{ "text": "Type Safe 這家公司出了一個模型叫做 Jev，\n它一個字都不會寫。\n你不能叫它寫文章，不能叫它寫程式，\n它連自己為什麼這樣判斷都不會解釋給你聽。\n因為 Jev 把所有力氣都花在一件事情上面：\n你給它一組自己定好的選項，它在裡面選一個，\n然後告訴你它有多確定。\n問一次不到一秒，而且輸出完全免費，\n因為它根本沒有產生文字。\nJev 是什麼、為什麼可以做到這件事、背後是誰做的，\n我在 EP404 已經完整講過一次了。\n如果你還沒聽過，可以直接點過去。\n今天這一集不重新講那些。\n今天要給你的是五個最多人在做、\n而且實際跑起來確實有用的用法，每一個用法該怎麼做，\n我會直接跟你說，\n包括為什麼那個地方用 Jev 會比用大模型好。\n在講用法之前，先花一分鐘把 Jev 的形狀講清楚，\n因為後面五個用法全部都是這個形狀的變化。\n你送過去的是一份資料，加上一個具名的問題清單，\n每一題要標題型、要給它判斷的說明，還有選項。\n題型有三種。\n第一種是是非題，Jev 回你一個零到一之間的校準機率。\n這個數字本身就是它的確定程度，\n不會再另外給你一個信心分數。\n第二種是選擇題，Jev 回你選中的那一個，\n加上每一個選項各自的機率，\n然後另外給你一個 confidence。\n這個分佈等一下在路由那個用法會變成關鍵。\n第三種是評分題，回你一個分數，可以是小數，\n一樣另外給你一個 confidence。\n級距最少兩級，最多十級。\n還有一個數字你先知道比較好。\n有一份獨立評測拿 6000 多筆資料比過，\nJev 跟頂級模型的判斷一致率是 91.5%，\n跟它同價位的那個小模型是 93.5%。\n所以 Jev 強的是快跟便宜，它沒有比較準。\n這件事很重要，因為它直接決定了下一頁那條規則。\n那條規則是這樣：Jev 真正適合的地方，\n是那些會被快速運算很多次、而且不太需要推理的判斷。\n因為它的單次優勢其實沒有那麼明顯，\n一次判斷省你幾秒鐘、省你幾分美元。\n可是當這個判斷要發生幾千次、\n或者卡在流程的每一步前面的時候，\n那個差距就被加成上去了。\n反過來說，如果一件事情只發生幾次，\n那用 Jev 省下來的東西很有限，\n你反而要擔心它答得對不對。\n這條規則還有另外半邊：那個判斷不能需要推理。\nJev 每一次判斷都是當下的直覺，它不記得上一步，\n也不會替你多想一層。\n所以會被快速運算很多次、而且不太需要推理，\n兩個條件都成立的時候，它才是對的選擇。\n接下來五個用法，我每一個都會告訴你它在那裡要跑幾次。\n我們今天就來拆解 Jev 最多人用、最好用的五個用法。\n大量資訊需要分類、自動分流、讓它自己點瀏覽器、\n決策前的把關，還有指令守門員。\n五個我都會給你實際的做法，\n你聽完挑一個今天晚上就可以試。\n第一個用法，也是最多人在做的一個：\n你手上有一大疊東西要分類。\nJev 在這件事上面的優勢最明顯，因為量越大，\n它跟大模型的差距就被加成得越開。\n如果你今天只想試一個，我建議從這一個開始。\n它不需要接進任何系統，\n你手上有一份現成的資料就可以跑，\n而且跑完你馬上看得到結果對不對。\n最典型的例子是問卷的開放題。\n選擇題 5 分鐘就統計完了，那一欄「還有什麼想說的」\n常常躺在那裡 3 個月沒有人看，因為那是文字，\n沒有人有空一筆一筆讀。\n把它丟給 Jev，題型選選擇題，\n選項就是你自己定的那幾類：這是抱怨價格、\n這是功能許願、這是稱讚、這是在問問題。\n2,000 多筆幾十秒就分完。\n說明就寫「這一則回覆主要在講哪一件事」，這樣就夠了。\n真正決定結果好壞的是選項要互相排斥。\n如果你放了兩個意思重疊的類別，它會在兩個中間搖擺，\n你就會看到一整排信心很低的結果。\n另外有一件事很容易漏掉： 如果你希望它能說「都不適合」，\n你要自己把這個選項也放進去。\n你不放，它一定會在你給的選項裡面硬選一個，\n而且信心還可能不低。\n同一個用法還有一個變化型，我覺得比分類更值得學：\ncluster。\ncluster 跟分類的差別是這樣：分類是你先想好類別，\n可是有時候你根本不知道該分成哪幾類。\n這時候做法反過來：不給它類別，\n改成兩兩成對問一題是非題「這 兩件事情是不是同一個主題」，\n然後讓相關的自己聚成一團。\n分完群之後，你再找一個便宜的語言模型，\n幫每一團取一個名字。\nJev 負責判斷關係，語言模型負責命名，各做各最擅長的。\n有人拿這個做法去跑自己專案一整年的 PR，\n想知道團隊的力氣到底花在哪裡。\n100 多個 PR 花了 1.1 分美元；\n2,000 個 PR 花 9 分美元，跑大概 2 分鐘。\n這種問題以前做得到，\n只是做一次的代價高到讓人算一算就放棄了，\n所以大部分團隊從來沒真的做過。\n這個用法每一筆都要跑一次，\n所以它是五個裡面最划算的一個。\n同樣 2,000 筆，大模型要跑幾 10 分鐘、\n花你幾 10 美元，而且你八成不會真的去跑。\n門檻怎麼設？\n這裡的兩種錯代價其實差不多，一筆被歸錯類，\n你之後翻回來重標就好。\n所以門檻可以設得比較高，\n讓它不確定的那一小疊進一個你每天花 10 分鐘看的地方，\n其餘的讓它自己分完。\n有一個實作上的限制要知道：\nJev 一次吃得下 6 萬 4000 個 token。\n幾千筆不能一次送過去，要自己切成好幾批。\n切好平行送出去，總時間會從一分多鐘掉到幾秒，\n而且成本完全一樣。\n第二個用法是路由：一件事情進來，決定它該交給誰處理。\n而且我要先講的那個目的地不是人，是模型。\n這一個適合已經有東西在線上跑的人。\n如果你的流程每天會被觸發幾百次以上，\n這裡省下來的錢通常比前面那個用法還多，\n因為它是每一個請求都要跑一次。\n你手上如果不只一個模型，有便宜的也有貴的，\n那「這一題要用哪一個」本身就是一個判斷。\n大部分人的做法是隨便選一個，\n或者叫一個貴的模型來決定。\n後面這個做法有個很明顯的問題：你光是決定要用哪一個，\n就已經先付了一次貴模型的錢，還要等它想完。\n這個判斷是擋在一堆昂貴的東西前面的，\n它自己不能也很昂貴。\n換成 Jev，題型選選擇題，\n選項就是你手上那幾個模型的定位：\n這一題是簡單的事實查詢、需要多步推理、\n還是需要寫長篇內容。\n說明裡面把每一個模型擅長什麼各寫一句。\n有人把面板做成底下掛兩個計數器：你實際花了多少，\n跟全部都送給貴模型會花多少。\n這一個每一個請求都要跑一次，所以請求量越大，\n那兩個數字差得越開。\n同一個做法換成派給人也成立。\n進來一張工單，內容是付款一直失敗，\nJev 同時算出三個目的地的機率。\n這一題的說明要寫得比前面細一點，\n因為你要教它每個目的地負責什麼：帳務處理付款、發票、\n退費；技術處理打不開、跑不動、出現錯誤訊息；\n業務處理報價跟方案。\n可是換成另外一張工單，帳務 0.42、技術 0.39。\n最高的那個雖然還是帳務，但它跟第二名只差 0.03。\n所以這個用法你要看的東西跟前面都不一樣。\n你要看的是第一名跟第二名差多少，不是最高那個有多高。\n差距很小就代表 Jev 其實分不出來，\n那就讓它停下來問人，\n不要讓它自己猜一個然後整條流程跑錯方向。\n這也是為什麼選擇題會把整個分佈都回給你，\n而不是只回一個答案。\n第三個用法，是讓你的工具自己開瀏覽器去做事。\n現在很多工具會自己上網查資料、自己跑報表、\n自己填表單。\n它們慢的原因，\n是每一步要點哪裡都要先問一次那個會寫文章的大模型。\n這一個的門檻比前面兩個高一點，\n因為你要有一個會自己操作瀏覽器的工具才用得到。\n但如果你已經有了，這是感受差距最明顯的一個，\n因為慢是你自己在旁邊看著的。\n換成 Jev 以後做法不一樣：\n每一步都把這個畫面上可以點的東西，\n重新列成一份純文字的清單。\n然後讓 Jev 從清單裡面選一個，點下去，\n再從新的畫面列一份新的清單，\n一直重複到它判斷目標達成或者被擋住為止。\n需要打字的時候再交給一個小模型去填。\n這裡是每一步都要跑一次。\n一個流程十幾步，每一步差個三秒，\n整條流程就差了快一分鐘。\n而且這個判斷本質上是「從清單裡選一個」，\n根本不需要會寫字的能力。\n有一件事要講清楚： 最後判斷「這個功能到底做對了沒有」，\n還是要留給主要那個模型。\nJev 只負責決定下一步點哪裡。\n這個用法我看過最實用的一個做法，是拿它來測權限。\n分別用管理員、主管、一般員工登入，\n讓它自己把每個身分該進得去、\n不該進得去的頁面都走一遍，\n看有沒有哪個人進得去他不該進去的地方。\n這種測試人工做很煩，所以大部分團隊根本不做。\n門檻在這裡是由動作本身決定的。\n只是讀一下、查一下，門檻可以低一點，\n做錯了再做一次就好。\n會改到資料、會送出去給別人看、會花到錢的，\n門檻就要高一點。\n第四個用法，是東西要送出去給別人看之前，\n先自動看一遍。\n你現在大概已經讓 AI 幫你擬回信、擬報價、擬產品文案、\n擬週報了。\n這些東西你不太敢直接送出去，\n所以你每一份都自己看過一遍。\n這個用法就是幫你評估 AI 生 成的文案是不是可以直接送出。\n這裡不是幫你省錢，但幫你省下了你的注意力，\n從每一份都要確認，變成只看它不確定的那幾份。\n做法是同一份草稿，一次問三題是非題。\n三題可以這樣寫。\n第一題寫「收件人、金額、日期這三個欄位， 內容裡都有而且都不是空的嗎」。\n第二題寫「這段文字的語氣， 適合寄給一個還沒成交的客戶嗎」。\n第三題寫「文章裡提到的數字或說法，有沒有附上出處」。\n這裡是一次問好幾題，而且方向跟前面相反：\n因為 Jev 是拿同一份資料一起判斷完的，\n多問幾題幾乎不會變慢。\n換成大模型，你要嘛問 3 次，\n要嘛叫它寫一段長輸出再自己拆解出來讀。\n三題都過就直接送出去。\n有一題不確定，就需要讓你先審核。\n有明顯不通過的，就退回去重做。\n要注意的是 Jev 只會說不行，不會告訴你哪裡不行。\n所以退回去的時候，你要自己把是哪一題沒過一起帶上，\n不然它不知道要改什麼。\n同一個位子還有一個評分題的用法，工程團隊應該都有感：\nbug 進來的時候先自動分嚴重程度。\n題型選評分題，級距就設三級或五級：這是只影響外觀的、\n是某個功能壞掉的、還是整個卡住不能用的。\n說明裡面把每一級實際長什麼樣子各寫一句。\n有一個原則我建議你守著：這個分數拿來排順序，\n不要拿來自動做處置。\n讓它決定今天早上誰先被看到，\n不要讓它決定哪一張可以直接關掉。\n第五個用法，如果你有在用那種會自己跑指令的工具，\n這一個我認為最值得先做。\n它擋的是那種平常不會發生，\n但是一旦發生就讓你很崩潰的事情。\n前面四個用法都是在幫你省時間跟省錢，這一個是保險。\n它平常毫無存在感，但是如果用到了，\n你會非常感謝自己有做這條保險。\n做法是在那個會跑指令的工具被呼叫之前，掛一道檢查，\n把整條指令送給 Jev，問一題是非題。\n你可能會想，這種事寫一份黑名單不就好了。\n問題是黑名單只擋得到你想得到的那幾條。\n通常真正出事的那次，是你從來沒想過的指令。\n所以那一題不要寫成「是不是這幾條指令」，\n要寫成「這條指令會不會造成不可逆的破壞」，\n然後在說明裡面給它 3 到 5 個例子，\n剩下的讓它自己推斷。\n同一套也可以拿來擋寫入，\n例如不准它去動放金鑰的那個檔案。\n這裡是每一次動作之前都要跑一次。\n這道檢查卡在每一步前面，所以它必須快到你感覺不到，\n不然你會自己把它拿掉。\n門檻在這裡要拉到很高，而且只要 Jev 不確定，\n就一律讓它停下來問人。\n理由很簡單：這裡的兩種錯代價差太多了。\n擋錯了，你多按一次確認；放錯了，是資料回不來。\n所以這裡寧可多擋，不要少擋。\n不過有一件事你要先知道：這個用法我看到有作者分享過，\n他實際跑下來有幾次成功、也有幾次失敗。\n所以不要把它當成最後一道防線，它是一道很便宜、\n可以一直開著的篩子，\n你原本該有的備份跟權限設定還是要有。\n前面每一個用法我都給了一個門檻的數字，\n但那些都是別人的參考值。\n確切來說，你自己的 threshold 要怎麼得到？\n我們這邊來聊一下。\n做法很簡單。\n你自己先人工判斷 100 到 200 筆，每一類都要有幾十筆。\n然後在幾個不同的門檻底下，分別算算看兩種錯各有幾筆：\n一種是該放過卻被擋下來，一種是該擋卻被放過去。\n這兩個數字會往相反的方向跑。\n門檻拉高，被擋下來的變多，你自己要看的東西就變多；\n門檻拉低，被放過去的錯就變多。\n選一個你可以接受的，那就是你的門檻。\n而你之後遇到第六個、第七個想法的時候，\n也可以自己判斷值不值得換成 Jev：\n就看那個判斷會不會被快速運算很多次，\n以及它需不需要推理。\n兩個都成立才是它的主場。\n那實際上要怎麼接起來？\n看你手上有沒有在寫程式，兩條路。\n如果你有在用那種可以自己跑程式的 AI，\n官方有出一個現成的 skill，兩行指令就裝好了。\n第一行是 claude plugin marketplace add，\n後面接 typesafe 槓 ai 斜線 skills；\n第二行是 claude plugin install，\n後面接 typesafe at typesafe 槓 ai。\n這兩行你搜 typesafe agent skill 就找得到，\n官方文件上有完整的。\n裝完把你的 key 放進專案的 `.env` 檔案裡面。\n有一件事很多人會誤會：\n這個 skill 本身不執行任何東西，它是一本規則書。\n你用白話描述你要判斷什麼，你的 AI 讀那本規則書，\n幫你把題目跟判斷標準寫出來，\n再寫一小段程式把資料送過去，最後把結果讀回來給你。\n如果你完全不想碰程式碼，\n市面上的自動化平台已經可以把 Jev 當成流程裡的一個步驟，\n前面接你的信箱、表單或者試算表，後面接它的判斷結果。\n今天晚上你可以先做三件事情。\n第一，從第一個用法開始，因為它最划算也最不會出事。\n找一疊你早就知道該看、可是一直沒時間看的東西，\n可能是問卷、可能是客訴、可能是你專案裡一整年的紀錄。\n第二，把你想分的類別寫成互相排斥的選項，\n然後自己加一個「都不適合」進去。\n如果你連該分成哪幾類都還不確定，\n那就用 cluster 那個做法，兩兩問它相不相關。\n第三，第 1 個禮拜先不要讓它真的動手。\n先讓它判斷，結果印在旁邊，你自己對一遍。\n對完那一輪，你就順便把門檻量出來了。\n今天講的五個用法，我幫你整理一次，\n順便把每一個要跑幾次再講一遍。\n大量資訊需要分類，每一筆都要跑一次，量越大越划算，\n而且不知道該分幾類的時候可以改成兩兩問相關。\n自動分流，每一個請求都要跑一次，\n而且你要看的是第一名跟第二名的差距。\n讓它自己點瀏覽器，每一步都要跑一次，\n最後那個做對了沒有還是留給大模型。\n決策前的把關，一次問三題幾乎不變慢。\n指令守門員，每一次動作之前都要跑一次，\n所以門檻要拉到很高。\n所以整支片你如果只帶走一句話，我希望是這一句：\n一個判斷值不值得換成 Jev，就看兩件事，\n它會不會被快速運算很多次，以及它需不需要推理。\n這五個裡面如果有一個你等一下就想試，幫我按個訂閱、\n開個小鈴鐺。\n所有平台都叫 AI 懶人報，習慣用聽的就搜 Podcast。\n我們下一集見。", "language": "zh", "duration": 886.0, "segments": [ { "id": 0, "start": 0.0, "end": 3.079, "text": "Type Safe 這家公司出了一個模型叫做 Jev，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 1, "start": 3.079, "end": 4.68, "text": "它一個字都不會寫。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 2, "start": 4.68, "end": 6.949, "text": "你不能叫它寫文章，不能叫它寫程式，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 3, "start": 6.949, "end": 10.28, "text": "它連自己為什麼這樣判斷都不會解釋給你聽。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 4, "start": 10.28, "end": 13.446, "text": "因為 Jev 把所有力氣都花在一件事情上面：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 5, "start": 13.446, "end": 16.467, "text": "你給它一組自己定好的選項，它在裡面選一個，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 6, "start": 16.467, "end": 18.2, "text": "然後告訴你它有多確定。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 7, "start": 18.2, "end": 20.164, "text": "問一次不到一秒，而且輸出完全免費，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 8, "start": 20.164, "end": 22.08, "text": "因為它根本沒有產生文字。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 9, "start": 22.08, "end": 24.989, "text": "Jev 是什麼、為什麼可以做到這件事、背後是誰做的，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 10, "start": 24.989, "end": 27.24, "text": "我在 EP404 已經完整講過一次了。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 11, "start": 27.24, "end": 28.5, "text": "如果你還沒聽過，可以直接點過去。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 12, "start": 28.5, "end": 30.24, "text": "今天這一集不重新講那些。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 13, "start": 30.24, "end": 32.148, "text": "今天要給你的是五個最多人在做、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 14, "start": 32.148, "end": 35.328, "text": "而且實際跑起來確實有用的用法，每一個用法該怎麼做，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 15, "start": 35.328, "end": 36.346, "text": "我會直接跟你說，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 16, "start": 36.346, "end": 39.48, "text": "包括為什麼那個地方用 Jev 會比用大模型好。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 17, "start": 39.48, "end": 42.922, "text": "在講用法之前，先花一分鐘把 Jev 的形狀講清楚，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 18, "start": 42.922, "end": 45.82, "text": "因為後面五個用法全部都是這個形狀的變化。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 19, "start": 45.82, "end": 49.038, "text": "你送過去的是一份資料，加上一個具名的問題清單，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 20, "start": 49.038, "end": 52.54, "text": "每一題要標題型、要給它判斷的說明，還有選項。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 21, "start": 52.54, "end": 53.52, "text": "題型有三種。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 22, "start": 53.52, "end": 57.6, "text": "第一種是是非題，Jev 回你一個零到一之間的校準機率。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 23, "start": 57.6, "end": 59.838, "text": "這個數字本身就是它的確定程度，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 24, "start": 59.838, "end": 62.28, "text": "不會再另外給你一個信心分數。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 25, "start": 62.28, "end": 65.201, "text": "第二種是選擇題，Jev 回你選中的那一個，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 26, "start": 65.201, "end": 67.01, "text": "加上每一個選項各自的機率，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 27, "start": 67.01, "end": 70.1, "text": "然後另外給你一個 confidence。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 28, "start": 70.1, "end": 73.46, "text": "這個分佈等一下在路由那個用法會變成關鍵。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 29, "start": 73.46, "end": 75.858, "text": "第三種是評分題，回你一個分數，可以是小數，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 30, "start": 75.858, "end": 78.64, "text": "一樣另外給你一個 confidence。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 31, "start": 78.64, "end": 80.82, "text": "級距最少兩級，最多十級。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 32, "start": 80.82, "end": 82.82, "text": "還有一個數字你先知道比較好。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 33, "start": 82.82, "end": 85.784, "text": "有一份獨立評測拿 6000 多筆資料比過，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 34, "start": 85.784, "end": 89.03, "text": "Jev 跟頂級模型的判斷一致率是 91.5%，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 35, "start": 89.03, "end": 91.964, "text": "跟它同價位的那個小模型是 93.5%。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 36, "start": 91.964, "end": 95.76, "text": "所以 Jev 強的是快跟便宜，它沒有比較準。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 37, "start": 95.76, "end": 98.84, "text": "這件事很重要，因為它直接決定了下一頁那條規則。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 38, "start": 98.84, "end": 101.9, "text": "那條規則是這樣：Jev 真正適合的地方，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 39, "start": 101.9, "end": 105.92, "text": "是那些會被快速運算很多次、而且不太需要推理的判斷。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 40, "start": 105.92, "end": 108.489, "text": "因為它的單次優勢其實沒有那麼明顯，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 41, "start": 108.489, "end": 111.24, "text": "一次判斷省你幾秒鐘、省你幾分美元。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 42, "start": 111.24, "end": 113.455, "text": "可是當這個判斷要發生幾千次、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 43, "start": 113.455, "end": 115.985, "text": "或者卡在流程的每一步前面的時候，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 44, "start": 115.985, "end": 118.24, "text": "那個差距就被加成上去了。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 45, "start": 118.24, "end": 120.511, "text": "反過來說，如果一件事情只發生幾次，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 46, "start": 120.511, "end": 122.783, "text": "那用 Jev 省下來的東西很有限，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 47, "start": 122.783, "end": 124.92, "text": "你反而要擔心它答得對不對。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 48, "start": 124.92, "end": 128.06, "text": "這條規則還有另外半邊：那個判斷不能需要推理。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 49, "start": 128.06, "end": 131.579, "text": "Jev 每一次判斷都是當下的直覺，它不記得上一步，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 50, "start": 131.579, "end": 133.2, "text": "也不會替你多想一層。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 51, "start": 133.2, "end": 136.333, "text": "所以會被快速運算很多次、而且不太需要推理，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 52, "start": 136.333, "end": 139.66, "text": "兩個條件都成立的時候，它才是對的選擇。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 53, "start": 139.66, "end": 143.62, "text": "接下來五個用法，我每一個都會告訴你它在那裡要跑幾次。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 54, "start": 143.62, "end": 147.92, "text": "我們今天就來拆解 Jev 最多人用、最好用的五個用法。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 55, "start": 147.92, "end": 153.046, "text": "大量資訊需要分類、自動分流、讓它自己點瀏覽器、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 56, "start": 153.046, "end": 156.98, "text": "決策前的把關，還有指令守門員。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 57, "start": 156.98, "end": 158.882, "text": "五個我都會給你實際的做法，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 58, "start": 158.882, "end": 161.24, "text": "你聽完挑一個今天晚上就可以試。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 59, "start": 161.24, "end": 163.278, "text": "第一個用法，也是最多人在做的一個：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 60, "start": 163.278, "end": 165.46, "text": "你手上有一大疊東西要分類。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 61, "start": 165.46, "end": 169.112, "text": "Jev 在這件事上面的優勢最明顯，因為量越大，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 62, "start": 169.112, "end": 172.22, "text": "它跟大模型的差距就被加成得越開。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 63, "start": 172.22, "end": 175.16, "text": "如果你今天只想試一個，我建議從這一個開始。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 64, "start": 175.16, "end": 176.719, "text": "它不需要接進任何系統，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 65, "start": 176.719, "end": 178.986, "text": "你手上有一份現成的資料就可以跑，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 66, "start": 178.986, "end": 181.62, "text": "而且跑完你馬上看得到結果對不對。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 67, "start": 181.62, "end": 183.94, "text": "最典型的例子是問卷的開放題。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 68, "start": 183.94, "end": 187.018, "text": "選擇題 5 分鐘就統計完了，那一欄「還有什麼想說的」", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 69, "start": 187.018, "end": 189.74, "text": "常常躺在那裡 3 個月沒有人看，因為那是文字，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 70, "start": 189.74, "end": 191.44, "text": "沒有人有空一筆一筆讀。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 71, "start": 191.44, "end": 194.389, "text": "把它丟給 Jev，題型選選擇題，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 72, "start": 194.389, "end": 198.075, "text": "選項就是你自己定的那幾類：這是抱怨價格、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 73, "start": 198.075, "end": 201.78, "text": "這是功能許願、這是稱讚、這是在問問題。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 74, "start": 201.78, "end": 204.02, "text": "2,000 多筆幾十秒就分完。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 75, "start": 204.02, "end": 208.3, "text": "說明就寫「這一則回覆主要在講哪一件事」，這樣就夠了。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 76, "start": 208.3, "end": 212.38, "text": "真正決定結果好壞的是選項要互相排斥。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 77, "start": 212.38, "end": 216.711, "text": "如果你放了兩個意思重疊的類別，它會在兩個中間搖擺，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 78, "start": 216.711, "end": 219.7, "text": "你就會看到一整排信心很低的結果。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 79, "start": 219.7, "end": 223.471, "text": "另外有一件事很容易漏掉： 如果你希望它能說「都不適合」，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 80, "start": 223.471, "end": 225.56, "text": "你要自己把這個選項也放進去。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 81, "start": 225.56, "end": 228.217, "text": "你不放，它一定會在你給的選項裡面硬選一個，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 82, "start": 228.217, "end": 229.64, "text": "而且信心還可能不低。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 83, "start": 229.64, "end": 232.716, "text": "同一個用法還有一個變化型，我覺得比分類更值得學：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 84, "start": 232.716, "end": 234.38, "text": "cluster。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 85, "start": 234.38, "end": 237.414, "text": "cluster 跟分類的差別是這樣：分類是你先想好類別，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 86, "start": 237.414, "end": 239.48, "text": "可是有時候你根本不知道該分成哪幾類。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 87, "start": 239.48, "end": 242.048, "text": "這時候做法反過來：不給它類別，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 88, "start": 242.048, "end": 246.841, "text": "改成兩兩成對問一題是非題「這 兩件事情是不是同一個主題」，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 89, "start": 246.841, "end": 249.52, "text": "然後讓相關的自己聚成一團。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 90, "start": 249.52, "end": 252.097, "text": "分完群之後，你再找一個便宜的語言模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 91, "start": 252.097, "end": 253.62, "text": "幫每一團取一個名字。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 92, "start": 253.62, "end": 258.56, "text": "Jev 負責判斷關係，語言模型負責命名，各做各最擅長的。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 93, "start": 258.56, "end": 262.078, "text": "有人拿這個做法去跑自己專案一整年的 PR，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 94, "start": 262.078, "end": 265.12, "text": "想知道團隊的力氣到底花在哪裡。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 95, "start": 265.12, "end": 267.03, "text": "100 多個 PR 花了 1.1 分美元；", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 96, "start": 267.03, "end": 269.72, "text": "2,000 個 PR 花 9 分美元，跑大概 2 分鐘。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 97, "start": 269.72, "end": 271.128, "text": "這種問題以前做得到，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 98, "start": 271.128, "end": 273.944, "text": "只是做一次的代價高到讓人算一算就放棄了，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 99, "start": 273.944, "end": 276.22, "text": "所以大部分團隊從來沒真的做過。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 100, "start": 276.22, "end": 278.268, "text": "這個用法每一筆都要跑一次，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 101, "start": 278.268, "end": 280.82, "text": "所以它是五個裡面最划算的一個。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 102, "start": 280.82, "end": 283.278, "text": "同樣 2,000 筆，大模型要跑幾 10 分鐘、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 103, "start": 283.278, "end": 286.08, "text": "花你幾 10 美元，而且你八成不會真的去跑。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 104, "start": 286.08, "end": 287.0, "text": "門檻怎麼設？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 105, "start": 287.0, "end": 290.592, "text": "這裡的兩種錯代價其實差不多，一筆被歸錯類，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 106, "start": 290.592, "end": 292.68, "text": "你之後翻回來重標就好。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 107, "start": 292.68, "end": 294.762, "text": "所以門檻可以設得比較高，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 108, "start": 294.762, "end": 299.619, "text": "讓它不確定的那一小疊進一個你每天花 10 分鐘看的地方，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 109, "start": 299.619, "end": 301.74, "text": "其餘的讓它自己分完。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 110, "start": 301.74, "end": 302.874, "text": "有一個實作上的限制要知道：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 111, "start": 302.874, "end": 306.562, "text": "Jev 一次吃得下 6 萬 4000 個 token。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 112, "start": 306.562, "end": 309.604, "text": "幾千筆不能一次送過去，要自己切成好幾批。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 113, "start": 309.604, "end": 312.914, "text": "切好平行送出去，總時間會從一分多鐘掉到幾秒，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 114, "start": 312.914, "end": 314.78, "text": "而且成本完全一樣。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 115, "start": 314.78, "end": 319.04, "text": "第二個用法是路由：一件事情進來，決定它該交給誰處理。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 116, "start": 319.04, "end": 322.4, "text": "而且我要先講的那個目的地不是人，是模型。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 117, "start": 322.4, "end": 324.66, "text": "這一個適合已經有東西在線上跑的人。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 118, "start": 324.66, "end": 327.354, "text": "如果你的流程每天會被觸發幾百次以上，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 119, "start": 327.354, "end": 330.198, "text": "這裡省下來的錢通常比前面那個用法還多，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 120, "start": 330.198, "end": 332.78, "text": "因為它是每一個請求都要跑一次。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 121, "start": 332.78, "end": 335.655, "text": "你手上如果不只一個模型，有便宜的也有貴的，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 122, "start": 335.655, "end": 338.82, "text": "那「這一題要用哪一個」本身就是一個判斷。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 123, "start": 338.82, "end": 340.817, "text": "大部分人的做法是隨便選一個，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 124, "start": 340.817, "end": 342.84, "text": "或者叫一個貴的模型來決定。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 125, "start": 342.84, "end": 346.699, "text": "後面這個做法有個很明顯的問題：你光是決定要用哪一個，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 126, "start": 346.699, "end": 350.04, "text": "就已經先付了一次貴模型的錢，還要等它想完。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 127, "start": 350.04, "end": 352.211, "text": "這個判斷是擋在一堆昂貴的東西前面的，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 128, "start": 352.211, "end": 353.78, "text": "它自己不能也很昂貴。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 129, "start": 353.78, "end": 356.297, "text": "換成 Jev，題型選選擇題，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 130, "start": 356.297, "end": 359.174, "text": "選項就是你手上那幾個模型的定位：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 131, "start": 359.174, "end": 362.591, "text": "這一題是簡單的事實查詢、需要多步推理、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 132, "start": 362.591, "end": 364.62, "text": "還是需要寫長篇內容。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 133, "start": 364.62, "end": 368.16, "text": "說明裡面把每一個模型擅長什麼各寫一句。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 134, "start": 368.16, "end": 372.037, "text": "有人把面板做成底下掛兩個計數器：你實際花了多少，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 135, "start": 372.037, "end": 374.46, "text": "跟全部都送給貴模型會花多少。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 136, "start": 374.46, "end": 378.347, "text": "這一個每一個請求都要跑一次，所以請求量越大，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 137, "start": 378.347, "end": 380.5, "text": "那兩個數字差得越開。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 138, "start": 380.5, "end": 382.8, "text": "同一個做法換成派給人也成立。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 139, "start": 382.8, "end": 385.312, "text": "進來一張工單，內容是付款一直失敗，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 140, "start": 385.312, "end": 387.98, "text": "Jev 同時算出三個目的地的機率。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 141, "start": 387.98, "end": 390.627, "text": "這一題的說明要寫得比前面細一點，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 142, "start": 390.627, "end": 394.928, "text": "因為你要教它每個目的地負責什麼：帳務處理付款、發票、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 143, "start": 394.928, "end": 398.568, "text": "退費；技術處理打不開、跑不動、出現錯誤訊息；", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 144, "start": 398.568, "end": 400.7, "text": "業務處理報價跟方案。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 145, "start": 400.7, "end": 403.696, "text": "可是換成另外一張工單，帳務 0.42、技術 0.39。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 146, "start": 403.696, "end": 408.442, "text": "最高的那個雖然還是帳務，但它跟第二名只差 0.03。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 147, "start": 408.442, "end": 411.46, "text": "所以這個用法你要看的東西跟前面都不一樣。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 148, "start": 411.46, "end": 414.92, "text": "你要看的是第一名跟第二名差多少，不是最高那個有多高。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 149, "start": 414.92, "end": 417.602, "text": "差距很小就代表 Jev 其實分不出來，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 150, "start": 417.602, "end": 419.014, "text": "那就讓它停下來問人，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 151, "start": 419.014, "end": 422.02, "text": "不要讓它自己猜一個然後整條流程跑錯方向。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 152, "start": 422.02, "end": 424.715, "text": "這也是為什麼選擇題會把整個分佈都回給你，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 153, "start": 424.715, "end": 426.58, "text": "而不是只回一個答案。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 154, "start": 426.58, "end": 430.24, "text": "第三個用法，是讓你的工具自己開瀏覽器去做事。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 155, "start": 430.24, "end": 433.247, "text": "現在很多工具會自己上網查資料、自己跑報表、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 156, "start": 433.247, "end": 434.48, "text": "自己填表單。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 157, "start": 434.48, "end": 435.687, "text": "它們慢的原因，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 158, "start": 435.687, "end": 440.22, "text": "是每一步要點哪裡都要先問一次那個會寫文章的大模型。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 159, "start": 440.22, "end": 442.424, "text": "這一個的門檻比前面兩個高一點，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 160, "start": 442.424, "end": 445.98, "text": "因為你要有一個會自己操作瀏覽器的工具才用得到。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 161, "start": 445.98, "end": 449.099, "text": "但如果你已經有了，這是感受差距最明顯的一個，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 162, "start": 449.099, "end": 451.26, "text": "因為慢是你自己在旁邊看著的。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 163, "start": 451.26, "end": 453.538, "text": "換成 Jev 以後做法不一樣：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 164, "start": 453.538, "end": 456.119, "text": "每一步都把這個畫面上可以點的東西，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 165, "start": 456.119, "end": 458.9, "text": "重新列成一份純文字的清單。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 166, "start": 458.9, "end": 462.119, "text": "然後讓 Jev 從清單裡面選一個，點下去，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 167, "start": 462.119, "end": 464.265, "text": "再從新的畫面列一份新的清單，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 168, "start": 464.265, "end": 467.66, "text": "一直重複到它判斷目標達成或者被擋住為止。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 169, "start": 467.66, "end": 470.64, "text": "需要打字的時候再交給一個小模型去填。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 170, "start": 470.64, "end": 472.18, "text": "這裡是每一步都要跑一次。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 171, "start": 472.18, "end": 474.381, "text": "一個流程十幾步，每一步差個三秒，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 172, "start": 474.381, "end": 476.18, "text": "整條流程就差了快一分鐘。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 173, "start": 476.18, "end": 478.946, "text": "而且這個判斷本質上是「從清單裡選一個」，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 174, "start": 478.946, "end": 480.76, "text": "根本不需要會寫字的能力。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 175, "start": 480.76, "end": 484.339, "text": "有一件事要講清楚： 最後判斷「這個功能到底做對了沒有」，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 176, "start": 484.339, "end": 486.1, "text": "還是要留給主要那個模型。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 177, "start": 486.1, "end": 489.24, "text": "Jev 只負責決定下一步點哪裡。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 178, "start": 489.24, "end": 492.94, "text": "這個用法我看過最實用的一個做法，是拿它來測權限。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 179, "start": 492.94, "end": 495.263, "text": "分別用管理員、主管、一般員工登入，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 180, "start": 495.263, "end": 497.177, "text": "讓它自己把每個身分該進得去、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 181, "start": 497.177, "end": 498.953, "text": "不該進得去的頁面都走一遍，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 182, "start": 498.953, "end": 502.16, "text": "看有沒有哪個人進得去他不該進去的地方。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 183, "start": 502.16, "end": 506.16, "text": "這種測試人工做很煩，所以大部分團隊根本不做。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 184, "start": 506.16, "end": 508.48, "text": "門檻在這裡是由動作本身決定的。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 185, "start": 508.48, "end": 511.031, "text": "只是讀一下、查一下，門檻可以低一點，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 186, "start": 511.031, "end": 512.62, "text": "做錯了再做一次就好。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 187, "start": 512.62, "end": 515.699, "text": "會改到資料、會送出去給別人看、會花到錢的，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 188, "start": 515.699, "end": 517.38, "text": "門檻就要高一點。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 189, "start": 517.38, "end": 520.978, "text": "第四個用法，是東西要送出去給別人看之前，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 190, "start": 520.978, "end": 522.5, "text": "先自動看一遍。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 191, "start": 522.5, "end": 526.636, "text": "你現在大概已經讓 AI 幫你擬回信、擬報價、擬產品文案、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 192, "start": 526.636, "end": 527.64, "text": "擬週報了。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 193, "start": 527.64, "end": 529.295, "text": "這些東西你不太敢直接送出去，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 194, "start": 529.295, "end": 531.08, "text": "所以你每一份都自己看過一遍。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 195, "start": 531.08, "end": 535.36, "text": "這個用法就是幫你評估 AI 生 成的文案是不是可以直接送出。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 196, "start": 535.36, "end": 538.567, "text": "這裡不是幫你省錢，但幫你省下了你的注意力，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 197, "start": 538.567, "end": 542.14, "text": "從每一份都要確認，變成只看它不確定的那幾份。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 198, "start": 542.14, "end": 545.52, "text": "做法是同一份草稿，一次問三題是非題。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 199, "start": 545.52, "end": 546.58, "text": "三題可以這樣寫。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 200, "start": 546.58, "end": 552.66, "text": "第一題寫「收件人、金額、日期這三個欄位， 內容裡都有而且都不是空的嗎」。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 201, "start": 552.66, "end": 557.34, "text": "第二題寫「這段文字的語氣， 適合寄給一個還沒成交的客戶嗎」。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 202, "start": 557.34, "end": 561.98, "text": "第三題寫「文章裡提到的數字或說法，有沒有附上出處」。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 203, "start": 561.98, "end": 564.656, "text": "這裡是一次問好幾題，而且方向跟前面相反：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 204, "start": 564.656, "end": 567.466, "text": "因為 Jev 是拿同一份資料一起判斷完的，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 205, "start": 567.466, "end": 569.08, "text": "多問幾題幾乎不會變慢。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 206, "start": 569.08, "end": 570.99, "text": "換成大模型，你要嘛問 3 次，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 207, "start": 570.99, "end": 573.58, "text": "要嘛叫它寫一段長輸出再自己拆解出來讀。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 208, "start": 573.58, "end": 575.68, "text": "三題都過就直接送出去。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 209, "start": 575.68, "end": 578.24, "text": "有一題不確定，就需要讓你先審核。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 210, "start": 578.24, "end": 581.1, "text": "有明顯不通過的，就退回去重做。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 211, "start": 581.1, "end": 583.84, "text": "要注意的是 Jev 只會說不行，不會告訴你哪裡不行。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 212, "start": 583.84, "end": 586.917, "text": "所以退回去的時候，你要自己把是哪一題沒過一起帶上，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 213, "start": 586.917, "end": 588.62, "text": "不然它不知道要改什麼。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 214, "start": 588.62, "end": 592.569, "text": "同一個位子還有一個評分題的用法，工程團隊應該都有感：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 215, "start": 592.569, "end": 595.56, "text": "bug 進來的時候先自動分嚴重程度。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 216, "start": 595.56, "end": 599.134, "text": "題型選評分題，級距就設三級或五級：這是只影響外觀的、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 217, "start": 599.134, "end": 602.04, "text": "是某個功能壞掉的、還是整個卡住不能用的。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 218, "start": 602.04, "end": 604.72, "text": "說明裡面把每一級實際長什麼樣子各寫一句。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 219, "start": 604.72, "end": 607.91, "text": "有一個原則我建議你守著：這個分數拿來排順序，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 220, "start": 607.91, "end": 609.52, "text": "不要拿來自動做處置。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 221, "start": 609.52, "end": 611.748, "text": "讓它決定今天早上誰先被看到，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 222, "start": 611.748, "end": 614.46, "text": "不要讓它決定哪一張可以直接關掉。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 223, "start": 614.46, "end": 617.692, "text": "第五個用法，如果你有在用那種會自己跑指令的工具，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 224, "start": 617.692, "end": 619.5, "text": "這一個我認為最值得先做。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 225, "start": 619.5, "end": 621.297, "text": "它擋的是那種平常不會發生，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 226, "start": 621.297, "end": 623.66, "text": "但是一旦發生就讓你很崩潰的事情。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 227, "start": 623.66, "end": 627.38, "text": "前面四個用法都是在幫你省時間跟省錢，這一個是保險。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 228, "start": 627.38, "end": 629.801, "text": "它平常毫無存在感，但是如果用到了，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 229, "start": 629.801, "end": 632.32, "text": "你會非常感謝自己有做這條保險。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 230, "start": 632.32, "end": 636.447, "text": "做法是在那個會跑指令的工具被呼叫之前，掛一道檢查，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 231, "start": 636.447, "end": 639.84, "text": "把整條指令送給 Jev，問一題是非題。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 232, "start": 639.84, "end": 642.72, "text": "你可能會想，這種事寫一份黑名單不就好了。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 233, "start": 642.72, "end": 645.84, "text": "問題是黑名單只擋得到你想得到的那幾條。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 234, "start": 645.84, "end": 650.28, "text": "通常真正出事的那次，是你從來沒想過的指令。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 235, "start": 650.28, "end": 653.086, "text": "所以那一題不要寫成「是不是這幾條指令」，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 236, "start": 653.086, "end": 656.032, "text": "要寫成「這條指令會不會造成不可逆的破壞」，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 237, "start": 656.032, "end": 658.837, "text": "然後在說明裡面給它 3 到 5 個例子，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 238, "start": 658.837, "end": 660.44, "text": "剩下的讓它自己推斷。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 239, "start": 660.44, "end": 662.203, "text": "同一套也可以拿來擋寫入，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 240, "start": 662.203, "end": 664.76, "text": "例如不准它去動放金鑰的那個檔案。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 241, "start": 664.76, "end": 667.62, "text": "這裡是每一次動作之前都要跑一次。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 242, "start": 667.62, "end": 671.378, "text": "這道檢查卡在每一步前面，所以它必須快到你感覺不到，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 243, "start": 671.378, "end": 673.2, "text": "不然你會自己把它拿掉。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 244, "start": 673.2, "end": 677.088, "text": "門檻在這裡要拉到很高，而且只要 Jev 不確定，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 245, "start": 677.088, "end": 679.1, "text": "就一律讓它停下來問人。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 246, "start": 679.1, "end": 682.5, "text": "理由很簡單：這裡的兩種錯代價差太多了。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 247, "start": 682.5, "end": 685.6, "text": "擋錯了，你多按一次確認；放錯了，是資料回不來。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 248, "start": 685.6, "end": 687.92, "text": "所以這裡寧可多擋，不要少擋。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 249, "start": 687.92, "end": 691.468, "text": "不過有一件事你要先知道：這個用法我看到有作者分享過，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 250, "start": 691.468, "end": 694.24, "text": "他實際跑下來有幾次成功、也有幾次失敗。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 251, "start": 694.24, "end": 697.864, "text": "所以不要把它當成最後一道防線，它是一道很便宜、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 252, "start": 697.864, "end": 699.439, "text": "可以一直開著的篩子，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 253, "start": 699.439, "end": 702.72, "text": "你原本該有的備份跟權限設定還是要有。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 254, "start": 702.72, "end": 705.676, "text": "前面每一個用法我都給了一個門檻的數字，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 255, "start": 705.676, "end": 707.7, "text": "但那些都是別人的參考值。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 256, "start": 707.7, "end": 710.14, "text": "確切來說，你自己的 threshold 要怎麼得到？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 257, "start": 710.14, "end": 711.4, "text": "我們這邊來聊一下。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 258, "start": 711.4, "end": 712.48, "text": "做法很簡單。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 259, "start": 712.48, "end": 717.08, "text": "你自己先人工判斷 100 到 200 筆，每一類都要有幾十筆。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 260, "start": 717.08, "end": 721.258, "text": "然後在幾個不同的門檻底下，分別算算看兩種錯各有幾筆：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 261, "start": 721.258, "end": 725.4, "text": "一種是該放過卻被擋下來，一種是該擋卻被放過去。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 262, "start": 725.4, "end": 727.74, "text": "這兩個數字會往相反的方向跑。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 263, "start": 727.74, "end": 731.595, "text": "門檻拉高，被擋下來的變多，你自己要看的東西就變多；", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 264, "start": 731.595, "end": 734.14, "text": "門檻拉低，被放過去的錯就變多。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 265, "start": 734.14, "end": 736.92, "text": "選一個你可以接受的，那就是你的門檻。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 266, "start": 736.92, "end": 739.558, "text": "而你之後遇到第六個、第七個想法的時候，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 267, "start": 739.558, "end": 742.057, "text": "也可以自己判斷值不值得換成 Jev：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 268, "start": 742.057, "end": 744.556, "text": "就看那個判斷會不會被快速運算很多次，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 269, "start": 744.556, "end": 746.08, "text": "以及它需不需要推理。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 270, "start": 746.08, "end": 748.32, "text": "兩個都成立才是它的主場。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 271, "start": 748.32, "end": 749.7, "text": "那實際上要怎麼接起來？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 272, "start": 749.7, "end": 752.12, "text": "看你手上有沒有在寫程式，兩條路。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 273, "start": 752.12, "end": 754.736, "text": "如果你有在用那種可以自己跑程式的 AI，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 274, "start": 754.736, "end": 758.2, "text": "官方有出一個現成的 skill，兩行指令就裝好了。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 275, "start": 758.2, "end": 761.209, "text": "第一行是 claude plugin marketplace add，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 276, "start": 761.209, "end": 763.616, "text": "後面接 typesafe 槓 ai 斜線 skills；", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 277, "start": 763.616, "end": 765.937, "text": "第二行是 claude plugin install，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 278, "start": 765.937, "end": 768.618, "text": "後面接 typesafe at typesafe 槓 ai。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 279, "start": 768.618, "end": 773.403, "text": "這兩行你搜 typesafe agent skill 就找得到，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 280, "start": 773.403, "end": 775.085, "text": "官方文件上有完整的。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 281, "start": 775.085, "end": 778.42, "text": "裝完把你的 key 放進專案的 `.env` 檔案裡面。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 282, "start": 778.42, "end": 779.832, "text": "有一件事很多人會誤會：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 283, "start": 779.832, "end": 783.5, "text": "這個 skill 本身不執行任何東西，它是一本規則書。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 284, "start": 783.5, "end": 787.594, "text": "你用白話描述你要判斷什麼，你的 AI 讀那本規則書，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 285, "start": 787.594, "end": 789.798, "text": "幫你把題目跟判斷標準寫出來，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 286, "start": 789.798, "end": 793.96, "text": "再寫一小段程式把資料送過去，最後把結果讀回來給你。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 287, "start": 793.96, "end": 795.596, "text": "如果你完全不想碰程式碼，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 288, "start": 795.596, "end": 799.686, "text": "市面上的自動化平台已經可以把 Jev 當成流程裡的一個步驟，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 289, "start": 799.686, "end": 803.4, "text": "前面接你的信箱、表單或者試算表，後面接它的判斷結果。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 290, "start": 803.4, "end": 805.72, "text": "今天晚上你可以先做三件事情。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 291, "start": 805.72, "end": 809.08, "text": "第一，從第一個用法開始，因為它最划算也最不會出事。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 292, "start": 809.08, "end": 812.47, "text": "找一疊你早就知道該看、可是一直沒時間看的東西，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 293, "start": 812.47, "end": 817.0, "text": "可能是問卷、可能是客訴、可能是你專案裡一整年的紀錄。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 294, "start": 817.0, "end": 820.009, "text": "第二，把你想分的類別寫成互相排斥的選項，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 295, "start": 820.009, "end": 822.58, "text": "然後自己加一個「都不適合」進去。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 296, "start": 822.58, "end": 824.387, "text": "如果你連該分成哪幾類都還不確定，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 297, "start": 824.387, "end": 827.92, "text": "那就用 cluster 那個做法，兩兩問它相不相關。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 298, "start": 827.92, "end": 830.22, "text": "第三，第 1 個禮拜先不要讓它真的動手。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 299, "start": 830.22, "end": 832.96, "text": "先讓它判斷，結果印在旁邊，你自己對一遍。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 300, "start": 832.96, "end": 836.02, "text": "對完那一輪，你就順便把門檻量出來了。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 301, "start": 836.02, "end": 838.515, "text": "今天講的五個用法，我幫你整理一次，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 302, "start": 838.515, "end": 840.9, "text": "順便把每一個要跑幾次再講一遍。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 303, "start": 840.9, "end": 844.246, "text": "大量資訊需要分類，每一筆都要跑一次，量越大越划算，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 304, "start": 844.246, "end": 847.5, "text": "而且不知道該分幾類的時候可以改成兩兩問相關。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 305, "start": 847.5, "end": 849.473, "text": "自動分流，每一個請求都要跑一次，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 306, "start": 849.473, "end": 852.06, "text": "而且你要看的是第一名跟第二名的差距。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 307, "start": 852.06, "end": 854.833, "text": "讓它自己點瀏覽器，每一步都要跑一次，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 308, "start": 854.833, "end": 857.9, "text": "最後那個做對了沒有還是留給大模型。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 309, "start": 857.9, "end": 860.82, "text": "決策前的把關，一次問三題幾乎不變慢。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 310, "start": 860.82, "end": 863.985, "text": "指令守門員，每一次動作之前都要跑一次，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 311, "start": 863.985, "end": 866.28, "text": "所以門檻要拉到很高。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 312, "start": 866.28, "end": 869.363, "text": "所以整支片你如果只帶走一句話，我希望是這一句：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 313, "start": 869.363, "end": 872.177, "text": "一個判斷值不值得換成 Jev，就看兩件事，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 314, "start": 872.177, "end": 875.92, "text": "它會不會被快速運算很多次，以及它需不需要推理。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 315, "start": 875.92, "end": 879.722, "text": "這五個裡面如果有一個你等一下就想試，幫我按個訂閱、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 316, "start": 879.722, "end": 880.8, "text": "開個小鈴鐺。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 317, "start": 880.8, "end": 883.604, "text": "所有平台都叫 AI 懶人報，習慣用聽的就搜 Podcast。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 318, "start": 883.604, "end": 885.37, "text": "我們下一集見。", "source": "video_source_subtitle", "authority": "video_source_subtitle" } ], "transcript_selection": { "schema": "videonote.transcript_selection.v1", "status": "SELECTED", "requested_language": "zh", "duration": 886.0, "reference_path": "/Users/sagawa/AI_WORK/video_notes/url/20261001_084428__EP408_-_Jev_最好用的五個用法_每一個都給你可以照抄的範例_什麼樣的判斷才該換成它/reference_subtitle.vtt", "reference_kind": "manual", "reference_language": "zh-TW-p7rCI5_Nyzo", "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": 319, "duration_coverage": 0.999289 } } }