{ "text": "最近有一個叫Jev的AI Model非常爆紅\n大家如果有上網\n應該都見到很多人討論這個AI Model\n這個叫Jev的AI Model\n其實是由一間叫TypeSafe AI的公司發佈的\n這間公司已經成立了2年\n只不過最近才發佈了這個Jev模型\n同時亦宣佈完成了一輪融資\n應該是Seed Round 籌集了大約4000萬美元\n這個叫Jev的AI Model 首先它不是大型語言模型\n例如ChatGPT、Gemini這些才是大型語言模型\n它們的性質像一個Chatbot 會跟你聊天\n它們執行你的任務之餘 亦可能會說很多廢話\n因為LLM的本質就是不停重複\n去Generate the Next Token\n但是這個Jev的AI Model 完全不是這樣的東西\n它技術方面的東西我稍後再說\n我先說一下它的創辦人\nTypeSafe AI這間公司的創辦人\n就是Diogo Almeida\n他之前在OpenAI工作\n他亦有參與InstructGPT這篇論文\n是作者之一 並有參與這項研究\nInstructGPT當然是非常重要\n因為例如在GPT-3之前\n大型語言模型的能力或訓練目標\n都是很單純地預測下一個字、下一個Token\n但自從OpenAI研究了InstructGPT之後\n它能夠將大型語言模型的Output\n與人類的意圖保持一致\n例如人類交給大型語言模型的任務\n可能你叫模型寫一封辭職信或一份計劃書\n模型就會根據你的任務和文字\n或計劃書的目的 去寫出那段文字\n自從有了InstructGPT之後\n大型語言模型才真正可以做到這樣的功能\n並且做得很準確\n而且InstructGPT亦確立了 強化學習的技術模式\n這個叫Jev的AI Model本質上是一個Decision Engine\n或者更準確地說 它算是一個Classifier Model\n舉一個例子 你有一封電郵和電郵的文字\n你想知道這封電郵究竟是客戶電郵\n還是垃圾郵件\n使用方法就是將電郵的文字\n交給這個Jev的AI Model\n然後給它幾個選擇\nA是客戶電郵 還是B是垃圾郵件\n如果它告訴你這是客戶電郵\n你可能想知道這封電郵是投訴 還是一般查詢\n如果是投訴 客戶現在是否很生氣\n模型不是寫文章 亦不會長篇大論\n它回傳的Output只是你給的選擇\n它會作出選擇 或給每個選擇一個機率\n又或者你問它是或不是\n它會給出是與不是的機率或分數\n為甚麼大家熱烈討論這個模型 這麼興奮\n原因是現在尤其是使用AI Agent時\nAI Agent很多時會呼叫大型語言模型\n最大的問題是 這些AI Agent不停呼叫大型語言模型時\n哪怕你只是做很少的事\n模型在背後其實真的會做Inference\n用很多Token長篇大論 雖然大部分人都不會看\n但事實上它就是消耗了這麼多Token\n其實Jev這個AI Model的運作並不複雜\n它和大型語言模型最重要的分別在於\n你除了要給它問題或任務的文字之外\n還必須要給一系列的選擇\n選擇可能很簡單 是一個是非題\nJev就在Yes or No之間做選擇\n或者分配一個概率\n或者是一個MCQ A至E五個答案中\n哪一個正確 或者機會率最大\n你最多可以給255個選擇\n記住選擇是必須的\n舉一個例子 現在一間診所裡有一位45歲女病人\n她的BP是184/112 沒有特別症狀\n亦沒有特別不適 測試結果正常\n究竟這個女人需不需要去急症室\n或者她的情況有多危急 有多緊急\n以1至5評分 5最緊急 1不緊急\nJev的應用場景就是這樣\n為方便一些做Tech的朋友試用Jev的AI Model\n我亦寫了一些Code作為Example\n給大家參考試用 亦已放上GitHub\n而最重要的是 我亦測試了一些任務\n當然我們做醫學人工智能\n剛才我提到那位女士血壓這麼高的情況下\n究竟有多危急 需不需要去A&E\n就是這些例子 當然這些例子本身是有答案的\n然後我們測試Jev在這些情境上表現好不好\nLLM的表現又好不好\n又或者你將Qwen 27B的Model 放到你的Mac上運行\n那就不需要Token 當然會較慢\n但它不需要花錢 亦是用你自己的電腦\n究竟效果又如何\n我們亦將結果放上去了\nI'm the king of the keys No crown on my dome\nChatGPT spitting scripts Making AI my home\nGemini twin flows Split brains in the zone\nDeepSeek visions Seeing truths never shown\nClaude in my corner Whisper moves like a sage\nLines so clean Got em popping off the page\nI don't CD That's alphabet soup\nI'm the vibe of the code Making systems salute\nI'm the king of coding No crown no throne\nTurn words into magic I'm a vibe on my own\nChatbots and cosmos Most the digital roam\nThe king of coding in the cloud I'm home\n接著我說一下我對Jev這個AI Model的看法\n根據TypeSafe AI所說\nJev這個AI Model的功能或定位\n又或者說它針對的痛點\n就是現今我們使用大型語言模型時\n尤其是使用Agent的時候\nAgent每一個步驟都會呼叫大型語言模型\n無論你用哪一個API 哪間公司的模型\n其實你每次呼叫的時候\n哪怕只是一個很簡單的決定\n或者一個很簡單的任務\n背後經常都會消耗很多Token\n它的論點是 這些Token是否真的必須這樣消耗\n因為老實說 你的AI Agent在背後做的事\n它在背後回傳的那些文字\n大部分時間只是AI Agent自己看 你不會看\n對大部分人來說 一些很Minor的決定\n你最重要是告訴我是或不是 做或不做\n或者用甚麼工具就可以\n為甚麼需要浪費這麼多Token\n這就是TypeSafe AI對Jev功能的看法\n又或者可以說是他們的Pitch\n毫無疑問 這個Pitch非常成功\n因為大家都感受到那種痛點\n因為大家用AI Agent去寫Program 寫Code\n你真的覺得好像在燒錢一樣\n但我的看法是 這是一個很好的Pitch\n但算不上是一個重大的技術突破\n因為如果從技術角度去看 我覺得Jev像一個Classifier Model\n當然它有理解語言的部分\n我相信這是Embedding Model做的事\n亦涉及一些Re-ranker的Algorithm\n但我剛才說的這幾樣東西絕對不是新技術\n因為你也可以想像 你說的是4000萬美元的Funding\n如果你要做大型語言模型\n尤其要與現在市面上的Frontier Model競爭\n根本不夠錢 4000萬美元不可能做LLM\n你唯一可能做到突破的\n就是在剛才說的Classifier Model\n再加Embedding Model的層次上去做\n所以我覺得這個Pitch很好 亦是很好的Marketing\n至少現在這麼多人在討論\n取得Traction之後再籌集更多資金\n再想下一步怎樣做\n這可能是他們現在的策略\n所以我不覺得這是一個重大的技術突破\n反而另一位 我想是他的同事 以前OpenAI的CTO 那位女士 我不記得她叫甚麼名字\n以前OpenAI的CTO\n那位女士 我不記得她叫甚麼名字\n她出來亦成立了一間AI公司\n那個反而算是一個技術突破\n因為她說的是用影像\n原生地作為語言模型的Input\n這其實是很難做到的\n我不是說她一定會成功\n但純粹從技術來說 前OpenAI CTO\n那位女士做的反而真的是技術突破\n而Jev不是\n所以要視乎你做的事、任務或決定有多複雜\n舉個例子 純粹做電郵分類\n可能你的公司每日有幾百封Email\n或者你做網店 每日有幾百個查詢\n你要對那些Email、查詢或客戶訊息\n純粹做分類 例如你的Ticket System\n要看那個Ticket有多Urgent\n如果很急 可能你會分為極急\n立即處理 客戶很生氣就先處理\n一般查詢就稍後\n如果純粹是電郵或客戶訊息\n用Jev這類模型 我覺得是對的\n因為一來成本會較便宜\n二來反應時間會較快\n而且這類System很多時未必即時自動回覆\n可能都會經過人手客服再覆核\n在這類情況下用Jev是合理的\n或者用Jev分類後再交給LLM處理亦可以\n但如果你做的事或任務有一定複雜程度\n你可能仍需要用回傳統語言模型\n就算你的任務純粹是做決定\n不涉及生成文章或回覆客戶Email\n不需要生成文字 純粹做決定\n我想你仍是用回傳統語言模型\n例如Claude、GPT這類Frontier Model\n它們的判斷做得好\n原因是它們可能會一步一步地思考\n它們有Chain of Thought\n你可以選擇Thinking是High還是Low\n用多少思考力去做\n複雜的任務或判斷\n你可能真的需要看當中的細節\n或者重點 才可以作出結論\n而這個過程是做Chain of Thought\n真的要消耗Token\n你不可以說是浪費 不可以說不是必須\n很多時有些Tasks要看很多細節\n你不看細節 不讓它一步一步思考\n就有很大機會出錯\n例如一個男人有性病\n他是否一定出去鬼混 未必\n表面看似是這樣 但不一定\n他的女朋友濫交又行不行\n或者那個男人走過砵蘭街\n被水淋到後染上性病又行不行 很無辜\n我說到這裡 一定有些有識之士會在留言更正我\n如果你走過砵蘭街被水淋到\n在醫學上是絕對不會染上性病的\n因為DeepSeek、Gemini都是這樣說\n其實我想說 我做這些節目\n或這些題目這麼悶\n有時搞搞笑 搞搞氣氛都可以\n我覺得有時大家缺乏了一些幽默感\n就像上次我的影片 我說梅建國\n真的有網友很認真地留言說\n梅建國其實是網民搞笑的\nAllen竟然當真\n然後說這就證明了\nAI幻覺的最佳例證\n老實說 美斯是否真的叫梅建國 我真的不知道\n我不肯定 但我可以很肯定地告訴大家\n美斯是王敏德\n另一個問題是 OpenAI、Gemini\n或DeepSeek這類模型\n現在亦有推出所謂Mini或Flash版本\n它們針對的應用場景\n是一些不需要太多思考又要快的\n任務或工作\n當然Jev可以反駁\n這些Flash Model、Mini Model都會消耗Token\n但就我個人的使用經驗而言\n未必代表大多數人\n但這類Flash Model\n無論你的prompt怎樣寫 我覺得都夠快亦夠便宜\n當然你不可以將它與Jev這類\nClassifier Model性質的東西比較\n但你說Flash Model是否真的很貴 我又覺得不是\n如果要我衡量任務的準確性\n即做出來的準確性 或判斷的準確性\n相對速度的Trade Off\n我寧願做慢一點 而不犧牲準確性\n如果你為了省錢\n而用Jev這類Model令你的決策出錯\n我覺得不太值得\n就算要等AI Agent 你先喝杯咖啡吧\n我覺得其實不是那麼Time Critical的事\n我覺得交回LLM去做會比較放心\n這就是我的結論\n但正如我剛才所說 他們的Pitch做得好\n因為這的確是一個痛點\n如果用Jev做一個Cheap Filter\n在大型語言模型之前先過濾\n或者有些任務很簡單就可以作決定\n我覺得用Jev是合理的\n但如果完全用Jev取代所有Decision Making\n這點我就不同意了\n這個叫Jev的AI Model 其實是向Jevons Paradox致敬\nJevons Paradox是一個經濟學現象\n它說的是當技術進步到一定程度\n令某種資源的使用效率提高時\n該資源的總消耗量 我們很多時以為\n提高效率 整體上會用少了那些資源\n但很多時發現並不是\n其中一個例子是引擎的改進\n技術進步令引擎更省油\n省油之後 我們以為對燃油的整體需求會減少\n其實不但沒有減少 反而上升\n另一個例子是電訊 即數據傳輸技術\n例如4G、5G\n技術進步令數據傳輸速度更快\n或成本降低\n我們會否以為人們會用少了數據\n答案當然不是 現在的人不只看文字\n不只看圖片 就算看影片都可能要求高\n高質素的4K、8K串流影片\n或者會做更多以前沒有的事\n例如線上會議或線上遊戲等等\n數據的總消耗量\n並沒有因為技術提升而下跌\n其實Jevons Paradox就是說這樣的事\n好 時間到了 多謝大家", "language": "zh", "duration": 1168.0, "segments": [ { "id": 0, "start": 0.0, "end": 3.7, "text": "最近有一個叫Jev的AI Model非常爆紅", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 1, "start": 3.96, "end": 5.46, "text": "大家如果有上網", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 2, "start": 5.64, "end": 8.92, "text": "應該都見到很多人討論這個AI Model", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 3, "start": 9.7, "end": 11.5, "text": "這個叫Jev的AI Model", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 4, "start": 11.7, "end": 15.92, "text": "其實是由一間叫TypeSafe AI的公司發佈的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 5, "start": 15.92, "end": 18.64, "text": "這間公司已經成立了2年", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 6, "start": 18.86, "end": 22.92, "text": "只不過最近才發佈了這個Jev模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 7, "start": 23.18, "end": 27.98, "text": "同時亦宣佈完成了一輪融資", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 8, "start": 28.1, "end": 31.84, "text": "應該是Seed Round 籌集了大約4000萬美元", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 9, "start": 33.06, "end": 38.7, "text": "這個叫Jev的AI Model 首先它不是大型語言模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 10, "start": 38.94, "end": 44.44, "text": "例如ChatGPT、Gemini這些才是大型語言模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 11, "start": 44.58, "end": 47.8, "text": "它們的性質像一個Chatbot 會跟你聊天", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 12, "start": 48.68, "end": 53.5, "text": "它們執行你的任務之餘 亦可能會說很多廢話", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 13, "start": 53.84, "end": 57.86, "text": "因為LLM的本質就是不停重複", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 14, "start": 58.68, "end": 60.48, "text": "去Generate the Next Token", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 15, "start": 61.1, "end": 65.9, "text": "但是這個Jev的AI Model 完全不是這樣的東西", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 16, "start": 66.4, "end": 69.08, "text": "它技術方面的東西我稍後再說", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 17, "start": 69.08, "end": 71.72, "text": "我先說一下它的創辦人", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 18, "start": 71.72, "end": 75.86, "text": "TypeSafe AI這間公司的創辦人", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 19, "start": 75.86, "end": 77.7, "text": "就是Diogo Almeida", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 20, "start": 78.33, "end": 81.44, "text": "他之前在OpenAI工作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 21, "start": 81.44, "end": 88.34, "text": "他亦有參與InstructGPT這篇論文", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 22, "start": 89.54, "end": 91.78, "text": "是作者之一 並有參與這項研究", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 23, "start": 91.98, "end": 95.16, "text": "InstructGPT當然是非常重要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 24, "start": 95.72, "end": 100.82, "text": "因為例如在GPT-3之前", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 25, "start": 101.68, "end": 108.92, "text": "大型語言模型的能力或訓練目標", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 26, "start": 109.14, "end": 112.98, "text": "都是很單純地預測下一個字、下一個Token", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 27, "start": 113.58, "end": 120.68, "text": "但自從OpenAI研究了InstructGPT之後", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 28, "start": 120.76, "end": 126.04, "text": "它能夠將大型語言模型的Output", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 29, "start": 126.04, "end": 132.5, "text": "與人類的意圖保持一致", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 30, "start": 132.5, "end": 137.34, "text": "例如人類交給大型語言模型的任務", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 31, "start": 137.5, "end": 142.9, "text": "可能你叫模型寫一封辭職信或一份計劃書", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 32, "start": 142.9, "end": 147.14, "text": "模型就會根據你的任務和文字", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 33, "start": 147.68, "end": 151.84, "text": "或計劃書的目的 去寫出那段文字", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 34, "start": 153.24, "end": 157.56, "text": "自從有了InstructGPT之後", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 35, "start": 158.18, "end": 162.82, "text": "大型語言模型才真正可以做到這樣的功能", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 36, "start": 162.82, "end": 164.56, "text": "並且做得很準確", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 37, "start": 164.72, "end": 171.74, "text": "而且InstructGPT亦確立了 強化學習的技術模式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 38, "start": 174.36, "end": 178.94, "text": "這個叫Jev的AI Model本質上是一個Decision Engine", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 39, "start": 179.42, "end": 183.7, "text": "或者更準確地說 它算是一個Classifier Model", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 40, "start": 184.28, "end": 190.32, "text": "舉一個例子 你有一封電郵和電郵的文字", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 41, "start": 190.56, "end": 194.5, "text": "你想知道這封電郵究竟是客戶電郵", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 42, "start": 194.5, "end": 197.36, "text": "還是垃圾郵件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 43, "start": 198.16, "end": 203.32, "text": "使用方法就是將電郵的文字", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 44, "start": 203.32, "end": 205.12, "text": "交給這個Jev的AI Model", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 45, "start": 205.12, "end": 209.32, "text": "然後給它幾個選擇", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 46, "start": 209.34, "end": 211.98, "text": "A是客戶電郵 還是B是垃圾郵件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 47, "start": 212.5, "end": 215.18, "text": "如果它告訴你這是客戶電郵", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 48, "start": 215.18, "end": 221.5, "text": "你可能想知道這封電郵是投訴 還是一般查詢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 49, "start": 222.14, "end": 225.14, "text": "如果是投訴 客戶現在是否很生氣", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 50, "start": 225.78, "end": 230.0, "text": "模型不是寫文章 亦不會長篇大論", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 51, "start": 230.0, "end": 234.92, "text": "它回傳的Output只是你給的選擇", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 52, "start": 235.04, "end": 238.8, "text": "它會作出選擇 或給每個選擇一個機率", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 53, "start": 238.96, "end": 242.96, "text": "又或者你問它是或不是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 54, "start": 242.96, "end": 245.56, "text": "它會給出是與不是的機率或分數", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 55, "start": 247.38, "end": 252.0, "text": "為甚麼大家熱烈討論這個模型 這麼興奮", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 56, "start": 252.0, "end": 258.26, "text": "原因是現在尤其是使用AI Agent時", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 57, "start": 259.4, "end": 265.24, "text": "AI Agent很多時會呼叫大型語言模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 58, "start": 266.12, "end": 271.64, "text": "最大的問題是 這些AI Agent不停呼叫大型語言模型時", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 59, "start": 271.88, "end": 274.64, "text": "哪怕你只是做很少的事", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 60, "start": 274.88, "end": 278.64, "text": "模型在背後其實真的會做Inference", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 61, "start": 278.64, "end": 283.0, "text": "用很多Token長篇大論 雖然大部分人都不會看", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 62, "start": 283.56, "end": 286.6, "text": "但事實上它就是消耗了這麼多Token", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 63, "start": 289.42, "end": 293.16, "text": "其實Jev這個AI Model的運作並不複雜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 64, "start": 293.34, "end": 297.02, "text": "它和大型語言模型最重要的分別在於", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 65, "start": 297.02, "end": 301.82, "text": "你除了要給它問題或任務的文字之外", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 66, "start": 301.82, "end": 305.86, "text": "還必須要給一系列的選擇", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 67, "start": 306.1, "end": 308.22, "text": "選擇可能很簡單 是一個是非題", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 68, "start": 308.22, "end": 312.24, "text": "Jev就在Yes or No之間做選擇", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 69, "start": 312.32, "end": 314.54, "text": "或者分配一個概率", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 70, "start": 315.04, "end": 320.08, "text": "或者是一個MCQ A至E五個答案中", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 71, "start": 320.82, "end": 323.4, "text": "哪一個正確 或者機會率最大", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 72, "start": 324.02, "end": 329.38, "text": "你最多可以給255個選擇", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 73, "start": 330.2, "end": 333.02, "text": "記住選擇是必須的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 74, "start": 333.58, "end": 340.98, "text": "舉一個例子 現在一間診所裡有一位45歲女病人", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 75, "start": 341.12, "end": 345.48, "text": "她的BP是184/112 沒有特別症狀", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 76, "start": 345.48, "end": 350.52, "text": "亦沒有特別不適 測試結果正常", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 77, "start": 350.74, "end": 355.64, "text": "究竟這個女人需不需要去急症室", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 78, "start": 355.76, "end": 360.1, "text": "或者她的情況有多危急 有多緊急", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 79, "start": 360.1, "end": 364.18, "text": "以1至5評分 5最緊急 1不緊急", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 80, "start": 364.32, "end": 367.8, "text": "Jev的應用場景就是這樣", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 81, "start": 368.72, "end": 376.54, "text": "為方便一些做Tech的朋友試用Jev的AI Model", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 82, "start": 376.54, "end": 381.48, "text": "我亦寫了一些Code作為Example", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 83, "start": 382.04, "end": 385.54, "text": "給大家參考試用 亦已放上GitHub", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 84, "start": 385.82, "end": 389.94, "text": "而最重要的是 我亦測試了一些任務", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 85, "start": 389.98, "end": 392.04, "text": "當然我們做醫學人工智能", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 86, "start": 397.06, "end": 400.78, "text": "剛才我提到那位女士血壓這麼高的情況下", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 87, "start": 400.96, "end": 403.66, "text": "究竟有多危急 需不需要去A&E", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 88, "start": 403.96, "end": 409.3, "text": "就是這些例子 當然這些例子本身是有答案的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 89, "start": 409.94, "end": 417.5, "text": "然後我們測試Jev在這些情境上表現好不好", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 90, "start": 418.16, "end": 419.8, "text": "LLM的表現又好不好", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 91, "start": 419.86, "end": 425.34, "text": "又或者你將Qwen 27B的Model 放到你的Mac上運行", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 92, "start": 425.34, "end": 427.56, "text": "那就不需要Token 當然會較慢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 93, "start": 427.66, "end": 430.62, "text": "但它不需要花錢 亦是用你自己的電腦", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 94, "start": 431.0, "end": 433.24, "text": "究竟效果又如何", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 95, "start": 433.36, "end": 436.12, "text": "我們亦將結果放上去了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 96, "start": 447.38, "end": 449.68, "text": "I'm the king of the keys No crown on my dome", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 97, "start": 449.68, "end": 452.78, "text": "ChatGPT spitting scripts Making AI my home", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 98, "start": 452.78, "end": 455.78, "text": "Gemini twin flows Split brains in the zone", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 99, "start": 455.78, "end": 458.42, "text": "DeepSeek visions Seeing truths never shown", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 100, "start": 458.42, "end": 460.9, "text": "Claude in my corner Whisper moves like a sage", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 101, "start": 460.9, "end": 463.58, "text": "Lines so clean Got em popping off the page", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 102, "start": 463.58, "end": 465.78, "text": "I don't CD That's alphabet soup", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 103, "start": 465.78, "end": 468.42, "text": "I'm the vibe of the code Making systems salute", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 104, "start": 468.42, "end": 471.22, "text": "I'm the king of coding No crown no throne", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 105, "start": 471.22, "end": 473.98, "text": "Turn words into magic I'm a vibe on my own", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 106, "start": 473.98, "end": 476.58, "text": "Chatbots and cosmos Most the digital roam", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 107, "start": 476.58, "end": 479.42, "text": "The king of coding in the cloud I'm home", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 108, "start": 482.48, "end": 486.74, "text": "接著我說一下我對Jev這個AI Model的看法", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 109, "start": 487.9, "end": 489.78, "text": "根據TypeSafe AI所說", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 110, "start": 489.96, "end": 495.56, "text": "Jev這個AI Model的功能或定位", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 111, "start": 496.54, "end": 499.4, "text": "又或者說它針對的痛點", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 112, "start": 499.4, "end": 502.46, "text": "就是現今我們使用大型語言模型時", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 113, "start": 502.56, "end": 505.32, "text": "尤其是使用Agent的時候", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 114, "start": 505.56, "end": 510.02, "text": "Agent每一個步驟都會呼叫大型語言模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 115, "start": 510.2, "end": 512.98, "text": "無論你用哪一個API 哪間公司的模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 116, "start": 513.12, "end": 515.6, "text": "其實你每次呼叫的時候", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 117, "start": 515.76, "end": 519.88, "text": "哪怕只是一個很簡單的決定", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 118, "start": 520.32, "end": 522.12, "text": "或者一個很簡單的任務", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 119, "start": 522.2, "end": 525.98, "text": "背後經常都會消耗很多Token", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 120, "start": 526.8, "end": 531.08, "text": "它的論點是 這些Token是否真的必須這樣消耗", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 121, "start": 531.18, "end": 535.24, "text": "因為老實說 你的AI Agent在背後做的事", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 122, "start": 535.88, "end": 539.48, "text": "它在背後回傳的那些文字", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 123, "start": 539.48, "end": 543.5, "text": "大部分時間只是AI Agent自己看 你不會看", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 124, "start": 543.98, "end": 549.34, "text": "對大部分人來說 一些很Minor的決定", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 125, "start": 549.86, "end": 552.2, "text": "你最重要是告訴我是或不是 做或不做", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 126, "start": 552.2, "end": 554.26, "text": "或者用甚麼工具就可以", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 127, "start": 554.4, "end": 557.3, "text": "為甚麼需要浪費這麼多Token", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 128, "start": 557.5, "end": 561.7, "text": "這就是TypeSafe AI對Jev功能的看法", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 129, "start": 561.88, "end": 564.34, "text": "又或者可以說是他們的Pitch", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 130, "start": 565.48, "end": 568.6, "text": "毫無疑問 這個Pitch非常成功", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 131, "start": 568.66, "end": 571.96, "text": "因為大家都感受到那種痛點", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 132, "start": 572.08, "end": 577.68, "text": "因為大家用AI Agent去寫Program 寫Code", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 133, "start": 577.68, "end": 579.9, "text": "你真的覺得好像在燒錢一樣", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 134, "start": 580.86, "end": 585.38, "text": "但我的看法是 這是一個很好的Pitch", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 135, "start": 585.8, "end": 590.04, "text": "但算不上是一個重大的技術突破", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 136, "start": 590.48, "end": 596.96, "text": "因為如果從技術角度去看 我覺得Jev像一個Classifier Model", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 137, "start": 597.24, "end": 600.48, "text": "當然它有理解語言的部分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 138, "start": 600.56, "end": 602.64, "text": "我相信這是Embedding Model做的事", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 139, "start": 602.76, "end": 607.22, "text": "亦涉及一些Re-ranker的Algorithm", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 140, "start": 607.92, "end": 613.44, "text": "但我剛才說的這幾樣東西絕對不是新技術", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 141, "start": 613.74, "end": 619.58, "text": "因為你也可以想像 你說的是4000萬美元的Funding", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 142, "start": 619.86, "end": 623.82, "text": "如果你要做大型語言模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 143, "start": 624.06, "end": 627.24, "text": "尤其要與現在市面上的Frontier Model競爭", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 144, "start": 627.34, "end": 631.18, "text": "根本不夠錢 4000萬美元不可能做LLM", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 145, "start": 631.6, "end": 638.14, "text": "你唯一可能做到突破的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 146, "start": 638.14, "end": 640.72, "text": "就是在剛才說的Classifier Model", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 147, "start": 640.98, "end": 644.12, "text": "再加Embedding Model的層次上去做", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 148, "start": 644.84, "end": 648.26, "text": "所以我覺得這個Pitch很好 亦是很好的Marketing", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 149, "start": 648.4, "end": 650.28, "text": "至少現在這麼多人在討論", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 150, "start": 650.48, "end": 653.84, "text": "取得Traction之後再籌集更多資金", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 151, "start": 654.02, "end": 655.6, "text": "再想下一步怎樣做", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 152, "start": 655.72, "end": 657.84, "text": "這可能是他們現在的策略", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 153, "start": 658.6, "end": 663.72, "text": "所以我不覺得這是一個重大的技術突破", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 154, "start": 663.8, "end": 665.98, "text": "反而另一位 我想是他的同事 以前OpenAI的CTO 那位女士 我不記得她叫甚麼名字", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "反而另一位 我想是他的同事", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 155, "start": 666.06, "end": 669.42, "text": "以前OpenAI的CTO", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 156, "start": 669.5, "end": 671.78, "text": "那位女士 我不記得她叫甚麼名字", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 157, "start": 672.56, "end": 675.84, "text": "她出來亦成立了一間AI公司", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 158, "start": 675.96, "end": 678.18, "text": "那個反而算是一個技術突破", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 159, "start": 678.26, "end": 680.2, "text": "因為她說的是用影像", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 160, "start": 681.14, "end": 685.2, "text": "原生地作為語言模型的Input", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 161, "start": 685.7, "end": 687.36, "text": "這其實是很難做到的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 162, "start": 687.52, "end": 690.7, "text": "我不是說她一定會成功", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 163, "start": 690.7, "end": 694.62, "text": "但純粹從技術來說 前OpenAI CTO", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 164, "start": 695.4, "end": 698.52, "text": "那位女士做的反而真的是技術突破", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 165, "start": 698.7, "end": 700.2, "text": "而Jev不是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 166, "start": 703.14, "end": 709.0, "text": "所以要視乎你做的事、任務或決定有多複雜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 167, "start": 709.38, "end": 712.2, "text": "舉個例子 純粹做電郵分類", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 168, "start": 712.2, "end": 714.92, "text": "可能你的公司每日有幾百封Email", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 169, "start": 715.62, "end": 718.98, "text": "或者你做網店 每日有幾百個查詢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 170, "start": 719.1, "end": 725.84, "text": "你要對那些Email、查詢或客戶訊息", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 171, "start": 726.0, "end": 728.5, "text": "純粹做分類 例如你的Ticket System", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 172, "start": 728.5, "end": 730.98, "text": "要看那個Ticket有多Urgent", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 173, "start": 731.48, "end": 734.7, "text": "如果很急 可能你會分為極急", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 174, "start": 734.7, "end": 736.8, "text": "立即處理 客戶很生氣就先處理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 175, "start": 736.8, "end": 738.56, "text": "一般查詢就稍後", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 176, "start": 739.3, "end": 743.5, "text": "如果純粹是電郵或客戶訊息", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 177, "start": 743.86, "end": 746.7, "text": "用Jev這類模型 我覺得是對的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 178, "start": 746.7, "end": 749.7, "text": "因為一來成本會較便宜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 179, "start": 749.82, "end": 752.2, "text": "二來反應時間會較快", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 180, "start": 752.2, "end": 756.4, "text": "而且這類System很多時未必即時自動回覆", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 181, "start": 756.94, "end": 759.42, "text": "可能都會經過人手客服再覆核", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 182, "start": 759.54, "end": 763.7, "text": "在這類情況下用Jev是合理的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 183, "start": 763.7, "end": 767.74, "text": "或者用Jev分類後再交給LLM處理亦可以", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 184, "start": 768.64, "end": 773.2, "text": "但如果你做的事或任務有一定複雜程度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 185, "start": 773.24, "end": 776.5, "text": "你可能仍需要用回傳統語言模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 186, "start": 776.5, "end": 780.74, "text": "就算你的任務純粹是做決定", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 187, "start": 780.86, "end": 784.18, "text": "不涉及生成文章或回覆客戶Email", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 188, "start": 784.52, "end": 786.96, "text": "不需要生成文字 純粹做決定", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 189, "start": 787.04, "end": 789.3, "text": "我想你仍是用回傳統語言模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 190, "start": 789.4, "end": 792.24, "text": "例如Claude、GPT這類Frontier Model", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 191, "start": 792.54, "end": 794.38, "text": "它們的判斷做得好", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 192, "start": 795.18, "end": 798.2, "text": "原因是它們可能會一步一步地思考", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 193, "start": 798.2, "end": 800.34, "text": "它們有Chain of Thought", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 194, "start": 801.08, "end": 805.98, "text": "你可以選擇Thinking是High還是Low", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 195, "start": 805.98, "end": 808.96, "text": "用多少思考力去做", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 196, "start": 809.5, "end": 812.2, "text": "複雜的任務或判斷", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 197, "start": 812.2, "end": 816.04, "text": "你可能真的需要看當中的細節", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 198, "start": 816.1, "end": 818.98, "text": "或者重點 才可以作出結論", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 199, "start": 819.14, "end": 822.82, "text": "而這個過程是做Chain of Thought", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 200, "start": 822.82, "end": 825.62, "text": "真的要消耗Token", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 201, "start": 825.62, "end": 828.68, "text": "你不可以說是浪費 不可以說不是必須", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 202, "start": 829.3, "end": 835.1, "text": "很多時有些Tasks要看很多細節", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 203, "start": 835.1, "end": 838.36, "text": "你不看細節 不讓它一步一步思考", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 204, "start": 838.36, "end": 840.64, "text": "就有很大機會出錯", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 205, "start": 840.64, "end": 842.82, "text": "例如一個男人有性病", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 206, "start": 842.9, "end": 845.98, "text": "他是否一定出去鬼混 未必", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 207, "start": 846.38, "end": 848.2, "text": "表面看似是這樣 但不一定", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 208, "start": 848.2, "end": 850.32, "text": "他的女朋友濫交又行不行", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 209, "start": 850.74, "end": 852.32, "text": "或者那個男人走過砵蘭街", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 210, "start": 852.48, "end": 855.2, "text": "被水淋到後染上性病又行不行 很無辜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 211, "start": 856.62, "end": 862.96, "text": "我說到這裡 一定有些有識之士會在留言更正我", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 212, "start": 863.18, "end": 866.14, "text": "如果你走過砵蘭街被水淋到", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 213, "start": 866.22, "end": 868.8, "text": "在醫學上是絕對不會染上性病的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 214, "start": 868.96, "end": 871.5, "text": "因為DeepSeek、Gemini都是這樣說", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 215, "start": 873.28, "end": 877.72, "text": "其實我想說 我做這些節目", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 216, "start": 877.9, "end": 880.0, "text": "或這些題目這麼悶", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 217, "start": 880.0, "end": 883.64, "text": "有時搞搞笑 搞搞氣氛都可以", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 218, "start": 883.64, "end": 887.82, "text": "我覺得有時大家缺乏了一些幽默感", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 219, "start": 888.62, "end": 892.26, "text": "就像上次我的影片 我說梅建國", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 220, "start": 893.26, "end": 898.76, "text": "真的有網友很認真地留言說", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 221, "start": 899.84, "end": 903.1, "text": "梅建國其實是網民搞笑的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 222, "start": 903.26, "end": 905.12, "text": "Allen竟然當真", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 223, "start": 905.2, "end": 908.88, "text": "然後說這就證明了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 224, "start": 909.04, "end": 912.08, "text": "AI幻覺的最佳例證", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 225, "start": 912.92, "end": 918.22, "text": "老實說 美斯是否真的叫梅建國 我真的不知道", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 226, "start": 918.22, "end": 922.44, "text": "我不肯定 但我可以很肯定地告訴大家", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 227, "start": 922.96, "end": 925.66, "text": "美斯是王敏德", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 228, "start": 928.26, "end": 931.36, "text": "另一個問題是 OpenAI、Gemini", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 229, "start": 931.66, "end": 935.18, "text": "或DeepSeek這類模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 230, "start": 935.28, "end": 938.96, "text": "現在亦有推出所謂Mini或Flash版本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 231, "start": 939.06, "end": 942.2, "text": "它們針對的應用場景", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 232, "start": 942.2, "end": 945.8, "text": "是一些不需要太多思考又要快的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 233, "start": 946.16, "end": 948.5, "text": "任務或工作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 234, "start": 949.3, "end": 952.72, "text": "當然Jev可以反駁", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 235, "start": 953.1, "end": 956.84, "text": "這些Flash Model、Mini Model都會消耗Token", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 236, "start": 957.7, "end": 961.98, "text": "但就我個人的使用經驗而言", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 237, "start": 962.06, "end": 963.68, "text": "未必代表大多數人", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 238, "start": 963.82, "end": 967.94, "text": "但這類Flash Model", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 239, "start": 968.52, "end": 972.6, "text": "無論你的prompt怎樣寫 我覺得都夠快亦夠便宜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 240, "start": 972.8, "end": 977.08, "text": "當然你不可以將它與Jev這類", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 241, "start": 977.76, "end": 980.42, "text": "Classifier Model性質的東西比較", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 242, "start": 980.62, "end": 985.76, "text": "但你說Flash Model是否真的很貴 我又覺得不是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 243, "start": 985.86, "end": 990.12, "text": "如果要我衡量任務的準確性", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 244, "start": 990.34, "end": 993.92, "text": "即做出來的準確性 或判斷的準確性", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 245, "start": 995.38, "end": 998.34, "text": "相對速度的Trade Off", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 246, "start": 998.48, "end": 1003.56, "text": "我寧願做慢一點 而不犧牲準確性", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 247, "start": 1003.72, "end": 1005.74, "text": "如果你為了省錢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 248, "start": 1006.64, "end": 1014.58, "text": "而用Jev這類Model令你的決策出錯", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 249, "start": 1014.72, "end": 1015.98, "text": "我覺得不太值得", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 250, "start": 1016.46, "end": 1019.72, "text": "就算要等AI Agent 你先喝杯咖啡吧", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 251, "start": 1019.84, "end": 1023.22, "text": "我覺得其實不是那麼Time Critical的事", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 252, "start": 1023.34, "end": 1028.5, "text": "我覺得交回LLM去做會比較放心", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 253, "start": 1029.48, "end": 1032.06, "text": "這就是我的結論", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 254, "start": 1032.2, "end": 1034.7, "text": "但正如我剛才所說 他們的Pitch做得好", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 255, "start": 1034.7, "end": 1037.66, "text": "因為這的確是一個痛點", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 256, "start": 1037.84, "end": 1042.12, "text": "如果用Jev做一個Cheap Filter", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 257, "start": 1042.3, "end": 1045.2, "text": "在大型語言模型之前先過濾", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 258, "start": 1045.2, "end": 1048.2, "text": "或者有些任務很簡單就可以作決定", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 259, "start": 1048.2, "end": 1050.98, "text": "我覺得用Jev是合理的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 260, "start": 1051.02, "end": 1057.04, "text": "但如果完全用Jev取代所有Decision Making", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 261, "start": 1057.1, "end": 1058.8, "text": "這點我就不同意了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 262, "start": 1061.82, "end": 1066.32, "text": "這個叫Jev的AI Model 其實是向Jevons Paradox致敬", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 263, "start": 1066.32, "end": 1072.16, "text": "Jevons Paradox是一個經濟學現象", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 264, "start": 1072.16, "end": 1075.62, "text": "它說的是當技術進步到一定程度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 265, "start": 1075.94, "end": 1079.8, "text": "令某種資源的使用效率提高時", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 266, "start": 1080.42, "end": 1084.1, "text": "該資源的總消耗量 我們很多時以為", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 267, "start": 1085.28, "end": 1089.22, "text": "提高效率 整體上會用少了那些資源", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 268, "start": 1089.88, "end": 1092.18, "text": "但很多時發現並不是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 269, "start": 1092.78, "end": 1096.7, "text": "其中一個例子是引擎的改進", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 270, "start": 1096.7, "end": 1101.36, "text": "技術進步令引擎更省油", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 271, "start": 1102.78, "end": 1108.58, "text": "省油之後 我們以為對燃油的整體需求會減少", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 272, "start": 1108.7, "end": 1112.42, "text": "其實不但沒有減少 反而上升", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 273, "start": 1112.62, "end": 1118.66, "text": "另一個例子是電訊 即數據傳輸技術", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 274, "start": 1119.14, "end": 1121.68, "text": "例如4G、5G", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 275, "start": 1122.48, "end": 1128.3, "text": "技術進步令數據傳輸速度更快", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 276, "start": 1129.16, "end": 1131.04, "text": "或成本降低", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 277, "start": 1131.16, "end": 1134.62, "text": "我們會否以為人們會用少了數據", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 278, "start": 1134.8, "end": 1138.22, "text": "答案當然不是 現在的人不只看文字", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 279, "start": 1138.3, "end": 1141.44, "text": "不只看圖片 就算看影片都可能要求高", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 280, "start": 1141.54, "end": 1145.3, "text": "高質素的4K、8K串流影片", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 281, "start": 1145.8, "end": 1148.06, "text": "或者會做更多以前沒有的事", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 282, "start": 1148.86, "end": 1151.5, "text": "例如線上會議或線上遊戲等等", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 283, "start": 1152.36, "end": 1155.28, "text": "數據的總消耗量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 284, "start": 1155.38, "end": 1159.64, "text": "並沒有因為技術提升而下跌", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 285, "start": 1159.8, "end": 1164.82, "text": "其實Jevons Paradox就是說這樣的事", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 286, "start": 1165.3, "end": 1167.44, "text": "好 時間到了 多謝大家", "source": "video_source_subtitle", "authority": "video_source_subtitle" } ], "transcript_selection": { "schema": "videonote.transcript_selection.v1", "status": "SELECTED", "requested_language": "zh", "duration": 1168.0, "reference_path": "/Users/sagawa/AI_WORK/video_notes/url/20260928_153332__科技專題_Jev很火_不聊天_不寫字_只做選擇_擊中LLM痛點_應用場景測試_Allen/reference_subtitle.vtt", "reference_kind": "manual", "reference_language": "zh-Hant", "selected_source": "video_source_subtitle", "source_caption_kind": "manual", "source_caption_processing": "verbatim_direct", "youtube_overlap_clamps": [], "general_high_confidence_repairs": [], "decision_reason": "complete_source_subtitle_fast_path_no_whisper", "whisper_rejected": false, "source_profile": { "segment_count": 287, "duration_coverage": 0.894632 } }, "presegmentation_correction": { "applied": true, "correction_count": 1, "deterministic_fix_count": 0 } }