써야 하는 도구다.","vibe coding은 진입장벽을 낮추고, agentic engineering은 품질을 지킨다.","에이전트 시대의 경쟁력은 큰 프로젝트를 끝내는 능력이다."],"keyClips":[{"clipId":"96jN2OCOfLs:c1:1-18","startSegmentIndex":1,"endSegmentIndex":18,"startTime":600.99,"endTime":744.68,"durationSeconds":143.7,"preview":"검증 가능성의 힘","mustSee":true},{"clipId":"96jN2OCOfLs:c1:19-26","startSegmentIndex":19,"endSegmentIndex":26,"startTime":744.68,"endTime":816.64,"durationSeconds":72,"preview":"데이터가 능력을 만든다","mustSee":false},{"clipId":"96jN2OCOfLs:c1:28-37","startSegmentIndex":28,"endSegmentIndex":37,"startTime":819.84,"endTime":909.16,"durationSeconds":89.3,"preview":"창업자의 유효지대","mustSee":false},{"clipId":"96jN2OCOfLs:c1:38-43","startSegmentIndex":38,"endSegmentIndex":43,"startTime":909.16,"endTime":946.6,"durationSeconds":37.4,"preview":"자동화의 경계는 좁다","mustSee":false},{"clipId":"96jN2OCOfLs:c1:44-60","startSegmentIndex":44,"endSegmentIndex":60,"startTime":946.6,"endTime":1038.48,"durationSeconds":91.9,"preview":"바이브와 에이전틱","mustSee":true},{"clipId":"96jN2OCOfLs:c1:61-73","startSegmentIndex":61,"endSegmentIndex":73,"startTime":1038.48,"endTime":1169.12,"durationSeconds":130.6,"preview":"AI 네이티브 개발법","mustSee":false},{"clipId":"96jN2OCOfLs:c1:74-77","startSegmentIndex":74,"endSegmentIndex":77,"startTime":1169.12,"endTime":1201.12,"durationSeconds":32,"preview":"인간의 남는 역할","mustSee":true}],"curatedSegments":[{"segmentIndex":25,"text":"And so that's why I think I'm stressing this dimension of it as we are slightly at the mercy of whatever the labs are doing, whatever they happen to put into the mix and you have to actually explore this thing that they give you that has no manual and it works in certain settings but maybe not in some settings and you have to kind of explore it a little bit and if you're in the circuits that were part of the RL, you fly and if you're in the circuits that are out of the data distribution, you're going to struggle and you have to kind of figure out which circuits you're in your application.","startTime":777.12,"endTime":807.12,"durationSeconds":30,"level":"C2","overallScore":8.8,"rationale":"실무 적용 한계와 탐색 원칙을 압축함."},{"segmentIndex":34,"text":"Um and these are jagged forms of intelligence that are shaped by data and reward functions, but not by intrinsic motivation or fun or curiosity or empowerment, uh things that kind of came about via evolution.","startTime":1425.76,"endTime":1439.6,"durationSeconds":14,"level":"C1","overallScore":8.6,"rationale":"AI 지능의 성격을 깊게 규정함."},{"segmentIndex":67,"text":"It was something along the lines of um you can outsource your thinking, but you can't outsource your understanding.","startTime":1690.24,"endTime":1696.48,"durationSeconds":6,"level":"B2","overallScore":8.6,"rationale":"대조적 통찰과 기억할 표현이 강함."},{"segmentIndex":15,"text":"I think that a lot of people actually I tried to stress this on uh Twitter and or X because I think a lot of people experienced AI uh last year as ChatGPT adjacent thing, uh but you really had to look again, and you had to look as of December uh because things have changed fundamentally and uh especially on this like agentic coherent workflow that really started to actually work.","startTime":103.36,"endTime":125.04,"durationSeconds":22,"level":"C1","overallScore":8,"rationale":"변화 시점과 이유를 통찰적으로 설명."},{"segmentIndex":27,"text":"So software 3.0 is kind of about uh you know, your programming now turns to prompting and what's in the context window is your lever over the interpreter that is the LLM that is kind of like interpreting your context and uh performing computation in the digital information space.","startTime":200.36,"endTime":215.56,"durationSeconds":15,"level":"C2","overallScore":8,"rationale":"새 패러다임을 밀도 높게 설명함."},{"segmentIndex":35,"text":"And the reason this is a lot more powerful is you're working now in the software 3.0 paradigm where you don't have to precisely uh spell out, you know, all the individual details of that setup.","startTime":260.92,"endTime":269.36,"durationSeconds":8,"level":"C1","overallScore":8,"rationale":"왜 강력한지 원리를 분명히 설명함."},{"segmentIndex":36,"text":"The agent has its own intelligence that it packages up and then it kind of like follows the instructions and it looks at your environment, your computer, and it kind of like performs intelligent actions to make things work and debugs things in the loop.","startTime":269.36,"endTime":279.84000000000003,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"에이전트 작동 방식을 구체화함."},{"segmentIndex":44,"text":"And then I saw the software 3.0 version of this, which is which blew my mind, which is literally just take your photo, give it to Gemini, and say use Nano Banana to overlay the things onto the menu.\"","startTime":337.56,"endTime":350,"durationSeconds":12,"level":"C1","overallScore":7.8,"rationale":"기존 앱을 대체하는 전환이 강렬함."},{"segmentIndex":46,"text":"And this blew my mind because actually all of my menu gen is spurious.","startTime":359.52,"endTime":364.24,"durationSeconds":5,"level":"B2","overallScore":7.4,"rationale":"기존 작업의 무의미함을 통찰적으로 밝힘."},{"segmentIndex":47,"text":"It's working in the old paradigm that app shouldn't exist.","startTime":364.24,"endTime":367.44,"durationSeconds":3,"level":"B2","overallScore":7.6,"rationale":"패러다임 전환의 함의를 날카롭게 말함."},{"segmentIndex":49,"text":"It just um your neural network is doing more and more of the work, and your prompt or context is just the image, and the output is an image, and there's no need to have any of the app in between.","startTime":371.72,"endTime":382.56,"durationSeconds":11,"level":"C1","overallScore":7.8,"rationale":"중간 앱이 불필요한 이유를 설명함."},{"segmentIndex":50,"text":"Um so, I think that people have to kind of like reframe, you know, not to work in the existing paradigm of what things existed and just think about it as a speed up of what exists.","startTime":382.56,"endTime":393.2,"durationSeconds":11,"level":"C1","overallScore":8,"rationale":"사고 전환을 직접 요구하는 문장."},{"segmentIndex":51,"text":"It's actually like new things are available now. And going back to your programming question, it's not even I think that's also an example of working in the old mindset because it's not just about programming and programming becoming faster.","startTime":393.2,"endTime":403.44,"durationSeconds":10,"level":"C1","overallScore":8,"rationale":"질문 틀 자체를 재정의하는 통찰."},{"segmentIndex":56,"text":"But now you can just take these documents and uh basically uh recompile them in a different way, and uh reorder them, and create something that is uh new and interesting uh as a reframing of the data.","startTime":429.76,"endTime":440.28,"durationSeconds":11,"level":"C1","overallScore":7.8,"rationale":"새 처리 가능성을 구체적으로 설명함."},{"segmentIndex":58,"text":"Uh and so, I think this is uh something that I keep trying to get back to as to not only what can we do that existed that is faster now, but I think there's new opportunities of just things that couldn't be possible before.","startTime":443.48,"endTime":455,"durationSeconds":12,"level":"C1","overallScore":8,"rationale":"논지의 핵심을 잘 압축한 문장."},{"segmentIndex":60,"text":"If you extrapolate that further, what is the 2026 equivalent for building websites in the '90s, building mobile apps in the 2010s, building SaaS in the last cloud era?","startTime":469.88,"endTime":482.48,"durationSeconds":13,"level":"C1","overallScore":8,"rationale":"미래 비유를 묻는 통찰적 질문."},{"segmentIndex":71,"text":"So, you could imagine something really weird and foreign when where neural nets are doing most of the heavy lifting, they're using tool use as just like, you know, historical appendage for some kinds of like deterministic tasks.","startTime":553.96,"endTime":565.24,"durationSeconds":11,"level":"C1","overallScore":8,"rationale":"핵심 관점이 선명하고 표현도 풍부함."},{"segmentIndex":76,"text":"The fact that AI will automate faster and more easily domains where the output can be verified.","startTime":585.12,"endTime":591,"durationSeconds":6,"level":"B2","overallScore":7.8,"rationale":"검증가능성 원리를 명확히 제시함."},{"segmentIndex":4,"text":"So, they are given a verification rewards. And then because of the way that these models are trained, they end up basically uh, progressing and creating these like jagged entities that really peak in capability in kind of like verifiable domains like math and code and adjacent.","startTime":624.52,"endTime":638.76,"durationSeconds":14,"level":"C2","overallScore":7.8,"rationale":"모델 능력의 들쭉날쭉함 원인 설명."},{"segmentIndex":14,"text":"How is it possible that state-of-the-art Opus 4.7 will simultaneously refactor a 100,000 like >> [laughter] >> code base a line code base or find zero-day vulnerabilities and yet tells me to walk to this car wash?","startTime":700.64,"endTime":714.16,"durationSeconds":14,"level":"C2","overallScore":7.8,"rationale":"능력 대비 오류의 역설을 강하게 제시."}],"generatedAt":"2026-06-25T00:32:22.935Z","keyClipsTotalSec":778},{"videoId":"96jN2OCOfLs","chunkIndex":2,"totalChunks":3,"title":"Andrej Karpathy: From Vibe Coding to Agentic Engineering — Part 3 of 3","thumbnail":"https://i.ytimg.com/vi/96jN2OCOfLs/maxresdefault.jpg","duration":1789,"uploader":"Sequoia Capital","youtubeUrl":"https://www.youtube.com/watch?v=96jN2OCOfLs","keywords":["agentic-ai","llm","software-engineering","automation","human-ai-collaboration","product-design","education","future-of-work","prompting","ai-agents"],"normalizedKeywords":["엔지니어링","프로덕트","교육"],"targetAudience":[{"who":"엔지니어","why":"에이전트가 실수하는 지점과 설계자가 지켜야 할 원칙을 배울 수 있음"},{"who":"창업자","why":"에이전트 네이티브 제품과 인프라가 어디로 가는지 감을 잡을 수 있음"},{"who":"학생","why":"지능이 싸질수록 무엇을 깊게 배워야 하는지 이해에 도움 됨"}],"normalizedAudience":["엔지니어·개발자","창업자·스타트업","학생·주니어"],"summary":"이 대화는 에이전트가 실제로 일을 맡는 시대에 인간이 무엇을 맡아야 하는지에 대한 탐구다. 카르파티는 아직도 에이전트가 이메일 주소 같은 사소하지만 치명적인 디테일에서 자주 틀리기 때문에, 사람은 스펙·아키텍처·우선순위 같은 상위 설계를 책임져야 한다고 말한다. 동시에 코드 품질과 미학, 단순화 능력은 아직 모델이 잘 못하는 영역이라며, 인간의 taste와 judgment가 여전히 중요하다고 본다.\n\n이어 그는 LLM을 '동물'보다 '유령'에 가깝다고 묘사하며, 데이터와 보상함수로 만들어진 jagged intelligence를 이해하는 프레임이 사용·배포·평가 방식까지 바꿀 수 있다고 설명한다. 더 나아가 소프트웨어와 문서, 인프라가 사람보다 에이전트를 먼저 고려하는 'agent-native' 방향으로 재설계되어야 한다고 주장한다. 마지막으로 교육의 핵심은 사고를 외주화하는 것이 아니라 이해를 유지하는 것이라며, 지식베이스와 질문 도구는 이해를 강화하는 수단이어야 한다고 정리한다.","insights":["에이전트 시대엔 세부 실행보다 스펙 설계가 더 중요하다.","모델의 약점은 사소한 디테일보다 지속적 기억과 판단이다.","코드 생성보다 코드의 단순함과 미학을 아직 인간이 더 본다.","인프라는 사람용이 아니라 에이전트용으로 다시 짜여야 한다.","생각은 외주화돼도 이해는 외주화될 수 없다."],"keyClips":[{"clipId":"96jN2OCOfLs:c2:1-19","startSegmentIndex":1,"endSegmentIndex":19,"startTime":1200.87,"endTime":1333.6,"durationSeconds":132.7,"preview":"에이전트와 스펙 설계","mustSee":true},{"clipId":"96jN2OCOfLs:c2:20-30","startSegmentIndex":20,"endSegmentIndex":30,"startTime":1333.6,"endTime":1411.12,"durationSeconds":77.5,"preview":"taste와 코드 품질","mustSee":false},{"clipId":"96jN2OCOfLs:c2:31-43","startSegmentIndex":31,"endSegmentIndex":43,"startTime":1411.12,"endTime":1516.48,"durationSeconds":105.4,"preview":"유령 같은 지능","mustSee":false},{"clipId":"96jN2OCOfLs:c2:44-62","startSegmentIndex":44,"endSegmentIndex":62,"startTime":1516.48,"endTime":1658.04,"durationSeconds":141.6,"preview":"에이전트 네이티브 세계","mustSee":true},{"clipId":"96jN2OCOfLs:c2:64-78","startSegmentIndex":64,"endSegmentIndex":78,"startTime":1663.36,"endTime":1780.36,"durationSeconds":117,"preview":"이해는 외주화 못한다","mustSee":true}],"curatedSegments":[{"segmentIndex":25,"text":"And so that's why I think I'm stressing this dimension of it as we are slightly at the mercy of whatever the labs are doing, whatever they happen to put into the mix and you have to actually explore this thing that they give you that has no manual and it works in certain settings but maybe not in some settings and you have to kind of explore it a little bit and if you're in the circuits that were part of the RL, you fly and if you're in the circuits that are out of the data distribution, you're going to struggle and you have to kind of figure out which circuits you're in your application.","startTime":777.12,"endTime":807.12,"durationSeconds":30,"level":"C2","overallScore":8.8,"rationale":"실무 적용 한계와 탐색 원칙을 압축함."},{"segmentIndex":34,"text":"Um and these are jagged forms of intelligence that are shaped by data and reward functions, but not by intrinsic motivation or fun or curiosity or empowerment, uh things that kind of came about via evolution.","startTime":1425.76,"endTime":1439.6,"durationSeconds":14,"level":"C1","overallScore":8.6,"rationale":"AI 지능의 성격을 깊게 규정함."},{"segmentIndex":67,"text":"It was something along the lines of um you can outsource your thinking, but you can't outsource your understanding.","startTime":1690.24,"endTime":1696.48,"durationSeconds":6,"level":"B2","overallScore":8.6,"rationale":"대조적 통찰과 기억할 표현이 강함."},{"segmentIndex":15,"text":"I think that a lot of people actually I tried to stress this on uh Twitter and or X because I think a lot of people experienced AI uh last year as ChatGPT adjacent thing, uh but you really had to look again, and you had to look as of December uh because things have changed fundamentally and uh especially on this like agentic coherent workflow that really started to actually work.","startTime":103.36,"endTime":125.04,"durationSeconds":22,"level":"C1","overallScore":8,"rationale":"변화 시점과 이유를 통찰적으로 설명."},{"segmentIndex":27,"text":"So software 3.0 is kind of about uh you know, your programming now turns to prompting and what's in the context window is your lever over the interpreter that is the LLM that is kind of like interpreting your context and uh performing computation in the digital information space.","startTime":200.36,"endTime":215.56,"durationSeconds":15,"level":"C2","overallScore":8,"rationale":"새 패러다임을 밀도 높게 설명함."},{"segmentIndex":35,"text":"And the reason this is a lot more powerful is you're working now in the software 3.0 paradigm where you don't have to precisely uh spell out, you know, all the individual details of that setup.","startTime":260.92,"endTime":269.36,"durationSeconds":8,"level":"C1","overallScore":8,"rationale":"왜 강력한지 원리를 분명히 설명함."},{"segmentIndex":36,"text":"The agent has its own intelligence that it packages up and then it kind of like follows the instructions and it looks at your environment, your computer, and it kind of like performs intelligent actions to make things work and debugs things in the loop.","startTime":269.36,"endTime":279.84000000000003,"durationSeconds":10,"level":"C1","overallScore":7.8,"rationale":"에이전트 작동 방식을 구체화함."},{"segmentIndex":44,"text":"And then I saw the software 3.0 version of this, which is which blew my mind, which is literally just take your photo, give it to Gemini, and say use Nano Banana to overlay the things onto the menu.\"","startTime":337.56,"endTime":350,"durationSeconds":12,"level":"C1","overallScore":7.8,"rationale":"기존 앱을 대체하는 전환이 강렬함."},{"segmentIndex":46,"text":"And this blew my mind because actually all of my menu gen is spurious.","startTime":359.52,"endTime":364.24,"durationSeconds":5,"level":"B2","overallScore":7.4,"rationale":"기존 작업의 무의미함을 통찰적으로 밝힘."},{"segmentIndex":47,"text":"It's working in the old paradigm that app shouldn't exist.","startTime":364.24,"endTime":367.44,"durationSeconds":3,"level":"B2","overallScore":7.6,"rationale":"패러다임 전환의 함의를 날카롭게 말함."},{"segmentIndex":49,"text":"It just um your neural network is doing more and more of the work, and your prompt or context is just the image, and the output is an image, and there's no need to have any of the app in between.","startTime":371.72,"endTime":382.56,"durationSeconds":11,"level":"C1","overallScore":7.8,"rationale":"중간 앱이 불필요한 이유를 설명함."},{"segmentIndex":50,"text":"Um so, I think that people have to kind of like reframe, you know, not to work in the existing paradigm of what things existed and just think about it as a speed up of what exists.","startTime":382.56,"endTime":393.2,"durationSeconds":11,"level":"C1","overallScore":8,"rationale":"사고 전환을 직접 요구하는 문장."},{"segmentIndex":51,"text":"It's actually like new things are available now. And going back to your programming question, it's not even I think that's also an example of working in the old mindset because it's not just about programming and programming becoming faster.","startTime":393.2,"endTime":403.44,"durationSeconds":10,"level":"C1","overallScore":8,"rationale":"질문 틀 자체를 재정의하는 통찰."},{"segmentIndex":56,"text":"But now you can just take these documents and uh basically uh recompile them in a different way, and uh reorder them, and create something that is uh new and interesting uh as a reframing of the data.","startTime":429.76,"endTime":440.28,"durationSeconds":11,"level":"C1","overallScore":7.8,"rationale":"새 처리 가능성을 구체적으로 설명함."},{"segmentIndex":58,"text":"Uh and so, I think this is uh something that I keep trying to get back to as to not only what can we do that existed that is faster now, but I think there's new opportunities of just things that couldn't be possible before.","startTime":443.48,"endTime":455,"durationSeconds":12,"level":"C1","overallScore":8,"rationale":"논지의 핵심을 잘 압축한 문장."},{"segmentIndex":60,"text":"If you extrapolate that further, what is the 2026 equivalent for building websites in the '90s, building mobile apps in the 2010s, building SaaS in the last cloud era?","startTime":469.88,"endTime":482.48,"durationSeconds":13,"level":"C1","overallScore":8,"rationale":"미래 비유를 묻는 통찰적 질문."},{"segmentIndex":71,"text":"So, you could imagine something really weird and foreign when where neural nets are doing most of the heavy lifting, they're using tool use as just like, you know, historical appendage for some kinds of like deterministic tasks.","startTime":553.96,"endTime":565.24,"durationSeconds":11,"level":"C1","overallScore":8,"rationale":"핵심 관점이 선명하고 표현도 풍부함."},{"segmentIndex":76,"text":"The fact that AI will automate faster and more easily domains where the output can be verified.","startTime":585.12,"endTime":591,"durationSeconds":6,"level":"B2","overallScore":7.8,"rationale":"검증가능성 원리를 명확히 제시함."},{"segmentIndex":4,"text":"So, they are given a verification rewards. And then because of the way that these models are trained, they end up basically uh, progressing and creating these like jagged entities that really peak in capability in kind of like verifiable domains like math and code and adjacent.","startTime":624.52,"endTime":638.76,"durationSeconds":14,"level":"C2","overallScore":7.8,"rationale":"모델 능력의 들쭉날쭉함 원인 설명."},{"segmentIndex":14,"text":"How is it possible that state-of-the-art Opus 4.7 will simultaneously refactor a 100,000 like >> [laughter] >> code base a line code base or find zero-day vulnerabilities and yet tells me to walk to this car wash?","startTime":700.64,"endTime":714.16,"durationSeconds":14,"level":"C2","overallScore":7.8,"rationale":"능력 대비 오류의 역설을 강하게 제시."}],"generatedAt":"2026-06-25T00:32:48.651Z","keyClipsTotalSec":778}]}