{"text": "Dipsic Flash 從4.0到4.1的進化程度堪比去年Cloud Ops 4.5到4.6就這麼誇張。第一它更快了,它的回復速度明顯提升了。第二它更強了,它是個小模型,但是能力超過了上一代的V4 Pro。我們看一下這個圖,可以看到在Terminal Bank 3.0里,它超過了KMI K3和GLM 5.3。要知道模型的主要工作環境就是在終端里,所以這項測試非常有意義。然後在其他三個測試里,它甚至超過了OPS 5和GPT 5.6 Soul。而且不僅如此,它還更便宜了。它的輸入價格相對於V4 Flash 降了60%。我們先看一下我測的任務,然後我再說一下這一波它為什麼提升這麼大。主要得益於它的新架構。這個是我用兩輪對話讓它做的一個之前很火的一個遊戲。可以看到它完成的很好,我現在是在控制人物在行動。可以看到這裡我還可以去攻擊敵人。OK,這裡就被Clear掉了。然後來到了第二關。整個效果我覺得非常非常棒。基本上沒有什麼Bug。然後我用一段提示詞讓它附現了一下遊戲王的卡片。當我們不斷旋轉卡片的時候,它這個流光也是在不斷的變化。然後下面還可以調這個輻射的強度。我們可以調的很高。可以把它調眉。調眉之後就只有這個光的反射。你可以看到當我把卡片旋轉到一定角度,是可以看見這個人物的一個輪廓的。這個細節還原的非常好。其實還有一些別的角色。整體效果也非常棒。然後這個主體也是可以縮放的。包括深度。我感覺用這個模型去做一些輕度任務、中度任務都非常非常好。而且它的生成的速度很快。好,我們詳細看一下這個模型這次用的一個新架構。它在這裡說的這個新架構有一些不容易懂。我用一個比較容易懂的邏輯來講一下。是這樣的。你就想像模型工作的時候,一般分為讀體和答體。這兩個步驟。第一個步驟,獨體的時候,每個字都要進入它所有的神經網絡。如果有1萬個字,這1萬個字都得從這40層裡面走一遍。而且每一層神經網絡走完之後,它不是走完就完了。而是它要把每一個字記住。把這1萬個字全部都走完記住完之後,模型才開始輸出文字。然後輸出的每一個文字也會走完這40層神經網絡。現在DeepSick這個新架構,就是在獨體的時候可以不用走完這40層神經網絡。它只要走完前20層就行了。那後面20層就直接把前20層的記錄。那後面20層就直接把前20層的記錄,直接翻譯成自己的就行。這樣的話,成本就直接少了一半。那為什麼這樣可以呢?因為後面的20層神經網絡從剛開始訓練的時候,它就是這麼訓練的。它從誕生起就幹了這個活。那為什麼不直接把39層神經網絡都這麼訓練,只留一層讀動就可以了。那首先是因為前20層每一層都把前一層的理解再加工一遍,到了第二層。那這份理解已經經過了20道工序,就已經夠深了。如果只有一層那還是太淺了,模型還是沒有把這個題目給內化。而且說實話,翻譯過來的內容和自己寫的相比肯定還是有損失。但實測下來性能和老做法是差不多的。微軟之前也驗證了這條路。Deepseek這次是放大了做,但報告自己也承認,在極端情況下的編輯也還沒有完全摸清。然後簡單總結一下為什麼獨體省一半這麼有意義。因為現在AI幹的活是讀的多寫的少。它一次可能讀幾萬字,但是它回復可能就幾百字。所以讀題省了一半,速度就會快一大截。然後官方說明的第二點是,它有更少的緩存也更省成本。那這一點是怎麼理解呢?就我們剛剛說到模型讀題就是記筆記嘛。那答題就是翻筆記。那這份筆記就叫做KVCatcher。那這份筆記佔的空間就要你付錢。Deepseek這次是把每個字的筆記縮到了上一代的四分之一。所以你就能聊更長的內容,花更少的錢。如果你接著上次的聊,筆記就會直接從硬盤裡拿出來,就不用重新讀一遍了。那這份筆記就縮到了八分之一。這就是這次Deepseek V4.1 Flash它更快又更便宜的一個底層邏輯。好,我是達斯海,我會一直分享一些有用有價值的AI內容。如果你覺得這期有用,那我們下次再見。", "segments": [{"id": 0, "seek": 0, "start": 0.0, "end": 29.84, "text": "Dipsic Flash 從4.0到4.1的進化程度堪比去年Cloud Ops 4.5到4.6就這麼誇張。第一它更快了,它的回復速度明顯提升了。第二它更強了,它是個小模型,但是能力超過了上一代的V4 Pro。我們看一下這個圖,可以看到在Terminal Bank 3.0里,它超過了KMI K3和GLM 5.3。要知道模型的主要工作環境就是在終端里,所以這項測試非常有意義。然後在其他三個測試里,它甚至超過了OPS 5和GPT 5.6 Soul。", "tokens": [50365, 35, 2600, 299, 20232, 220, 21068, 19, 13, 15, 4511, 19, 13, 16, 1546, 18214, 23756, 29649, 13127, 10726, 103, 11706, 6734, 5157, 32787, 422, 1878, 1017, 13, 20, 4511, 19, 13, 21, 3111, 21269, 3549, 229, 18889, 1543, 18049, 11284, 19002, 10251, 2289, 11, 45224, 8350, 2172, 102, 31217, 13127, 11100, 46605, 20949, 41670, 2289, 1543, 19693, 11284, 19002, 22752, 2289, 11, 11284, 1541, 3338, 7322, 41908, 39823, 11, 11189, 8225, 13486, 19869, 8816, 2289, 5708, 2257, 19105, 1546, 53, 19, 1705, 1543, 5884, 28324, 6287, 2523, 244, 11, 6723, 18032, 3581, 51, 966, 2071, 8915, 805, 13, 15, 15759, 11, 11284, 19869, 8816, 2289, 42, 13808, 591, 18, 12565, 19440, 44, 1025, 13, 18, 1543, 4275, 7758, 41908, 39823, 1546, 13557, 4275, 41315, 40063, 108, 47885, 5620, 3581, 29371, 11957, 107, 15759, 11, 7239, 2664, 8313, 227, 9592, 105, 22099, 14392, 2412, 9042, 37430, 1543, 10213, 3581, 9572, 5000, 10960, 3338, 9592, 105, 22099, 15759, 11, 11284, 17556, 20844, 19869, 8816, 2289, 46, 6273, 1025, 12565, 38, 47, 51, 1025, 13, 21, 13588, 1543, 51857], "temperature": 0, "avg_logprob": -0.17866528828938802, "compression_ratio": 1.225063938618926, "no_speech_prob": 7.703577359352565e-12}, {"id": 1, "seek": 2984, "start": 29.84, "end": 59.04, "text": "而且不僅如此,它還更便宜了。它的輸入價格相對於V4 Flash 降了60%。我們先看一下我測的任務,然後我再說一下這一波它為什麼提升這麼大。主要得益於它的新架構。這個是我用兩輪對話讓它做的一個之前很火的一個遊戲。可以看到它完成的很好,我現在是在控制人物在行動。可以看到這裡我還可以去攻擊敵人。OK,這裡就被Clear掉了。然後來到了第二關。", "tokens": [50365, 22942, 1960, 9502, 227, 48487, 11, 11284, 7824, 19002, 27364, 2415, 250, 2289, 1543, 45224, 12290, 116, 14028, 45441, 30921, 15106, 2855, 19488, 53, 19, 20232, 220, 47421, 2289, 4550, 4, 1543, 5884, 10108, 28324, 1654, 9592, 105, 1546, 26443, 39517, 11, 10213, 1654, 8623, 4622, 8861, 32260, 30806, 11284, 18987, 20949, 41670, 21269, 3582, 1543, 13557, 4275, 5916, 5419, 232, 19488, 45224, 12560, 7360, 114, 43362, 1543, 41427, 1654, 9254, 16313, 12290, 103, 2855, 11103, 21195, 11284, 10907, 1546, 8990, 32442, 4563, 26171, 1546, 8990, 43046, 1543, 6723, 18032, 11284, 41509, 1546, 23801, 11, 1654, 12648, 1541, 3581, 48707, 25491, 4035, 23516, 3581, 8082, 12572, 1543, 6723, 18032, 18771, 1654, 7824, 6723, 6734, 41530, 37813, 7017, 113, 4035, 1543, 9443, 11, 18771, 3111, 23238, 34, 5797, 29327, 2289, 1543, 10213, 3763, 21381, 19693, 14899, 1543, 51825], "temperature": 0, "avg_logprob": -0.13401677797166564, "compression_ratio": 1.2924791086350975, "no_speech_prob": 1.6506457659049012e-11}, {"id": 2, "seek": 5984, "start": 59.84, "end": 63.84, "text": "整個效果我覺得非常非常棒。基本上沒有什麼Bug。", "tokens": [50365, 27662, 3338, 43076, 9319, 17689, 14392, 14392, 23514, 1543, 37946, 5708, 6963, 7598, 33, 697, 1543, 50565], "temperature": 0, "avg_logprob": -0.16542005214561412, "compression_ratio": 1.4580645161290322, "no_speech_prob": 2.081025456124319e-11}, {"id": 3, "seek": 5984, "start": 63.84, "end": 67.84, "text": "然後我用一段提示詞讓它附現了一下遊戲王的卡片。", "tokens": [50565, 10213, 1654, 9254, 2257, 28427, 20949, 25696, 6716, 252, 21195, 11284, 8842, 226, 9581, 2289, 8861, 43046, 19940, 1546, 32681, 16668, 1543, 50765], "temperature": 0, "avg_logprob": -0.16542005214561412, "compression_ratio": 1.4580645161290322, "no_speech_prob": 2.081025456124319e-11}, {"id": 4, "seek": 5984, "start": 67.84, "end": 72.84, "text": "當我們不斷旋轉卡片的時候,它這個流光也是在不斷的變化。", "tokens": [50765, 13118, 5884, 1960, 4307, 115, 4479, 233, 32739, 32681, 16668, 20643, 11, 11284, 6287, 27854, 20690, 22021, 3581, 1960, 4307, 115, 1546, 26327, 23756, 1543, 51015], "temperature": 0, "avg_logprob": -0.16542005214561412, "compression_ratio": 1.4580645161290322, "no_speech_prob": 2.081025456124319e-11}, {"id": 5, "seek": 5984, "start": 72.84, "end": 76.84, "text": "然後下面還可以調這個輻射的強度。我們可以調的很高。", "tokens": [51015, 10213, 47150, 7824, 6723, 32943, 6287, 12290, 119, 1530, 226, 1546, 22752, 13127, 1543, 5884, 6723, 32943, 1546, 4563, 12979, 1543, 51215], "temperature": 0, "avg_logprob": -0.16542005214561412, "compression_ratio": 1.4580645161290322, "no_speech_prob": 2.081025456124319e-11}, {"id": 6, "seek": 5984, "start": 76.84, "end": 78.84, "text": "可以把它調眉。", "tokens": [51215, 6723, 42061, 32943, 2862, 231, 1543, 51315], "temperature": 0, "avg_logprob": -0.16542005214561412, "compression_ratio": 1.4580645161290322, "no_speech_prob": 2.081025456124319e-11}, {"id": 7, "seek": 5984, "start": 78.84, "end": 86.84, "text": "調眉之後就只有這個光的反射。你可以看到當我把卡片旋轉到一定角度,是可以看見這個人物的一個輪廓的。", "tokens": [51315, 32943, 2862, 231, 26362, 3111, 35244, 6287, 20690, 1546, 22138, 1530, 226, 1543, 42766, 18032, 13118, 1654, 16075, 32681, 16668, 4479, 233, 32739, 4511, 24272, 29389, 13127, 11, 1541, 6723, 4200, 10470, 6287, 4035, 23516, 1546, 8990, 12290, 103, 17527, 241, 1546, 1543, 51715], "temperature": 0, "avg_logprob": -0.16542005214561412, "compression_ratio": 1.4580645161290322, "no_speech_prob": 2.081025456124319e-11}, {"id": 8, "seek": 8684, "start": 86.84, "end": 95.84, "text": "這個細節還原的非常好。其實還有一些別的角色。整體效果也非常棒。然後這個主體也是可以縮放的。包括深度。", "tokens": [50365, 6287, 41588, 27694, 7824, 19683, 1546, 14392, 2131, 1543, 14139, 15569, 38515, 16158, 1546, 29389, 17673, 1543, 27662, 23987, 43076, 9319, 6404, 14392, 23514, 1543, 10213, 6287, 13557, 23987, 22021, 6723, 21030, 106, 12744, 1546, 1543, 41828, 24043, 13127, 1543, 50815], "temperature": 0, "avg_logprob": -0.10903177632913962, "compression_ratio": 1.5366568914956011, "no_speech_prob": 1.56489821101502e-11}, {"id": 9, "seek": 8684, "start": 95.84, "end": 106.84, "text": "我感覺用這個模型去做一些輕度任務、中度任務都非常非常好。而且它的生成的速度很快。好,我們詳細看一下這個模型這次用的一個新架構。", "tokens": [50815, 1654, 31742, 9254, 6287, 41908, 39823, 6734, 10907, 38515, 34801, 13127, 26443, 39517, 1231, 5975, 13127, 26443, 39517, 7182, 14392, 14392, 2131, 1543, 22942, 45224, 8244, 11336, 1546, 31217, 13127, 4563, 10251, 1543, 2131, 11, 5884, 6716, 111, 41588, 28324, 6287, 41908, 39823, 2664, 9487, 9254, 1546, 8990, 12560, 7360, 114, 43362, 1543, 51365], "temperature": 0, "avg_logprob": -0.10903177632913962, "compression_ratio": 1.5366568914956011, "no_speech_prob": 1.56489821101502e-11}, {"id": 10, "seek": 8684, "start": 106.84, "end": 114.84, "text": "它在這裡說的這個新架構有一些不容易懂。我用一個比較容易懂的邏輯來講一下。是這樣的。你就想像模型工作的時候,一般分為讀體和答體。", "tokens": [51365, 11284, 3581, 18771, 4622, 1546, 6287, 12560, 7360, 114, 43362, 32241, 13824, 1960, 49212, 29624, 1543, 1654, 9254, 8990, 25174, 49212, 29624, 1546, 3023, 237, 12290, 107, 3763, 11932, 8861, 1543, 1541, 39425, 1543, 41045, 7093, 12760, 41908, 39823, 41315, 20643, 11, 2257, 49640, 6627, 6344, 7422, 222, 23987, 12565, 28223, 23987, 1543, 51765], "temperature": 0, "avg_logprob": -0.10903177632913962, "compression_ratio": 1.5366568914956011, "no_speech_prob": 1.56489821101502e-11}, {"id": 11, "seek": 11484, "start": 114.84, "end": 123.84, "text": "這兩個步驟。第一個步驟,獨體的時候,每個字都要進入它所有的神經網絡。如果有1萬個字,這1萬個字都得從這40層裡面走一遍。", "tokens": [50365, 2664, 34623, 31429, 24023, 253, 1543, 40760, 31429, 24023, 253, 11, 163, 18304, 23987, 20643, 11, 23664, 3338, 22381, 7182, 4275, 18214, 14028, 11284, 39300, 1546, 20432, 15145, 26352, 6948, 94, 1543, 13119, 2412, 16, 14099, 3338, 22381, 11, 2664, 16, 14099, 3338, 22381, 7182, 5916, 21068, 2664, 5254, 9636, 97, 32399, 9575, 2257, 3330, 235, 1543, 50815], "temperature": 0, "avg_logprob": -0.11866713033139127, "compression_ratio": 1.648876404494382, "no_speech_prob": 2.1370837391399e-11}, {"id": 12, "seek": 11484, "start": 123.84, "end": 132.84, "text": "而且每一層神經網絡走完之後,它不是走完就完了。而是它要把每一個字記住。把這1萬個字全部都走完記住完之後,模型才開始輸出文字。", "tokens": [50815, 22942, 23664, 2257, 9636, 97, 20432, 15145, 26352, 6948, 94, 9575, 14128, 26362, 11, 11284, 7296, 9575, 14128, 3111, 41665, 1543, 11070, 1541, 11284, 4275, 16075, 23664, 8990, 22381, 16958, 21632, 1543, 16075, 2664, 16, 14099, 3338, 22381, 38714, 7182, 9575, 14128, 16958, 21632, 14128, 26362, 11, 41908, 39823, 18888, 21017, 12290, 116, 7781, 17174, 22381, 1543, 51265], "temperature": 0, "avg_logprob": -0.11866713033139127, "compression_ratio": 1.648876404494382, "no_speech_prob": 2.1370837391399e-11}, {"id": 13, "seek": 11484, "start": 132.84, "end": 142.84, "text": "然後輸出的每一個文字也會走完這40層神經網絡。現在DeepSick這個新架構,就是在獨體的時候可以不用走完這40層神經網絡。它只要走完前20層就行了。那後面20層就直接把前20層的記錄。", "tokens": [51265, 10213, 12290, 116, 7781, 1546, 23664, 8990, 17174, 22381, 6404, 6236, 9575, 14128, 2664, 5254, 9636, 97, 20432, 15145, 26352, 6948, 94, 1543, 12648, 11089, 595, 50, 618, 6287, 12560, 7360, 114, 43362, 11, 5620, 3581, 163, 18304, 23987, 20643, 6723, 24384, 9575, 14128, 2664, 5254, 9636, 97, 20432, 15145, 26352, 6948, 94, 1543, 11284, 14003, 4275, 9575, 14128, 8945, 2009, 9636, 97, 3111, 46606, 1543, 4184, 5661, 8833, 2009, 9636, 97, 3111, 43297, 16075, 8945, 2009, 9636, 97, 1546, 16958, 9567, 226, 1543, 51765], "temperature": 0, "avg_logprob": -0.11866713033139127, "compression_ratio": 1.648876404494382, "no_speech_prob": 2.1370837391399e-11}, {"id": 14, "seek": 14284, "start": 142.84, "end": 148.84, "text": "那後面20層就直接把前20層的記錄,直接翻譯成自己的就行。這樣的話,成本就直接少了一半。", "tokens": [50365, 4184, 5661, 8833, 2009, 9636, 97, 3111, 43297, 16075, 8945, 2009, 9636, 97, 1546, 16958, 9567, 226, 11, 43297, 42716, 11823, 107, 11336, 17645, 1546, 3111, 8082, 1543, 8377, 21358, 11, 11336, 8802, 3111, 43297, 15686, 2289, 2257, 30018, 1543, 50665], "temperature": 0, "avg_logprob": -0.08766661315667824, "compression_ratio": 1.421259842519685, "no_speech_prob": 1.8690360023554398e-11}, {"id": 15, "seek": 14284, "start": 148.84, "end": 159.84, "text": "那為什麼這樣可以呢?因為後面的20層神經網絡從剛開始訓練的時候,它就是這麼訓練的。它從誕生起就幹了這個活。那為什麼不直接把39層神經網絡都這麼訓練,只留一層讀動就可以了。", "tokens": [50665, 4184, 18987, 8377, 6723, 6240, 30, 11471, 5661, 8833, 1546, 2009, 9636, 97, 20432, 15145, 26352, 6948, 94, 21068, 16940, 21017, 5396, 241, 47027, 20643, 11, 11284, 5620, 21269, 5396, 241, 47027, 1546, 1543, 11284, 21068, 3549, 243, 8244, 9147, 3111, 29932, 2289, 6287, 25956, 1543, 4184, 18987, 1960, 43297, 16075, 12493, 9636, 97, 20432, 15145, 26352, 6948, 94, 7182, 21269, 5396, 241, 47027, 11, 14003, 24456, 2257, 9636, 97, 7422, 222, 12572, 36781, 2289, 1543, 51215], "temperature": 0, "avg_logprob": -0.08766661315667824, "compression_ratio": 1.421259842519685, "no_speech_prob": 1.8690360023554398e-11}, {"id": 16, "seek": 15984, "start": 159.84, "end": 172.84, "text": "那首先是因為前20層每一層都把前一層的理解再加工一遍,到了第二層。那這份理解已經經過了20道工序,就已經夠深了。如果只有一層那還是太淺了,模型還是沒有把這個題目給內化。", "tokens": [50365, 4184, 36490, 1541, 11471, 8945, 2009, 9636, 97, 23664, 2257, 9636, 97, 7182, 16075, 8945, 2257, 9636, 97, 1546, 13876, 17278, 8623, 9990, 23323, 2257, 3330, 235, 11, 21381, 19693, 9636, 97, 1543, 4184, 2664, 36266, 13876, 17278, 25338, 15145, 8816, 2289, 2009, 6025, 23323, 6346, 237, 11, 3111, 25338, 31649, 24043, 2289, 1543, 13119, 14003, 32241, 9636, 97, 4184, 25583, 9455, 14176, 118, 2289, 11, 41908, 39823, 25583, 6963, 16075, 6287, 14475, 11386, 17798, 28472, 23756, 1543, 51015], "temperature": 0, "avg_logprob": -0.09646058664089296, "compression_ratio": 1.2526315789473683, "no_speech_prob": 1.697774386355544e-11}, {"id": 17, "seek": 17284, "start": 172.84, "end": 179.84, "text": "而且說實話,翻譯過來的內容和自己寫的相比肯定還是有損失。但實測下來性能和老做法是差不多的。", "tokens": [50365, 22942, 4622, 10376, 11103, 11, 42716, 11823, 107, 49950, 1546, 28472, 25750, 12565, 17645, 4510, 104, 1546, 15106, 11706, 14356, 107, 12088, 25583, 2412, 22220, 235, 28564, 1543, 8395, 10376, 9592, 105, 27769, 21686, 8225, 12565, 10439, 10907, 11148, 1541, 37876, 1546, 1543, 50715], "temperature": 0, "avg_logprob": -0.11810719966888428, "compression_ratio": 1.3, "no_speech_prob": 1.3422880862368203e-11}, {"id": 18, "seek": 17284, "start": 179.84, "end": 186.84, "text": "微軟之前也驗證了這條路。Deepseek這次是放大了做,但報告自己也承認,在極端情況下的編輯也還沒有完全摸清。", "tokens": [50715, 39152, 14454, 253, 32442, 6404, 24023, 245, 11823, 231, 2289, 2664, 31106, 24658, 1543, 11089, 595, 405, 916, 2664, 9487, 1541, 12744, 3582, 2289, 10907, 11, 8395, 17803, 16846, 17645, 6404, 3416, 123, 22041, 11, 3581, 13387, 113, 11957, 107, 39391, 4438, 1546, 12608, 101, 12290, 107, 6404, 7824, 6963, 37100, 34783, 116, 21784, 1543, 51065], "temperature": 0, "avg_logprob": -0.11810719966888428, "compression_ratio": 1.3, "no_speech_prob": 1.3422880862368203e-11}, {"id": 19, "seek": 17284, "start": 186.84, "end": 194.84, "text": "然後簡單總結一下為什麼獨體省一半這麼有意義。因為現在AI幹的活是讀的多寫的少。它一次可能讀幾萬字,但是它回復可能就幾百字。", "tokens": [51065, 10213, 31995, 26575, 17144, 8861, 18987, 163, 18304, 23987, 2862, 223, 2257, 30018, 21269, 2412, 9042, 37430, 1543, 11471, 12648, 48698, 29932, 1546, 25956, 1541, 7422, 222, 1546, 6392, 4510, 104, 1546, 15686, 1543, 11284, 27505, 16657, 7422, 222, 23575, 14099, 22381, 11, 11189, 11284, 8350, 2172, 102, 16657, 3111, 23575, 31906, 22381, 1543, 51465], "temperature": 0, "avg_logprob": -0.11810719966888428, "compression_ratio": 1.3, "no_speech_prob": 1.3422880862368203e-11}, {"id": 20, "seek": 19484, "start": 194.84, "end": 202.84, "text": "所以讀題省了一半,速度就會快一大截。然後官方說明的第二點是,它有更少的緩存也更省成本。那這一點是怎麼理解呢?", "tokens": [50365, 7239, 7422, 222, 14475, 2862, 223, 2289, 2257, 30018, 11, 31217, 13127, 47778, 10251, 2257, 3582, 1486, 103, 1543, 10213, 31929, 9249, 4622, 11100, 1546, 19693, 8216, 1541, 11, 11284, 2412, 19002, 15686, 1546, 12608, 102, 39781, 6404, 19002, 2862, 223, 11336, 8802, 1543, 4184, 2664, 19202, 1541, 11614, 13876, 17278, 6240, 30, 50765], "temperature": 0, "avg_logprob": -0.13158842051251335, "compression_ratio": 1.3495440729483283, "no_speech_prob": 1.6272023659058554e-11}, {"id": 21, "seek": 19484, "start": 202.84, "end": 212.84, "text": "就我們剛剛說到模型讀題就是記筆記嘛。那答題就是翻筆記。那這份筆記就叫做KVCatcher。那這份筆記佔的空間就要你付錢。", "tokens": [50765, 3111, 5884, 32795, 4622, 4511, 41908, 39823, 7422, 222, 14475, 5620, 16958, 7973, 228, 16958, 20722, 1543, 4184, 28223, 14475, 5620, 42716, 7973, 228, 16958, 1543, 4184, 2664, 36266, 7973, 228, 16958, 3111, 19855, 10907, 42, 53, 34, 49871, 1543, 4184, 2664, 36266, 7973, 228, 16958, 1593, 242, 1546, 23322, 11016, 3111, 4275, 2166, 41469, 29483, 1543, 51265], "temperature": 0, "avg_logprob": -0.13158842051251335, "compression_ratio": 1.3495440729483283, "no_speech_prob": 1.6272023659058554e-11}, {"id": 22, "seek": 19484, "start": 212.84, "end": 219.84, "text": "Deepseek這次是把每個字的筆記縮到了上一代的四分之一。所以你就能聊更長的內容,花更少的錢。", "tokens": [51265, 11089, 595, 405, 916, 2664, 9487, 1541, 16075, 23664, 3338, 22381, 1546, 7973, 228, 16958, 21030, 106, 21381, 5708, 2257, 19105, 1546, 19425, 6627, 9574, 2257, 1543, 7239, 41045, 8225, 40096, 19002, 15353, 1546, 28472, 25750, 11, 20127, 19002, 15686, 1546, 29483, 1543, 51615], "temperature": 0, "avg_logprob": -0.13158842051251335, "compression_ratio": 1.3495440729483283, "no_speech_prob": 1.6272023659058554e-11}, {"id": 23, "seek": 21984, "start": 219.84, "end": 226.84, "text": "如果你接著上次的聊,筆記就會直接從硬盤裡拿出來,就不用重新讀一遍了。那這份筆記就縮到了八分之一。", "tokens": [50365, 45669, 14468, 19382, 5708, 9487, 1546, 40096, 11, 7973, 228, 16958, 47778, 43297, 21068, 38114, 105, 5419, 97, 11066, 24351, 29741, 11, 3111, 24384, 12624, 12560, 7422, 222, 2257, 3330, 235, 2289, 1543, 4184, 2664, 36266, 7973, 228, 16958, 3111, 21030, 106, 21381, 33453, 6627, 9574, 2257, 1543, 50715], "temperature": 0, "avg_logprob": -0.09424851521724413, "compression_ratio": 1.178082191780822, "no_speech_prob": 1.4521391034083564e-11}, {"id": 24, "seek": 21984, "start": 226.84, "end": 232.84, "text": "這就是這次Deepseek V4.1 Flash它更快又更便宜的一個底層邏輯。", "tokens": [50715, 2664, 5620, 2664, 9487, 11089, 595, 405, 916, 691, 19, 13, 16, 20232, 11284, 19002, 10251, 17047, 19002, 27364, 2415, 250, 1546, 8990, 22816, 9636, 97, 3023, 237, 12290, 107, 1543, 51015], "temperature": 0, "avg_logprob": -0.09424851521724413, "compression_ratio": 1.178082191780822, "no_speech_prob": 1.4521391034083564e-11}, {"id": 25, "seek": 21984, "start": 232.84, "end": 238.84, "text": "好,我是達斯海,我會一直分享一些有用有價值的AI內容。如果你覺得這期有用,那我們下次再見。", "tokens": [51015, 2131, 11, 15914, 33086, 30758, 20078, 11, 1654, 6236, 34448, 30855, 38515, 2412, 9254, 2412, 45441, 40242, 1546, 48698, 28472, 25750, 1543, 45669, 11274, 2664, 16786, 2412, 9254, 11, 46714, 35829, 30177, 1543, 51315], "temperature": 0, "avg_logprob": -0.09424851521724413, "compression_ratio": 1.178082191780822, "no_speech_prob": 1.4521391034083564e-11}], "language": "zh"}