{
  "text": "OpenAI Codex Harness開源了,模型只是發動機,Harness才是那套駕駛系統。\n8月20日,OpenAI聯合創始人兼總裁Greg Brockman又一次把話題帶回了Codex。\n他提到一個挺反直覺的案例,一套由Codex驅動的稅務系統,在試點中處理了大約7000份報稅表,把稅務準備時間縮短了1 3左右。\n但如果你只把這理解成OpenAI又升級了一次編程AI,可能就錯過重點了。\n這次真正值得聊的,甚至不是哪個新模型,而是模型背後那套執行系統Harness。\n很多人第一次聽到這個詞,會以為它就是個開發框架,其實不太準確。\n它更像是AI智能體的執行系統。\n今天一個能連續工作幾十分鐘甚至幾個小時的Agent,Code從來不只是好模型加好Prompt。\n他得知道什麼時候該讀文件,什麼時候該搜索信息,什麼時候該調用工具,\n任務失敗了怎麼恢復,上下輪越堆越長了怎麼壓縮,\n哪些操作可以自動跑,哪些必須停下來等人審批。\n他還要維護任務狀態,保存現成,處理權限,記錄過程,把工具執行的結果不斷送回模型,讓模型決定下一步該幹嘛。\n說白了,模型負責思考,Harness負責把這種思考真正落進現實世界。\n這也是為什麼Agent時代正在出現一個判斷,以後決定一個AI產品體驗好不好的,可能不只是用了哪個模型,而是這個模型是怎麼被組織起來工作的。\nOpenAI自己就給了個很典型的例子。\n在ARC-IGS這個公開測試集上,同樣是GPT5.6SAL這個模型,用官方Harness跑,得分13.3%。\n等系統進一步打開Retained Reasoning,也就是保留推理狀態和Compaction上下輪壓縮之後,得分直接提到38.3%,輸出的Token數量反而降到了大概1.6。\n注意,這不是說模型智商突然翻了三倍,真正值得關注的是,模型壓根沒換,系統表現卻發生了這麼大的變化。\n這說明大模型的競爭正在從單純比參數,比推理能力,往更深一層下沉。\n上下文怎麼管,工具怎麼組織,任務怎麼循環,錯誤怎麼恢復,人什麼時候該介入。\n模型是發動機,Harness才是變速箱,方向盤,導航,刹車,是整套駕駛控制系統。\n而Codex正在把這套能力越來越完整的開放給開發者。\n這裡有個容易被誤讀的地方,得澄清一下。Codex的Harness並不是昨天才突然開源的。\nCodex的代碼倉庫早就用的是Arpreci 2.0許可證。\n2025年,OpenAI就發布了Codex SDK,開發者可以把整套Agent能力欠進自己的工作流。\n到了2026年2月,又進一步公開了Codex App Server,一套雙向JSON RPC接口,\n能把Codex底層的Agent循環接近不同的產品界面裡。\n所以這次真正值得討論的,不是OpenAI突然把秘密武器免費送人,\n而是一個更大的趨勢正在成熟。Codex正在從一個產品慢慢變成可以嵌入其他產品的基礎能力。\n開發者大致有三條路可以走。\n最簡單的是Codex Exact,可以把Agent直接塞進腳本或者CI流程裡,幹一些邊界清楚的活。\n再往上是Codex SDK,應用可以創建、恢復現成、運行任務、接收結構化輸出,把Agent真正寫進業務邏輯裡。\n更深一層是App Server,它甚至不要求你的產品長得像Codex,你可以保留自己原本的界面,只是把底層的Agent循環接近來。\n這事看起來只是個技術架構上的變化,但可能會改寫AI軟件過去三年的一個基本套路。\n從ChatGPT爆發以來,行業裡幾乎形成了一種條件反射,只要用上AI就加一個聊天礦,客服要加,數據分析要加,CRM要加,財務軟件也要加,幾乎無一例外。\n結果就是底層模型越來越聰明,前端產品卻越來越像。\n但真實的工作根本不是圍繞聊天發生的。\n安全工程師面對的是告警列表,財務人員面對的是賬目表格,客服面對的是訂單和公班,物流面對的是貨單和庫存。\n醫生、律師、工程師,每個行業都有自己幾十年打磨出來的專業界面和操作習慣。\n凭什麼為了用AI,所有人都得離開這些系統,跑到一個空白聊天礦裡,把上下文重新說一遍。\n更合理的方向其實反過來,不是讓人去適應AI的界面,而是讓AI進到人本來就在用的界面裡去。\nCisco已經給出了一個現實案例,他在Cisco Cloud Control裡推出了App Builder,直接把Codex SDK嵌進平台,\n讓客戶用自然語言創建定制應用,不用先跳到另一個AI工具裡幹完活,再把結果搬回來。\n稅務場景則更能說明問題。\nOpenAI, Thrive Holdings和Crate一起做的Tax AI,\n壓根沒打算讓會計師每天跟機器人聊天,而是直接進入了原本就存在的報稅流程。\n這一季度試點處理了700份報稅表,從業人員的修改、審核、糾錯,又被轉化成新的評測信號,繼續推動系統改進。\n這可能才是Agent真正能走進企業核心業務的方式。\nAI不是另外多出來的一個軟件,而是變成了原有軟件裡的一層隱形執行能力。\n當然,這裡有個邊界得說清楚。\nCodex的代碼用Apache 2.0許可證開源,不代表OpenAI把GPT模型的權重也一併開源了。\n開源Harness也不代表模型推理、算力、企業部署這些就變成零成本了。\n更不代表公司可以毫無安全措施地把數據庫、內部API、敏感文件全都甩給Agent去處理。\n真正成熟的企業及Agent反而會越來越強調全線邊界、殺箱、審計、工具白名單,以及關鍵動作必須有人確認的Human in the Loop機制。\n而這或許才是Codex Harness最值得關注的地方。\n過去幾年我們一直在問,下一代模型會有多聰明?\n但接下來,軟件行業可能會越來越頻繁地面對另一個問題。\n怎麼把這份聰明,安全、穩定、可控地裝進真實世界?\n從這個角度看,Harness的價值就很清楚了。\n模型決定的是智能的上限,Harness決定的是這種智能能不能真正幹活。\n而應用開發者真正的護城河也會重新回到自己的數據、工作流、專業界面、全線體系和行業理解上。\n未來最強大的AI產品,可能未必天天提醒你這裡有個AI。\n它可能沒有巨大的輸入框,沒有閃爍的機器人頭像,也不會要求你重新學一套操作方式。\n你打開的還是那個熟悉的財務軟件,運營看板,IDE,客服系統。\n只是從某一天開始,這些軟件突然開始自己理解任務,自己調用工具,自己執行工作。\n只在真正需要你的時候才停下來,問你一句,這一步要執行嗎?",
  "language": "zh",
  "duration": 397.966813,
  "segments": [
    {
      "id": 0,
      "seek": 0,
      "start": 0.0,
      "end": 6.0,
      "text": "OpenAI Codex Harness開源了,模型只是發動機,Harness才是那套駕駛系統。",
      "tokens": [
        50365,
        45569,
        48698,
        15549,
        87,
        389,
        16937,
        8949,
        47402,
        2289,
        11,
        41908,
        39823,
        36859,
        14637,
        12572,
        17543,
        11,
        39,
        16937,
        18888,
        1541,
        4184,
        1881,
        245,
        165,
        100,
        243,
        165,
        100,
        249,
        25368,
        33725,
        1543,
        50665
      ],
      "temperature": 0,
      "avg_logprob": -0.10461287064985796,
      "compression_ratio": 1.1782729805013927,
      "no_speech_prob": 8.662922809654638e-12,
      "source": "whisper"
    },
    {
      "id": 1,
      "seek": 0,
      "start": 6.0,
      "end": 11.0,
      "text": "8月20日,OpenAI聯合創始人兼總裁Greg Brockman又一次把話題帶回了Codex。",
      "tokens": [
        50665,
        23,
        6939,
        2009,
        6890,
        11,
        45569,
        48698,
        43942,
        14245,
        5935,
        113,
        15476,
        4035,
        2347,
        120,
        26575,
        8083,
        223,
        48254,
        32093,
        1601,
        17047,
        27505,
        16075,
        11103,
        14475,
        32379,
        8350,
        2289,
        34,
        1429,
        87,
        1543,
        50915
      ],
      "temperature": 0,
      "avg_logprob": -0.10461287064985796,
      "compression_ratio": 1.1782729805013927,
      "no_speech_prob": 8.662922809654638e-12,
      "source": "whisper"
    },
    {
      "id": 2,
      "seek": 0,
      "start": 11.0,
      "end": 21.0,
      "text": "他提到一個挺反直覺的案例,一套由Codex驅動的稅務系統,在試點中處理了大約7000份報稅表,把稅務準備時間縮短了1 3左右。",
      "tokens": [
        50915,
        5000,
        20949,
        4511,
        8990,
        41046,
        22138,
        16186,
        9463,
        1546,
        28899,
        17797,
        11,
        2257,
        1881,
        245,
        23786,
        34,
        1429,
        87,
        24023,
        227,
        12572,
        1546,
        10415,
        227,
        39517,
        25368,
        33725,
        11,
        3581,
        22099,
        8216,
        5975,
        39289,
        13876,
        2289,
        3582,
        30889,
        22,
        1360,
        36266,
        17803,
        10415,
        227,
        17571,
        11,
        16075,
        10415,
        227,
        39517,
        36241,
        20788,
        21030,
        106,
        5881,
        255,
        2289,
        16,
        805,
        29457,
        1543,
        51415
      ],
      "temperature": 0,
      "avg_logprob": -0.10461287064985796,
      "compression_ratio": 1.1782729805013927,
      "no_speech_prob": 8.662922809654638e-12,
      "source": "whisper"
    },
    {
      "id": 3,
      "seek": 0,
      "start": 21.0,
      "end": 26.0,
      "text": "但如果你只把這理解成OpenAI又升級了一次編程AI,可能就錯過重點了。",
      "tokens": [
        51415,
        8395,
        45669,
        14003,
        16075,
        2664,
        13876,
        17278,
        11336,
        45569,
        48698,
        17047,
        41670,
        35315,
        2289,
        27505,
        12608,
        101,
        29649,
        48698,
        11,
        16657,
        3111,
        13133,
        8816,
        12624,
        8216,
        2289,
        1543,
        51665
      ],
      "temperature": 0,
      "avg_logprob": -0.10461287064985796,
      "compression_ratio": 1.1782729805013927,
      "no_speech_prob": 8.662922809654638e-12,
      "source": "whisper"
    },
    {
      "id": 4,
      "seek": 2600,
      "start": 26.0,
      "end": 32.0,
      "text": "這次真正值得聊的,甚至不是哪個新模型,而是模型背後那套執行系統Harness。",
      "tokens": [
        50365,
        2664,
        9487,
        6303,
        15789,
        40242,
        5916,
        40096,
        1546,
        11,
        17556,
        20844,
        7296,
        17028,
        3338,
        12560,
        41908,
        39823,
        11,
        11070,
        1541,
        41908,
        39823,
        46329,
        5661,
        4184,
        1881,
        245,
        16262,
        115,
        8082,
        25368,
        33725,
        39,
        16937,
        1543,
        50665
      ],
      "temperature": 0,
      "avg_logprob": -0.0698443188693354,
      "compression_ratio": 1.305263157894737,
      "no_speech_prob": 1.3890382767611786e-11,
      "source": "whisper"
    },
    {
      "id": 5,
      "seek": 2600,
      "start": 32.0,
      "end": 37.0,
      "text": "很多人第一次聽到這個詞,會以為它就是個開發框架,其實不太準確。",
      "tokens": [
        50665,
        20778,
        4035,
        18049,
        9487,
        21524,
        4511,
        6287,
        6716,
        252,
        11,
        6236,
        3588,
        6344,
        11284,
        5620,
        3338,
        8949,
        14637,
        21416,
        228,
        7360,
        114,
        11,
        14139,
        1960,
        9455,
        28220,
        24293,
        1543,
        50915
      ],
      "temperature": 0,
      "avg_logprob": -0.0698443188693354,
      "compression_ratio": 1.305263157894737,
      "no_speech_prob": 1.3890382767611786e-11,
      "source": "whisper"
    },
    {
      "id": 6,
      "seek": 2600,
      "start": 37.0,
      "end": 39.0,
      "text": "它更像是AI智能體的執行系統。",
      "tokens": [
        50915,
        11284,
        19002,
        12760,
        1541,
        48698,
        5094,
        118,
        8225,
        23987,
        1546,
        16262,
        115,
        8082,
        25368,
        33725,
        1543,
        51015
      ],
      "temperature": 0,
      "avg_logprob": -0.0698443188693354,
      "compression_ratio": 1.305263157894737,
      "no_speech_prob": 1.3890382767611786e-11,
      "source": "whisper"
    },
    {
      "id": 7,
      "seek": 2600,
      "start": 39.0,
      "end": 45.0,
      "text": "今天一個能連續工作幾十分鐘甚至幾個小時的Agent,Code從來不只是好模型加好Prompt。",
      "tokens": [
        51015,
        12074,
        8990,
        8225,
        36252,
        29583,
        41315,
        23575,
        20145,
        32611,
        17556,
        20844,
        23575,
        3338,
        46997,
        1546,
        32,
        6930,
        11,
        34,
        1429,
        21068,
        3763,
        1960,
        36859,
        2131,
        41908,
        39823,
        9990,
        2131,
        47,
        4397,
        662,
        1543,
        51315
      ],
      "temperature": 0,
      "avg_logprob": -0.0698443188693354,
      "compression_ratio": 1.305263157894737,
      "no_speech_prob": 1.3890382767611786e-11,
      "source": "whisper"
    },
    {
      "id": 8,
      "seek": 2600,
      "start": 45.0,
      "end": 50.0,
      "text": "他得知道什麼時候該讀文件,什麼時候該搜索信息,什麼時候該調用工具,",
      "tokens": [
        51315,
        5000,
        5916,
        7758,
        7598,
        14010,
        22033,
        7422,
        222,
        17174,
        20485,
        11,
        7598,
        14010,
        22033,
        162,
        14987,
        7732,
        95,
        17665,
        26460,
        11,
        7598,
        14010,
        22033,
        32943,
        9254,
        23323,
        39806,
        11,
        51565
      ],
      "temperature": 0,
      "avg_logprob": -0.0698443188693354,
      "compression_ratio": 1.305263157894737,
      "no_speech_prob": 1.3890382767611786e-11,
      "source": "whisper"
    },
    {
      "id": 9,
      "seek": 2600,
      "start": 50.0,
      "end": 54.0,
      "text": "任務失敗了怎麼恢復,上下輪越堆越長了怎麼壓縮,",
      "tokens": [
        51565,
        26443,
        39517,
        28564,
        7017,
        245,
        2289,
        11614,
        11845,
        95,
        2172,
        102,
        11,
        5708,
        4438,
        12290,
        103,
        25761,
        10726,
        228,
        25761,
        15353,
        2289,
        11614,
        45964,
        21030,
        106,
        11,
        51765
      ],
      "temperature": 0,
      "avg_logprob": -0.0698443188693354,
      "compression_ratio": 1.305263157894737,
      "no_speech_prob": 1.3890382767611786e-11,
      "source": "whisper"
    },
    {
      "id": 10,
      "seek": 5400,
      "start": 54.0,
      "end": 58.0,
      "text": "哪些操作可以自動跑,哪些必須停下來等人審批。",
      "tokens": [
        50365,
        17028,
        13824,
        13167,
        235,
        11914,
        6723,
        9722,
        12572,
        32585,
        11,
        17028,
        13824,
        28531,
        8313,
        230,
        36135,
        27769,
        10187,
        4035,
        4510,
        102,
        3416,
        117,
        1543,
        50565
      ],
      "temperature": 0,
      "avg_logprob": -0.05985308101034572,
      "compression_ratio": 1.1620553359683794,
      "no_speech_prob": 2.0937935413245512e-11,
      "source": "whisper"
    },
    {
      "id": 11,
      "seek": 5400,
      "start": 58.0,
      "end": 66.0,
      "text": "他還要維護任務狀態,保存現成,處理權限,記錄過程,把工具執行的結果不斷送回模型,讓模型決定下一步該幹嘛。",
      "tokens": [
        50565,
        5000,
        7824,
        4275,
        9261,
        255,
        48289,
        26443,
        39517,
        18637,
        222,
        45248,
        11,
        24302,
        39781,
        9581,
        11336,
        11,
        39289,
        13876,
        40212,
        43446,
        11,
        16958,
        9567,
        226,
        8816,
        29649,
        11,
        16075,
        23323,
        39806,
        16262,
        115,
        8082,
        1546,
        35181,
        1960,
        4307,
        115,
        29309,
        8350,
        41908,
        39823,
        11,
        21195,
        41908,
        39823,
        33540,
        12088,
        45271,
        31429,
        22033,
        44854,
        1543,
        50965
      ],
      "temperature": 0,
      "avg_logprob": -0.05985308101034572,
      "compression_ratio": 1.1620553359683794,
      "no_speech_prob": 2.0937935413245512e-11,
      "source": "whisper"
    },
    {
      "id": 12,
      "seek": 5400,
      "start": 66.0,
      "end": 71.0,
      "text": "說白了,模型負責思考,Harness負責把這種思考真正落進現實世界。",
      "tokens": [
        50965,
        4622,
        13558,
        2289,
        11,
        41908,
        39823,
        11561,
        254,
        11561,
        105,
        8870,
        26504,
        11,
        39,
        16937,
        11561,
        254,
        11561,
        105,
        16075,
        33005,
        8870,
        26504,
        6303,
        15789,
        30848,
        18214,
        9581,
        10376,
        24486,
        1543,
        51215
      ],
      "temperature": 0,
      "avg_logprob": -0.05985308101034572,
      "compression_ratio": 1.1620553359683794,
      "no_speech_prob": 2.0937935413245512e-11,
      "source": "whisper"
    },
    {
      "id": 13,
      "seek": 7100,
      "start": 71.0,
      "end": 83.0,
      "text": "這也是為什麼Agent時代正在出現一個判斷,以後決定一個AI產品體驗好不好的,可能不只是用了哪個模型,而是這個模型是怎麼被組織起來工作的。",
      "tokens": [
        50365,
        2664,
        22021,
        18987,
        32,
        6930,
        6611,
        19105,
        15789,
        3581,
        7781,
        9581,
        8990,
        2437,
        97,
        4307,
        115,
        11,
        3588,
        5661,
        33540,
        12088,
        8990,
        48698,
        33299,
        30246,
        23987,
        24023,
        245,
        34291,
        1546,
        11,
        16657,
        1960,
        36859,
        9254,
        2289,
        17028,
        3338,
        41908,
        39823,
        11,
        11070,
        1541,
        6287,
        41908,
        39823,
        1541,
        11614,
        23238,
        30052,
        23141,
        242,
        21670,
        41315,
        1546,
        1543,
        50965
      ],
      "temperature": 0,
      "avg_logprob": -0.06396377860725701,
      "compression_ratio": 1.1153846153846154,
      "no_speech_prob": 2.167715312806351e-11,
      "source": "whisper"
    },
    {
      "id": 14,
      "seek": 7100,
      "start": 83.0,
      "end": 85.0,
      "text": "OpenAI自己就給了個很典型的例子。",
      "tokens": [
        50965,
        45569,
        48698,
        17645,
        3111,
        17798,
        2289,
        3338,
        4563,
        2347,
        116,
        39823,
        1546,
        17797,
        7626,
        1543,
        51065
      ],
      "temperature": 0,
      "avg_logprob": -0.06396377860725701,
      "compression_ratio": 1.1153846153846154,
      "no_speech_prob": 2.167715312806351e-11,
      "source": "whisper"
    },
    {
      "id": 15,
      "seek": 8500,
      "start": 85.0,
      "end": 94.0,
      "text": "在ARC-IGS這個公開測試集上,同樣是GPT5.6SAL這個模型,用官方Harness跑,得分13.3%。",
      "tokens": [
        50365,
        3581,
        1899,
        34,
        12,
        40,
        24446,
        6287,
        13545,
        8949,
        9592,
        105,
        22099,
        26020,
        5708,
        11,
        13089,
        6180,
        1541,
        38,
        47,
        51,
        20,
        13,
        21,
        50,
        3427,
        6287,
        41908,
        39823,
        11,
        9254,
        31929,
        9249,
        39,
        16937,
        32585,
        11,
        5916,
        6627,
        7668,
        13,
        18,
        4,
        1543,
        50815
      ],
      "temperature": 0,
      "avg_logprob": -0.1662745475769043,
      "compression_ratio": 1.0072992700729928,
      "no_speech_prob": 2.062615009346125e-11,
      "source": "whisper"
    },
    {
      "id": 16,
      "seek": 8500,
      "start": 94.0,
      "end": 105.0,
      "text": "等系統進一步打開Retained Reasoning,也就是保留推理狀態和Compaction上下輪壓縮之後,得分直接提到38.3%,輸出的Token數量反而降到了大概1.6。",
      "tokens": [
        50815,
        10187,
        25368,
        33725,
        18214,
        2257,
        31429,
        12467,
        8949,
        49,
        302,
        3563,
        39693,
        278,
        11,
        6404,
        5620,
        24302,
        24456,
        33597,
        13876,
        18637,
        222,
        45248,
        12565,
        34,
        8586,
        2894,
        5708,
        4438,
        12290,
        103,
        45964,
        21030,
        106,
        26362,
        11,
        5916,
        6627,
        43297,
        20949,
        4511,
        12625,
        13,
        18,
        8923,
        12290,
        116,
        7781,
        1546,
        51,
        8406,
        30622,
        26748,
        22138,
        11070,
        47421,
        21381,
        32044,
        16,
        13,
        21,
        1543,
        51365
      ],
      "temperature": 0,
      "avg_logprob": -0.1662745475769043,
      "compression_ratio": 1.0072992700729928,
      "no_speech_prob": 2.062615009346125e-11,
      "source": "whisper"
    },
    {
      "id": 17,
      "seek": 10500,
      "start": 105.0,
      "end": 113.0,
      "text": "注意,這不是說模型智商突然翻了三倍,真正值得關注的是,模型壓根沒換,系統表現卻發生了這麼大的變化。",
      "tokens": [
        50365,
        40794,
        11,
        2664,
        7296,
        4622,
        41908,
        39823,
        5094,
        118,
        45581,
        40859,
        5823,
        42716,
        2289,
        10960,
        35477,
        11,
        6303,
        15789,
        40242,
        5916,
        14899,
        26432,
        24620,
        11,
        41908,
        39823,
        45964,
        31337,
        4913,
        36338,
        11,
        25368,
        33725,
        17571,
        9581,
        5322,
        119,
        14637,
        8244,
        2289,
        21269,
        39156,
        26327,
        23756,
        1543,
        50765
      ],
      "temperature": 0,
      "avg_logprob": -0.057722732938569166,
      "compression_ratio": 1.2759643916913948,
      "no_speech_prob": 2.0526413901933438e-11,
      "source": "whisper"
    },
    {
      "id": 18,
      "seek": 10500,
      "start": 113.0,
      "end": 119.0,
      "text": "這說明大模型的競爭正在從單純比參數,比推理能力,往更深一層下沉。",
      "tokens": [
        50765,
        2664,
        4622,
        11100,
        3582,
        41908,
        39823,
        1546,
        11957,
        114,
        8164,
        255,
        15789,
        3581,
        21068,
        24227,
        7732,
        242,
        11706,
        47375,
        30622,
        11,
        11706,
        33597,
        13876,
        8225,
        13486,
        11,
        29510,
        19002,
        24043,
        2257,
        9636,
        97,
        4438,
        3308,
        231,
        1543,
        51065
      ],
      "temperature": 0,
      "avg_logprob": -0.057722732938569166,
      "compression_ratio": 1.2759643916913948,
      "no_speech_prob": 2.0526413901933438e-11,
      "source": "whisper"
    },
    {
      "id": 19,
      "seek": 10500,
      "start": 119.0,
      "end": 125.0,
      "text": "上下文怎麼管,工具怎麼組織,任務怎麼循環,錯誤怎麼恢復,人什麼時候該介入。",
      "tokens": [
        51065,
        5708,
        4438,
        17174,
        11614,
        23131,
        11,
        23323,
        39806,
        11614,
        30052,
        23141,
        242,
        11,
        26443,
        39517,
        11614,
        2172,
        103,
        40063,
        108,
        11,
        13133,
        3549,
        97,
        11614,
        11845,
        95,
        2172,
        102,
        11,
        4035,
        7598,
        14010,
        22033,
        30312,
        14028,
        1543,
        51365
      ],
      "temperature": 0,
      "avg_logprob": -0.057722732938569166,
      "compression_ratio": 1.2759643916913948,
      "no_speech_prob": 2.0526413901933438e-11,
      "source": "whisper"
    },
    {
      "id": 20,
      "seek": 10500,
      "start": 125.0,
      "end": 132.0,
      "text": "模型是發動機,Harness才是變速箱,方向盤,導航,刹車,是整套駕駛控制系統。",
      "tokens": [
        51365,
        41908,
        39823,
        1541,
        14637,
        12572,
        17543,
        11,
        39,
        16937,
        18888,
        1541,
        26327,
        31217,
        11249,
        109,
        11,
        9249,
        24282,
        5419,
        97,
        11,
        25647,
        10256,
        103,
        11,
        2437,
        117,
        25941,
        11,
        1541,
        27662,
        1881,
        245,
        165,
        100,
        243,
        165,
        100,
        249,
        48707,
        25491,
        25368,
        33725,
        1543,
        51715
      ],
      "temperature": 0,
      "avg_logprob": -0.057722732938569166,
      "compression_ratio": 1.2759643916913948,
      "no_speech_prob": 2.0526413901933438e-11,
      "source": "whisper"
    },
    {
      "id": 21,
      "seek": 13200,
      "start": 132.0,
      "end": 137.0,
      "text": "而Codex正在把這套能力越來越完整的開放給開發者。",
      "tokens": [
        50365,
        11070,
        34,
        1429,
        87,
        15789,
        3581,
        16075,
        2664,
        1881,
        245,
        8225,
        13486,
        25761,
        3763,
        25761,
        14128,
        27662,
        1546,
        8949,
        12744,
        17798,
        8949,
        14637,
        12444,
        1543,
        50615
      ],
      "temperature": 0,
      "avg_logprob": -0.10464358042521649,
      "compression_ratio": 1.183381088825215,
      "no_speech_prob": 1.870195838471478e-11,
      "source": "whisper"
    },
    {
      "id": 22,
      "seek": 13200,
      "start": 137.0,
      "end": 143.0,
      "text": "這裡有個容易被誤讀的地方,得澄清一下。Codex的Harness並不是昨天才突然開源的。",
      "tokens": [
        50615,
        18771,
        2412,
        3338,
        49212,
        23238,
        3549,
        97,
        7422,
        222,
        1546,
        30146,
        11,
        5916,
        25621,
        226,
        21784,
        8861,
        1543,
        34,
        1429,
        87,
        1546,
        39,
        16937,
        35934,
        7296,
        47404,
        6135,
        18888,
        40859,
        5823,
        8949,
        47402,
        1546,
        1543,
        50915
      ],
      "temperature": 0,
      "avg_logprob": -0.10464358042521649,
      "compression_ratio": 1.183381088825215,
      "no_speech_prob": 1.870195838471478e-11,
      "source": "whisper"
    },
    {
      "id": 23,
      "seek": 13200,
      "start": 143.0,
      "end": 147.0,
      "text": "Codex的代碼倉庫早就用的是Arpreci 2.0許可證。",
      "tokens": [
        50915,
        34,
        1429,
        87,
        1546,
        19105,
        16337,
        120,
        2122,
        231,
        6346,
        104,
        21176,
        3111,
        9254,
        24620,
        10683,
        79,
        3326,
        568,
        13,
        15,
        40914,
        4429,
        11823,
        231,
        1543,
        51115
      ],
      "temperature": 0,
      "avg_logprob": -0.10464358042521649,
      "compression_ratio": 1.183381088825215,
      "no_speech_prob": 1.870195838471478e-11,
      "source": "whisper"
    },
    {
      "id": 24,
      "seek": 13200,
      "start": 147.0,
      "end": 154.0,
      "text": "2025年,OpenAI就發布了Codex SDK,開發者可以把整套Agent能力欠進自己的工作流。",
      "tokens": [
        51115,
        2009,
        6074,
        5157,
        11,
        45569,
        48698,
        3111,
        14637,
        34688,
        2289,
        34,
        1429,
        87,
        37135,
        11,
        8949,
        14637,
        12444,
        6723,
        16075,
        27662,
        1881,
        245,
        32,
        6930,
        8225,
        13486,
        5988,
        254,
        18214,
        17645,
        1546,
        41315,
        27854,
        1543,
        51465
      ],
      "temperature": 0,
      "avg_logprob": -0.10464358042521649,
      "compression_ratio": 1.183381088825215,
      "no_speech_prob": 1.870195838471478e-11,
      "source": "whisper"
    },
    {
      "id": 25,
      "seek": 13200,
      "start": 154.0,
      "end": 160.0,
      "text": "到了2026年2月,又進一步公開了Codex App Server,一套雙向JSON RPC接口,",
      "tokens": [
        51465,
        21381,
        2009,
        10880,
        5157,
        17,
        6939,
        11,
        17047,
        18214,
        2257,
        31429,
        13545,
        8949,
        2289,
        34,
        1429,
        87,
        3132,
        25684,
        11,
        2257,
        1881,
        245,
        6306,
        247,
        24282,
        41,
        10388,
        497,
        12986,
        14468,
        18144,
        11,
        51765
      ],
      "temperature": 0,
      "avg_logprob": -0.10464358042521649,
      "compression_ratio": 1.183381088825215,
      "no_speech_prob": 1.870195838471478e-11,
      "source": "whisper"
    },
    {
      "id": 26,
      "seek": 16000,
      "start": 160.0,
      "end": 164.0,
      "text": "能把Codex底層的Agent循環接近不同的產品界面裡。",
      "tokens": [
        50365,
        8225,
        16075,
        34,
        1429,
        87,
        22816,
        9636,
        97,
        1546,
        32,
        6930,
        2172,
        103,
        40063,
        108,
        14468,
        17463,
        47123,
        1546,
        33299,
        30246,
        21185,
        8833,
        11066,
        1543,
        50565
      ],
      "temperature": 0,
      "avg_logprob": -0.05069079399108887,
      "compression_ratio": 1.1553030303030303,
      "no_speech_prob": 1.677541612565836e-11,
      "source": "whisper"
    },
    {
      "id": 27,
      "seek": 16000,
      "start": 164.0,
      "end": 169.0,
      "text": "所以這次真正值得討論的,不是OpenAI突然把秘密武器免費送人,",
      "tokens": [
        50565,
        7239,
        2664,
        9487,
        6303,
        15789,
        40242,
        5916,
        5396,
        236,
        29340,
        1546,
        11,
        7296,
        45569,
        48698,
        40859,
        5823,
        16075,
        8204,
        246,
        47838,
        38194,
        34386,
        2347,
        235,
        50071,
        29309,
        4035,
        11,
        50815
      ],
      "temperature": 0,
      "avg_logprob": -0.05069079399108887,
      "compression_ratio": 1.1553030303030303,
      "no_speech_prob": 1.677541612565836e-11,
      "source": "whisper"
    },
    {
      "id": 28,
      "seek": 16000,
      "start": 169.0,
      "end": 177.0,
      "text": "而是一個更大的趨勢正在成熟。Codex正在從一個產品慢慢變成可以嵌入其他產品的基礎能力。",
      "tokens": [
        50815,
        11070,
        1541,
        8990,
        19002,
        39156,
        9759,
        101,
        7978,
        95,
        15789,
        3581,
        11336,
        21805,
        253,
        1543,
        34,
        1429,
        87,
        15789,
        3581,
        21068,
        8990,
        33299,
        30246,
        41439,
        26327,
        11336,
        6723,
        161,
        113,
        234,
        14028,
        9572,
        5000,
        33299,
        30246,
        1546,
        26008,
        14971,
        236,
        8225,
        13486,
        1543,
        51215
      ],
      "temperature": 0,
      "avg_logprob": -0.05069079399108887,
      "compression_ratio": 1.1553030303030303,
      "no_speech_prob": 1.677541612565836e-11,
      "source": "whisper"
    },
    {
      "id": 29,
      "seek": 16000,
      "start": 177.0,
      "end": 179.0,
      "text": "開發者大致有三條路可以走。",
      "tokens": [
        51215,
        8949,
        14637,
        12444,
        3582,
        6784,
        112,
        2412,
        10960,
        31106,
        24658,
        6723,
        9575,
        1543,
        51315
      ],
      "temperature": 0,
      "avg_logprob": -0.05069079399108887,
      "compression_ratio": 1.1553030303030303,
      "no_speech_prob": 1.677541612565836e-11,
      "source": "whisper"
    },
    {
      "id": 30,
      "seek": 17900,
      "start": 179.0,
      "end": 186.0,
      "text": "最簡單的是Codex Exact,可以把Agent直接塞進腳本或者CI流程裡,幹一些邊界清楚的活。",
      "tokens": [
        50365,
        8661,
        31995,
        24620,
        34,
        1429,
        87,
        7199,
        11,
        6723,
        16075,
        32,
        6930,
        43297,
        13331,
        252,
        18214,
        49595,
        8802,
        31148,
        25240,
        27854,
        29649,
        11066,
        11,
        29932,
        38515,
        14901,
        21185,
        38433,
        1546,
        25956,
        1543,
        50715
      ],
      "temperature": 0,
      "avg_logprob": -0.1253423061999646,
      "compression_ratio": 1.0619469026548674,
      "no_speech_prob": 1.657787969400193e-11,
      "source": "whisper"
    },
    {
      "id": 31,
      "seek": 17900,
      "start": 186.0,
      "end": 195.0,
      "text": "再往上是Codex SDK,應用可以創建、恢復現成、運行任務、接收結構化輸出,把Agent真正寫進業務邏輯裡。",
      "tokens": [
        50715,
        8623,
        29510,
        5708,
        1541,
        34,
        1429,
        87,
        37135,
        11,
        20481,
        9254,
        6723,
        5935,
        113,
        34157,
        1231,
        11845,
        95,
        2172,
        102,
        9581,
        11336,
        1231,
        32772,
        8082,
        26443,
        39517,
        1231,
        14468,
        18681,
        17144,
        43362,
        23756,
        12290,
        116,
        7781,
        11,
        16075,
        32,
        6930,
        6303,
        15789,
        4510,
        104,
        18214,
        27119,
        39517,
        3023,
        237,
        12290,
        107,
        11066,
        1543,
        51165
      ],
      "temperature": 0,
      "avg_logprob": -0.1253423061999646,
      "compression_ratio": 1.0619469026548674,
      "no_speech_prob": 1.657787969400193e-11,
      "source": "whisper"
    },
    {
      "id": 32,
      "seek": 19500,
      "start": 195.0,
      "end": 204.0,
      "text": "更深一層是App Server,它甚至不要求你的產品長得像Codex,你可以保留自己原本的界面,只是把底層的Agent循環接近來。",
      "tokens": [
        50365,
        19002,
        24043,
        2257,
        9636,
        97,
        1541,
        9132,
        25684,
        11,
        11284,
        17556,
        20844,
        11962,
        32718,
        18961,
        33299,
        30246,
        15353,
        5916,
        12760,
        34,
        1429,
        87,
        11,
        42766,
        24302,
        24456,
        17645,
        19683,
        8802,
        1546,
        21185,
        8833,
        11,
        36859,
        16075,
        22816,
        9636,
        97,
        1546,
        32,
        6930,
        2172,
        103,
        40063,
        108,
        14468,
        17463,
        3763,
        1543,
        50815
      ],
      "temperature": 0,
      "avg_logprob": -0.06582887276359227,
      "compression_ratio": 1.0655737704918034,
      "no_speech_prob": 2.0233056549634476e-11,
      "source": "whisper"
    },
    {
      "id": 33,
      "seek": 19500,
      "start": 204.0,
      "end": 210.0,
      "text": "這事看起來只是個技術架構上的變化,但可能會改寫AI軟件過去三年的一個基本套路。",
      "tokens": [
        50815,
        2664,
        6973,
        4200,
        21670,
        36859,
        3338,
        32502,
        9890,
        241,
        7360,
        114,
        43362,
        5708,
        1546,
        26327,
        23756,
        11,
        8395,
        16657,
        6236,
        34490,
        4510,
        104,
        48698,
        14454,
        253,
        20485,
        36910,
        10960,
        34128,
        8990,
        37946,
        1881,
        245,
        24658,
        1543,
        51115
      ],
      "temperature": 0,
      "avg_logprob": -0.06582887276359227,
      "compression_ratio": 1.0655737704918034,
      "no_speech_prob": 2.0233056549634476e-11,
      "source": "whisper"
    },
    {
      "id": 34,
      "seek": 21000,
      "start": 210.0,
      "end": 223.0,
      "text": "從ChatGPT爆發以來,行業裡幾乎形成了一種條件反射,只要用上AI就加一個聊天礦,客服要加,數據分析要加,CRM要加,財務軟件也要加,幾乎無一例外。",
      "tokens": [
        50365,
        21068,
        41683,
        38,
        47,
        51,
        45827,
        14637,
        3588,
        3763,
        11,
        8082,
        27119,
        11066,
        23575,
        2930,
        236,
        30900,
        11336,
        2289,
        2257,
        16516,
        31106,
        20485,
        22138,
        1530,
        226,
        11,
        14003,
        4275,
        9254,
        5708,
        48698,
        3111,
        9990,
        8990,
        40096,
        6135,
        14971,
        99,
        11,
        32316,
        27408,
        4275,
        9990,
        11,
        30622,
        36841,
        6627,
        7360,
        238,
        4275,
        9990,
        11,
        18547,
        44,
        4275,
        9990,
        11,
        11561,
        94,
        39517,
        14454,
        253,
        20485,
        6404,
        4275,
        9990,
        11,
        23575,
        2930,
        236,
        16976,
        2257,
        17797,
        12022,
        1543,
        51015
      ],
      "temperature": 0,
      "avg_logprob": -0.0532019403245714,
      "compression_ratio": 1.351032448377581,
      "no_speech_prob": 1.96870939689342e-11,
      "source": "whisper"
    },
    {
      "id": 35,
      "seek": 21000,
      "start": 223.0,
      "end": 227.0,
      "text": "結果就是底層模型越來越聰明,前端產品卻越來越像。",
      "tokens": [
        51015,
        35181,
        5620,
        22816,
        9636,
        97,
        41908,
        39823,
        25761,
        3763,
        25761,
        8171,
        108,
        11100,
        11,
        8945,
        11957,
        107,
        33299,
        30246,
        5322,
        119,
        25761,
        3763,
        25761,
        12760,
        1543,
        51215
      ],
      "temperature": 0,
      "avg_logprob": -0.0532019403245714,
      "compression_ratio": 1.351032448377581,
      "no_speech_prob": 1.96870939689342e-11,
      "source": "whisper"
    },
    {
      "id": 36,
      "seek": 21000,
      "start": 227.0,
      "end": 230.0,
      "text": "但真實的工作根本不是圍繞聊天發生的。",
      "tokens": [
        51215,
        8395,
        6303,
        10376,
        1546,
        41315,
        31337,
        8802,
        7296,
        2523,
        235,
        23141,
        252,
        40096,
        6135,
        14637,
        8244,
        1546,
        1543,
        51365
      ],
      "temperature": 0,
      "avg_logprob": -0.0532019403245714,
      "compression_ratio": 1.351032448377581,
      "no_speech_prob": 1.96870939689342e-11,
      "source": "whisper"
    },
    {
      "id": 37,
      "seek": 21000,
      "start": 230.0,
      "end": 238.0,
      "text": "安全工程師面對的是告警列表,財務人員面對的是賬目表格,客服面對的是訂單和公班,物流面對的是貨單和庫存。",
      "tokens": [
        51365,
        45466,
        23323,
        29649,
        16276,
        8833,
        2855,
        24620,
        16846,
        39499,
        43338,
        17571,
        11,
        11561,
        94,
        39517,
        4035,
        20992,
        8833,
        2855,
        24620,
        13352,
        105,
        11386,
        17571,
        30921,
        11,
        32316,
        27408,
        8833,
        2855,
        24620,
        28320,
        24227,
        12565,
        13545,
        44695,
        11,
        23516,
        27854,
        8833,
        2855,
        24620,
        11561,
        101,
        24227,
        12565,
        6346,
        104,
        39781,
        1543,
        51765
      ],
      "temperature": 0,
      "avg_logprob": -0.0532019403245714,
      "compression_ratio": 1.351032448377581,
      "no_speech_prob": 1.96870939689342e-11,
      "source": "whisper"
    },
    {
      "id": 38,
      "seek": 23800,
      "start": 238.0,
      "end": 244.0,
      "text": "醫生、律師、工程師,每個行業都有自己幾十年打磨出來的專業界面和操作習慣。",
      "tokens": [
        50365,
        48351,
        8244,
        1231,
        2172,
        233,
        16276,
        1231,
        23323,
        29649,
        16276,
        11,
        23664,
        3338,
        8082,
        27119,
        48121,
        17645,
        23575,
        20145,
        5157,
        12467,
        163,
        13033,
        29741,
        1546,
        46072,
        27119,
        21185,
        8833,
        12565,
        13167,
        235,
        11914,
        34025,
        12358,
        96,
        1543,
        50665
      ],
      "temperature": 0,
      "avg_logprob": -0.06630031746554088,
      "compression_ratio": 1.194300518134715,
      "no_speech_prob": 1.709527484849982e-11,
      "source": "whisper"
    },
    {
      "id": 39,
      "seek": 23800,
      "start": 244.0,
      "end": 251.0,
      "text": "凭什麼為了用AI,所有人都得離開這些系統,跑到一個空白聊天礦裡,把上下文重新說一遍。",
      "tokens": [
        50665,
        6336,
        255,
        7598,
        6344,
        2289,
        9254,
        48698,
        11,
        39300,
        4035,
        7182,
        5916,
        32662,
        8949,
        29869,
        25368,
        33725,
        11,
        32585,
        4511,
        8990,
        23322,
        13558,
        40096,
        6135,
        14971,
        99,
        11066,
        11,
        16075,
        5708,
        4438,
        17174,
        12624,
        12560,
        4622,
        2257,
        3330,
        235,
        1543,
        51015
      ],
      "temperature": 0,
      "avg_logprob": -0.06630031746554088,
      "compression_ratio": 1.194300518134715,
      "no_speech_prob": 1.709527484849982e-11,
      "source": "whisper"
    },
    {
      "id": 40,
      "seek": 23800,
      "start": 251.0,
      "end": 258.0,
      "text": "更合理的方向其實反過來,不是讓人去適應AI的界面,而是讓AI進到人本來就在用的界面裡去。",
      "tokens": [
        51015,
        19002,
        14245,
        13876,
        1546,
        9249,
        24282,
        14139,
        22138,
        49950,
        11,
        7296,
        21195,
        4035,
        6734,
        3330,
        102,
        20481,
        48698,
        1546,
        21185,
        8833,
        11,
        11070,
        1541,
        21195,
        48698,
        18214,
        4511,
        4035,
        8802,
        3763,
        3111,
        3581,
        9254,
        1546,
        21185,
        8833,
        11066,
        6734,
        1543,
        51365
      ],
      "temperature": 0,
      "avg_logprob": -0.06630031746554088,
      "compression_ratio": 1.194300518134715,
      "no_speech_prob": 1.709527484849982e-11,
      "source": "whisper"
    },
    {
      "id": 41,
      "seek": 23800,
      "start": 258.0,
      "end": 266.0,
      "text": "Cisco已經給出了一個現實案例,他在Cisco Cloud Control裡推出了App Builder,直接把Codex SDK嵌進平台,",
      "tokens": [
        51365,
        34,
        8610,
        25338,
        17798,
        7781,
        2289,
        8990,
        9581,
        10376,
        28899,
        17797,
        11,
        5000,
        3581,
        34,
        8610,
        8061,
        12912,
        11066,
        33597,
        7781,
        2289,
        9132,
        11875,
        260,
        11,
        43297,
        16075,
        34,
        1429,
        87,
        37135,
        161,
        113,
        234,
        18214,
        16716,
        15433,
        11,
        51765
      ],
      "temperature": 0,
      "avg_logprob": -0.06630031746554088,
      "compression_ratio": 1.194300518134715,
      "no_speech_prob": 1.709527484849982e-11,
      "source": "whisper"
    },
    {
      "id": 42,
      "seek": 26600,
      "start": 266.0,
      "end": 273.0,
      "text": "讓客戶用自然語言創建定制應用,不用先跳到另一個AI工具裡幹完活,再把結果搬回來。",
      "tokens": [
        50365,
        21195,
        32316,
        1486,
        114,
        9254,
        9722,
        5823,
        31348,
        12009,
        5935,
        113,
        34157,
        12088,
        25491,
        20481,
        9254,
        11,
        24384,
        10108,
        29366,
        4511,
        22762,
        8990,
        48698,
        23323,
        39806,
        11066,
        29932,
        14128,
        25956,
        11,
        8623,
        16075,
        35181,
        22220,
        105,
        46556,
        1543,
        50715
      ],
      "temperature": 0,
      "avg_logprob": -0.10211096445719402,
      "compression_ratio": 0.9565217391304348,
      "no_speech_prob": 1.87891941588747e-11,
      "source": "whisper"
    },
    {
      "id": 43,
      "seek": 26600,
      "start": 273.0,
      "end": 275.0,
      "text": "稅務場景則更能說明問題。",
      "tokens": [
        50715,
        10415,
        227,
        39517,
        21446,
        50218,
        46225,
        19002,
        8225,
        4622,
        11100,
        17197,
        1543,
        50815
      ],
      "temperature": 0,
      "avg_logprob": -0.10211096445719402,
      "compression_ratio": 0.9565217391304348,
      "no_speech_prob": 1.87891941588747e-11,
      "source": "whisper"
    },
    {
      "id": 44,
      "seek": 26600,
      "start": 275.0,
      "end": 279.0,
      "text": "OpenAI, Thrive Holdings和Crate一起做的Tax AI,",
      "tokens": [
        50815,
        45569,
        48698,
        11,
        334,
        8003,
        6962,
        1109,
        12565,
        34,
        4404,
        29567,
        10907,
        1546,
        51,
        2797,
        7318,
        11,
        51015
      ],
      "temperature": 0,
      "avg_logprob": -0.10211096445719402,
      "compression_ratio": 0.9565217391304348,
      "no_speech_prob": 1.87891941588747e-11,
      "source": "whisper"
    },
    {
      "id": 45,
      "seek": 27900,
      "start": 279.0,
      "end": 286.0,
      "text": "壓根沒打算讓會計師每天跟機器人聊天,而是直接進入了原本就存在的報稅流程。",
      "tokens": [
        50365,
        45964,
        31337,
        4913,
        12467,
        19497,
        21195,
        6236,
        30114,
        16276,
        23664,
        6135,
        9678,
        17543,
        34386,
        4035,
        40096,
        6135,
        11,
        11070,
        1541,
        43297,
        18214,
        14028,
        2289,
        19683,
        8802,
        3111,
        39781,
        3581,
        1546,
        17803,
        10415,
        227,
        27854,
        29649,
        1543,
        50715
      ],
      "temperature": 0,
      "avg_logprob": -0.07157809463972897,
      "compression_ratio": 1.1210762331838564,
      "no_speech_prob": 1.8425063558202837e-11,
      "source": "whisper"
    },
    {
      "id": 46,
      "seek": 27900,
      "start": 286.0,
      "end": 295.0,
      "text": "這一季度試點處理了700份報稅表,從業人員的修改、審核、糾錯,又被轉化成新的評測信號,繼續推動系統改進。",
      "tokens": [
        50715,
        32260,
        8052,
        96,
        13127,
        22099,
        8216,
        39289,
        13876,
        2289,
        18197,
        36266,
        17803,
        10415,
        227,
        17571,
        11,
        21068,
        27119,
        4035,
        20992,
        1546,
        7792,
        106,
        34490,
        1231,
        4510,
        102,
        8987,
        116,
        1231,
        15239,
        122,
        13133,
        11,
        17047,
        23238,
        32739,
        23756,
        11336,
        12560,
        1546,
        39115,
        9592,
        105,
        17665,
        18616,
        11,
        38459,
        33597,
        12572,
        25368,
        33725,
        34490,
        18214,
        1543,
        51165
      ],
      "temperature": 0,
      "avg_logprob": -0.07157809463972897,
      "compression_ratio": 1.1210762331838564,
      "no_speech_prob": 1.8425063558202837e-11,
      "source": "whisper"
    },
    {
      "id": 47,
      "seek": 29500,
      "start": 295.0,
      "end": 298.0,
      "text": "這可能才是Agent真正能走進企業核心業務的方式。",
      "tokens": [
        50365,
        2664,
        16657,
        18888,
        1541,
        32,
        6930,
        6303,
        15789,
        8225,
        9575,
        18214,
        7384,
        223,
        27119,
        8987,
        116,
        7945,
        27119,
        39517,
        1546,
        9249,
        27584,
        1543,
        50515
      ],
      "temperature": 0,
      "avg_logprob": -0.07155802210823434,
      "compression_ratio": 1.1014492753623188,
      "no_speech_prob": 2.1357226751006486e-11,
      "source": "whisper"
    },
    {
      "id": 48,
      "seek": 29500,
      "start": 298.0,
      "end": 304.0,
      "text": "AI不是另外多出來的一個軟件,而是變成了原有軟件裡的一層隱形執行能力。",
      "tokens": [
        50515,
        48698,
        7296,
        26202,
        6392,
        29741,
        1546,
        8990,
        14454,
        253,
        20485,
        11,
        11070,
        1541,
        26327,
        11336,
        2289,
        19683,
        2412,
        14454,
        253,
        20485,
        11066,
        1546,
        2257,
        9636,
        97,
        10673,
        109,
        30900,
        16262,
        115,
        8082,
        8225,
        13486,
        1543,
        50815
      ],
      "temperature": 0,
      "avg_logprob": -0.07155802210823434,
      "compression_ratio": 1.1014492753623188,
      "no_speech_prob": 2.1357226751006486e-11,
      "source": "whisper"
    },
    {
      "id": 49,
      "seek": 29500,
      "start": 304.0,
      "end": 306.0,
      "text": "當然,這裡有個邊界得說清楚。",
      "tokens": [
        50815,
        21707,
        11,
        18771,
        2412,
        3338,
        14901,
        21185,
        5916,
        4622,
        38433,
        1543,
        50915
      ],
      "temperature": 0,
      "avg_logprob": -0.07155802210823434,
      "compression_ratio": 1.1014492753623188,
      "no_speech_prob": 2.1357226751006486e-11,
      "source": "whisper"
    },
    {
      "id": 50,
      "seek": 29500,
      "start": 306.0,
      "end": 313.0,
      "text": "Codex的代碼用Apache 2.0許可證開源,不代表OpenAI把GPT模型的權重也一併開源了。",
      "tokens": [
        50915,
        34,
        1429,
        87,
        1546,
        19105,
        16337,
        120,
        9254,
        32,
        79,
        6000,
        568,
        13,
        15,
        40914,
        4429,
        11823,
        231,
        8949,
        47402,
        11,
        1960,
        19105,
        17571,
        45569,
        48698,
        16075,
        38,
        47,
        51,
        41908,
        39823,
        1546,
        40212,
        12624,
        6404,
        2257,
        1593,
        113,
        8949,
        47402,
        2289,
        1543,
        51265
      ],
      "temperature": 0,
      "avg_logprob": -0.07155802210823434,
      "compression_ratio": 1.1014492753623188,
      "no_speech_prob": 2.1357226751006486e-11,
      "source": "whisper"
    },
    {
      "id": 51,
      "seek": 31300,
      "start": 313.0,
      "end": 319.0,
      "text": "開源Harness也不代表模型推理、算力、企業部署這些就變成零成本了。",
      "tokens": [
        50365,
        8949,
        47402,
        39,
        16937,
        6404,
        1960,
        19105,
        17571,
        41908,
        39823,
        33597,
        13876,
        1231,
        19497,
        13486,
        1231,
        7384,
        223,
        27119,
        13470,
        16469,
        110,
        29869,
        3111,
        26327,
        11336,
        6306,
        114,
        11336,
        8802,
        2289,
        1543,
        50665
      ],
      "temperature": 0,
      "avg_logprob": -0.07827498854660406,
      "compression_ratio": 1.1850828729281768,
      "no_speech_prob": 1.7243038594183524e-11,
      "source": "whisper"
    },
    {
      "id": 52,
      "seek": 31300,
      "start": 319.0,
      "end": 326.0,
      "text": "更不代表公司可以毫無安全措施地把數據庫、內部API、敏感文件全都甩給Agent去處理。",
      "tokens": [
        50665,
        19002,
        1960,
        19105,
        17571,
        13545,
        32981,
        6723,
        7256,
        104,
        16976,
        45466,
        6900,
        103,
        4307,
        121,
        10928,
        16075,
        30622,
        36841,
        6346,
        104,
        1231,
        28472,
        13470,
        4715,
        40,
        1231,
        7017,
        237,
        9709,
        17174,
        20485,
        11319,
        7182,
        3526,
        102,
        17798,
        32,
        6930,
        6734,
        39289,
        13876,
        1543,
        51015
      ],
      "temperature": 0,
      "avg_logprob": -0.07827498854660406,
      "compression_ratio": 1.1850828729281768,
      "no_speech_prob": 1.7243038594183524e-11,
      "source": "whisper"
    },
    {
      "id": 53,
      "seek": 31300,
      "start": 326.0,
      "end": 336.0,
      "text": "真正成熟的企業及Agent反而會越來越強調全線邊界、殺箱、審計、工具白名單,以及關鍵動作必須有人確認的Human in the Loop機制。",
      "tokens": [
        51015,
        6303,
        15789,
        11336,
        21805,
        253,
        1546,
        7384,
        223,
        27119,
        25703,
        32,
        6930,
        22138,
        11070,
        6236,
        25761,
        3763,
        25761,
        22752,
        32943,
        11319,
        31699,
        14901,
        21185,
        1231,
        45528,
        11249,
        109,
        1231,
        4510,
        102,
        30114,
        1231,
        23323,
        39806,
        13558,
        15940,
        24227,
        11,
        40282,
        14899,
        41668,
        113,
        12572,
        11914,
        28531,
        8313,
        230,
        45820,
        24293,
        22041,
        1546,
        39,
        6829,
        294,
        264,
        45660,
        17543,
        25491,
        1543,
        51515
      ],
      "temperature": 0,
      "avg_logprob": -0.07827498854660406,
      "compression_ratio": 1.1850828729281768,
      "no_speech_prob": 1.7243038594183524e-11,
      "source": "whisper"
    },
    {
      "id": 54,
      "seek": 31300,
      "start": 336.0,
      "end": 340.0,
      "text": "而這或許才是Codex Harness最值得關注的地方。",
      "tokens": [
        51515,
        11070,
        2664,
        19780,
        40914,
        18888,
        1541,
        34,
        1429,
        87,
        389,
        16937,
        8661,
        40242,
        5916,
        14899,
        26432,
        1546,
        30146,
        1543,
        51715
      ],
      "temperature": 0,
      "avg_logprob": -0.07827498854660406,
      "compression_ratio": 1.1850828729281768,
      "no_speech_prob": 1.7243038594183524e-11,
      "source": "whisper"
    },
    {
      "id": 55,
      "seek": 34000,
      "start": 340.0,
      "end": 344.0,
      "text": "過去幾年我們一直在問,下一代模型會有多聰明?",
      "tokens": [
        50365,
        36910,
        23575,
        5157,
        5884,
        34448,
        3581,
        11361,
        11,
        45271,
        19105,
        41908,
        39823,
        6236,
        2412,
        6392,
        8171,
        108,
        11100,
        30,
        50565
      ],
      "temperature": 0,
      "avg_logprob": -0.07011800823789655,
      "compression_ratio": 1.3079019073569482,
      "no_speech_prob": 2.187994056768172e-11,
      "source": "whisper"
    },
    {
      "id": 56,
      "seek": 34000,
      "start": 344.0,
      "end": 348.0,
      "text": "但接下來,軟件行業可能會越來越頻繁地面對另一個問題。",
      "tokens": [
        50565,
        8395,
        40012,
        11,
        14454,
        253,
        20485,
        8082,
        27119,
        16657,
        6236,
        25761,
        3763,
        25761,
        29940,
        23141,
        223,
        10928,
        8833,
        2855,
        22762,
        8990,
        17197,
        1543,
        50765
      ],
      "temperature": 0,
      "avg_logprob": -0.07011800823789655,
      "compression_ratio": 1.3079019073569482,
      "no_speech_prob": 2.187994056768172e-11,
      "source": "whisper"
    },
    {
      "id": 57,
      "seek": 34000,
      "start": 348.0,
      "end": 353.0,
      "text": "怎麼把這份聰明,安全、穩定、可控地裝進真實世界?",
      "tokens": [
        50765,
        11614,
        16075,
        2664,
        36266,
        8171,
        108,
        11100,
        11,
        45466,
        1231,
        13740,
        102,
        12088,
        1231,
        4429,
        48707,
        10928,
        44453,
        18214,
        6303,
        10376,
        24486,
        30,
        51015
      ],
      "temperature": 0,
      "avg_logprob": -0.07011800823789655,
      "compression_ratio": 1.3079019073569482,
      "no_speech_prob": 2.187994056768172e-11,
      "source": "whisper"
    },
    {
      "id": 58,
      "seek": 34000,
      "start": 353.0,
      "end": 356.0,
      "text": "從這個角度看,Harness的價值就很清楚了。",
      "tokens": [
        51015,
        21068,
        6287,
        29389,
        13127,
        4200,
        11,
        39,
        16937,
        1546,
        45441,
        40242,
        3111,
        4563,
        38433,
        2289,
        1543,
        51165
      ],
      "temperature": 0,
      "avg_logprob": -0.07011800823789655,
      "compression_ratio": 1.3079019073569482,
      "no_speech_prob": 2.187994056768172e-11,
      "source": "whisper"
    },
    {
      "id": 59,
      "seek": 34000,
      "start": 356.0,
      "end": 362.0,
      "text": "模型決定的是智能的上限,Harness決定的是這種智能能不能真正幹活。",
      "tokens": [
        51165,
        41908,
        39823,
        33540,
        12088,
        24620,
        5094,
        118,
        8225,
        1546,
        5708,
        43446,
        11,
        39,
        16937,
        33540,
        12088,
        24620,
        33005,
        5094,
        118,
        8225,
        8225,
        28590,
        6303,
        15789,
        29932,
        25956,
        1543,
        51465
      ],
      "temperature": 0,
      "avg_logprob": -0.07011800823789655,
      "compression_ratio": 1.3079019073569482,
      "no_speech_prob": 2.187994056768172e-11,
      "source": "whisper"
    },
    {
      "id": 60,
      "seek": 34000,
      "start": 362.0,
      "end": 369.64,
      "text": "而應用開發者真正的護城河也會重新回到自己的數據、工作流、專業界面、全線體系和行業理解上。",
      "tokens": [
        51465,
        11070,
        20481,
        9254,
        8949,
        14637,
        12444,
        6303,
        15789,
        1546,
        48289,
        45277,
        3308,
        111,
        6404,
        6236,
        12624,
        12560,
        8350,
        4511,
        17645,
        1546,
        30622,
        36841,
        1231,
        41315,
        27854,
        1231,
        46072,
        27119,
        21185,
        8833,
        1231,
        11319,
        31699,
        23987,
        25368,
        12565,
        8082,
        27119,
        13876,
        17278,
        5708,
        1543,
        51847
      ],
      "temperature": 0,
      "avg_logprob": -0.07011800823789655,
      "compression_ratio": 1.3079019073569482,
      "no_speech_prob": 2.187994056768172e-11,
      "source": "whisper"
    },
    {
      "id": 61,
      "seek": 37000,
      "start": 370.0,
      "end": 374.0,
      "text": "未來最強大的AI產品,可能未必天天提醒你這裡有個AI。",
      "tokens": [
        50365,
        29954,
        3763,
        8661,
        22752,
        39156,
        48698,
        33299,
        30246,
        11,
        16657,
        29954,
        28531,
        6135,
        6135,
        20949,
        40796,
        2166,
        18771,
        2412,
        3338,
        48698,
        1543,
        50565
      ],
      "temperature": 0,
      "avg_logprob": -0.05814369987039005,
      "compression_ratio": 1.1441048034934498,
      "no_speech_prob": 2.3517760655566988e-11,
      "source": "whisper"
    },
    {
      "id": 62,
      "seek": 37000,
      "start": 374.0,
      "end": 380.0,
      "text": "它可能沒有巨大的輸入框,沒有閃爍的機器人頭像,也不會要求你重新學一套操作方式。",
      "tokens": [
        50565,
        11284,
        16657,
        6963,
        5238,
        101,
        39156,
        12290,
        116,
        14028,
        21416,
        228,
        11,
        6963,
        5196,
        225,
        8164,
        235,
        1546,
        17543,
        34386,
        4035,
        19921,
        12760,
        11,
        6404,
        21121,
        4275,
        32718,
        2166,
        12624,
        12560,
        21372,
        2257,
        1881,
        245,
        13167,
        235,
        11914,
        9249,
        27584,
        1543,
        50865
      ],
      "temperature": 0,
      "avg_logprob": -0.05814369987039005,
      "compression_ratio": 1.1441048034934498,
      "no_speech_prob": 2.3517760655566988e-11,
      "source": "whisper"
    },
    {
      "id": 63,
      "seek": 37000,
      "start": 380.0,
      "end": 386.0,
      "text": "你打開的還是那個熟悉的財務軟件,運營看板,IDE,客服系統。",
      "tokens": [
        50865,
        2166,
        12467,
        8949,
        1546,
        25583,
        20754,
        21805,
        253,
        14696,
        231,
        1546,
        11561,
        94,
        39517,
        14454,
        253,
        20485,
        11,
        32772,
        24184,
        253,
        4200,
        43664,
        11,
        29283,
        11,
        32316,
        27408,
        25368,
        33725,
        1543,
        51165
      ],
      "temperature": 0,
      "avg_logprob": -0.05814369987039005,
      "compression_ratio": 1.1441048034934498,
      "no_speech_prob": 2.3517760655566988e-11,
      "source": "whisper"
    },
    {
      "id": 64,
      "seek": 38600,
      "start": 386.0,
      "end": 393.0,
      "text": "只是從某一天開始,這些軟件突然開始自己理解任務,自己調用工具,自己執行工作。",
      "tokens": [
        50365,
        36859,
        21068,
        17238,
        238,
        2257,
        6135,
        21017,
        11,
        29869,
        14454,
        253,
        20485,
        40859,
        5823,
        21017,
        17645,
        13876,
        17278,
        26443,
        39517,
        11,
        17645,
        32943,
        9254,
        23323,
        39806,
        11,
        17645,
        16262,
        115,
        8082,
        41315,
        1543,
        50715
      ],
      "temperature": 0,
      "avg_logprob": -0.07878539675757998,
      "compression_ratio": 1.1481481481481481,
      "no_speech_prob": 2.261777397316589e-11,
      "source": "whisper"
    },
    {
      "id": 65,
      "seek": 38600,
      "start": 393.0,
      "end": 397.966813,
      "text": "只在真正需要你的時候才停下來,問你一句,這一步要執行嗎?",
      "tokens": [
        50715,
        14003,
        3581,
        6303,
        15789,
        35748,
        18961,
        14010,
        18888,
        36135,
        27769,
        11,
        11361,
        2166,
        2257,
        34592,
        11,
        32260,
        31429,
        4275,
        16262,
        115,
        8082,
        7434,
        30,
        50965
      ],
      "temperature": 0,
      "avg_logprob": -0.07878539675757998,
      "compression_ratio": 1.1481481481481481,
      "no_speech_prob": 2.261777397316589e-11,
      "source": "whisper"
    }
  ],
  "transcript_selection": {
    "schema": "videonote.transcript_selection.v1",
    "status": "SELECTED",
    "requested_language": "zh",
    "duration": 397.966813,
    "reference_path": null,
    "reference_kind": null,
    "reference_language": null,
    "selected_source": "whisper_segmented",
    "decision_reason": "source_unavailable_or_untrusted_whisper_passed",
    "whisper_rejected": false,
    "source_profile": {
      "segment_count": 0,
      "duration_coverage": 0.0,
      "usable_full": false,
      "usable_partial": false,
      "hallucination_region_count": 0,
      "hallucination_regions": [],
      "language": {
        "requested": "zh",
        "matched": false
      }
    },
    "whisper_profile": {
      "segment_count": 66,
      "multilingual": false,
      "first_start": 0.0,
      "final_end": 397.967,
      "duration_coverage": 1.0,
      "maximum_gap_seconds": 0.36,
      "inverted_count": 0,
      "language": {
        "requested": "zh",
        "matched": true,
        "latin": 488,
        "han": 1830,
        "kana": 0,
        "hangul": 0
      },
      "hallucination_region_count": 0,
      "hallucination_regions": [],
      "usable_full": true,
      "usable_partial": true
    },
    "comparison": {
      "window_seconds": 30.0,
      "compared_window_count": 0,
      "median_score": 0.0,
      "divergent_window_count": 0,
      "divergent_ratio": 0.0,
      "divergent_windows": []
    },
    "timestamp_alignment": {
      "applied": false,
      "estimated_offset_seconds": 0.0,
      "matched_segment_count": 0,
      "match_scores": []
    },
    "caption_timing_repair": null,
    "filled_source_profile": null,
    "whisper_duration_guard": {
      "media_duration": 397.967,
      "rejected_segment_count": 0,
      "rejected_segments": []
    }
  }
}