{"text": "欢迎来到WorkBetty进阶蓝皮书的22章。这一章我们聊一个特别实用的是,怎么把一本书,一系列视频蒸留成Agent能主动调用的Skill。核心就三步,读书,蒸留,用起来。很多人都有这个困惑,书读了不少,真要用的时候却想不起来。恋爱也一样,他训练时读过很多经典,但回答常常是一堆正确的废话,每个字都对,就是没有可落地的步骤。知识经流要解决的,就是这种学了用不上的问题。知识经流,简单来说,就是把书或视频里的方法论,提纯成一个个代处发条件,代执行步骤的Skill。就像化学经流暗费点分离混合物,它按框架、原则、暗粒、反粒,数于五个维度,把知识分离成纯净组分,只把真正有用的提纯成可执行单元。Tanger Scale用六个阶段把一整本书蒸留成一套Skill,先整书理解,在五个Agent并行提取,然后三重验证筛选,接着构造Skill,建立链接,最后压力测试。以文案创作完全手册为例,整本书最终变成一套可以调用,可以测试的Skill集合。阶段0很关键,不从摘金句开始,而是先读清整本书的骨架,这一步决定了后面提取的质量上限。阶段1让五个Agent从五个维度并行扫描全文,各自独立工作,互不干扰,避免单线阅读时的视角遗漏。阶段1.5是三重验证筛选,每个候选单元必须过三关,没过就淘汰,宁缺无赖。一本书通常50到100个候选,最后只留10到25个。阶段2构造Skill,最核心的是设计触发条件,没有触发条件,Agent根本不知道什么时候该调用它。阶段4座链接,把Skill之间的关系连成知识网络,这样遇到复杂问题,Agent能选一组Skill而不是单个。阶段5是压力测试,用诱饵测试故意给不该触发的场景,看Skill能不能忍住不激活,再用真实问题验证它给的是不是可落地的步骤。征流完的产物是一套结构清晰的Skill集合,有Rimi说明,有每个Skill的独立文件,有记录关系的Index,还有待测试用力的Test目录。而且测试用力兼容Darwin Scale,可以自动进化,持续优化,分数只升不降。有人会问,这和Rack有什么区别?Rack解决的是知识管理,让你能查到书里有什么,知识经流解决的是知识运用,让Agent在对的时刻主动拿出对的框架。当你不知道该问什么的时候,Rack帮不了你。他也吸收了Carpacy的LLM wiki思路,但目标不同,两者并不忽斥。Tanger Scale第二版增加了视频蒸留,先用Video Downloader把视频下载,提取音频,转写成文字。再进入六阶段Sub。支持YouTube和B站,长视频,推荐用ASR API转写。多个同主题视频还能合并蒸留,自动取重,而视频获取和文本蒸留职责分离,各自演进。什么材料适合蒸留?方法论密度高的书最适合,五星,访谈和课程视频也不错,长视频播客可用,金句散文类就一般了,小说叙事文学基本不适合。还有一个潜质条件,蒸留前最好自己先读过一遍,蒸留是阅读后的结构化工具,不是替代阅读。知识经流是Token消耗密集型的,主要花在多Agent并行提取和验证上,建议用轻量模型做提取,强模型做验证和构造。好在产物可以直接分享复用,把GitHub仓库地址给Agent,它就能自动安装使用,社区里同一本书不必每个人重复蒸留。最后提醒几个误区,训练过的书也要蒸留,因为小众和新书,AI大概率没读过,而且蒸留的价值在触发条件,蒸留完还是要读书,它是补充不是替代。AI给建议不等于能直接执行,决策还是人的责任,覆盖也不是越广越好,边界要控住。以温达的AI入门课为例,26个视频就能蒸留成一套Skill集合。知识经流是Skill的一种生产方式和Seg封装并行,产物可以在同一个Agent框架下混合使用。", "segments": [{"id": 0, "seek": 0, "start": 0.0, "end": 15.46, "text": "欢迎来到WorkBetty进阶蓝皮书的22章。这一章我们聊一个特别实用的是,怎么把一本书,一系列视频蒸留成Agent能主动调用的Skill。核心就三步,读书,蒸留,用起来。", "tokens": [50365, 28566, 17699, 6912, 4511, 28846, 33, 38204, 36700, 10034, 114, 164, 26152, 34582, 2930, 99, 1546, 7490, 11957, 254, 1543, 5562, 2257, 11957, 254, 15003, 40096, 20182, 17682, 18453, 24726, 9254, 24620, 11, 15282, 16075, 2257, 8802, 2930, 99, 11, 2257, 25368, 43338, 40656, 39752, 42356, 116, 24456, 11336, 32, 6930, 8225, 13557, 34961, 8897, 225, 9254, 1546, 50, 34213, 1543, 8987, 116, 7945, 3111, 10960, 31429, 11, 5233, 119, 2930, 99, 11, 42356, 116, 24456, 11, 9254, 9147, 6912, 1543, 51138], "temperature": 0, "avg_logprob": -0.16121112599092371, "compression_ratio": 1.0456852791878173, "no_speech_prob": 9.546981261099319e-12}, {"id": 1, "seek": 1546, "start": 15.46, "end": 32.5, "text": "很多人都有这个困惑,书读了不少,真要用的时候却想不起来。恋爱也一样,他训练时读过很多经典,但回答常常是一堆正确的废话,每个字都对,就是没有可落地的步骤。", "tokens": [50365, 20778, 4035, 48121, 15368, 3919, 108, 4799, 239, 11, 2930, 99, 5233, 119, 2289, 1960, 15686, 11, 6303, 4275, 9254, 49873, 5322, 112, 7093, 26451, 6912, 1543, 11845, 233, 27324, 6404, 2257, 14496, 11, 5000, 7422, 255, 10115, 225, 15729, 5233, 119, 16866, 20778, 30276, 2347, 116, 11, 8395, 8350, 28223, 11279, 11279, 1541, 2257, 10726, 228, 15789, 38114, 106, 1546, 6346, 253, 21596, 11, 23664, 7549, 22381, 7182, 8713, 11, 5620, 17944, 4429, 30848, 10928, 1546, 31429, 49657, 97, 1543, 51217], "temperature": 0, "avg_logprob": -0.056854274294791965, "compression_ratio": 1.206896551724138, "no_speech_prob": 1.0178719846154483e-11}, {"id": 2, "seek": 1546, "start": 33.18, "end": 37.24, "text": "知识经流要解决的,就是这种学了用不上的问题。", "tokens": [51251, 6498, 5233, 228, 30276, 27854, 4275, 17278, 5676, 111, 1546, 11, 5620, 5562, 39810, 29618, 2289, 9254, 1960, 5708, 1546, 34069, 1543, 51454], "temperature": 0, "avg_logprob": -0.056854274294791965, "compression_ratio": 1.206896551724138, "no_speech_prob": 1.0178719846154483e-11}, {"id": 3, "seek": 3724, "start": 37.24, "end": 48.14, "text": "知识经流,简单来说,就是把书或视频里的方法论,提纯成一个个代处发条件,代执行步骤的Skill。", "tokens": [50365, 6498, 5233, 228, 30276, 27854, 11, 11249, 222, 47446, 6912, 8090, 11, 5620, 16075, 2930, 99, 19780, 40656, 39752, 15759, 1546, 9249, 11148, 7422, 118, 11, 20949, 16853, 107, 11336, 20182, 7549, 19105, 1787, 226, 28926, 48837, 20485, 11, 19105, 3416, 100, 8082, 31429, 49657, 97, 1546, 50, 34213, 1543, 50910], "temperature": 0, "avg_logprob": -0.08081281185150146, "compression_ratio": 1.2258064516129032, "no_speech_prob": 1.5207751724588547e-11}, {"id": 4, "seek": 3724, "start": 48.78, "end": 61.36, "text": "就像化学经流暗费点分离混合物,它按框架、原则、暗粒、反粒,数于五个维度,把知识分离成纯净组分,只把真正有用的提纯成可执行单元。", "tokens": [50942, 3111, 12760, 23756, 29618, 30276, 27854, 23826, 245, 18464, 117, 12579, 6627, 18614, 119, 48640, 14245, 23516, 11, 11284, 26613, 21416, 228, 7360, 114, 1231, 19683, 2437, 247, 1231, 23826, 245, 18393, 240, 1231, 22138, 18393, 240, 11, 33188, 37732, 21001, 7549, 10115, 112, 13127, 11, 16075, 6498, 5233, 228, 6627, 18614, 119, 11336, 16853, 107, 6336, 222, 10115, 226, 6627, 11, 14003, 16075, 6303, 15789, 2412, 9254, 1546, 20949, 16853, 107, 11336, 4429, 3416, 100, 8082, 47446, 14812, 1543, 51571], "temperature": 0, "avg_logprob": -0.08081281185150146, "compression_ratio": 1.2258064516129032, "no_speech_prob": 1.5207751724588547e-11}, {"id": 5, "seek": 6136, "start": 61.36, "end": 76.68, "text": "Tanger Scale用六个阶段把一整本书蒸留成一套Skill,先整书理解,在五个Agent并行提取,然后三重验证筛选,接着构造Skill,建立链接,最后压力测试。", "tokens": [50365, 51, 3176, 42999, 9254, 30566, 7549, 10034, 114, 28427, 16075, 2257, 27662, 8802, 2930, 99, 42356, 116, 24456, 11336, 2257, 1881, 245, 50, 34213, 11, 10108, 27662, 2930, 99, 13876, 17278, 11, 3581, 21001, 7549, 32, 6930, 3509, 114, 8082, 20949, 29436, 11, 26636, 10960, 12624, 49657, 234, 5233, 223, 7973, 249, 2215, 231, 11, 14468, 20708, 7360, 226, 37583, 50, 34213, 11, 34157, 24409, 165, 241, 122, 14468, 11, 8661, 13547, 5014, 233, 13486, 11038, 233, 5233, 243, 1543, 51131], "temperature": 0, "avg_logprob": -0.07737071812152863, "compression_ratio": 1.144, "no_speech_prob": 1.755056690255774e-11}, {"id": 6, "seek": 6136, "start": 77.34, "end": 84.62, "text": "以文案创作完全手册为例,整本书最终变成一套可以调用,可以测试的Skill集合。", "tokens": [51164, 3588, 17174, 28899, 2437, 249, 11914, 37100, 11389, 5676, 234, 13992, 17797, 11, 27662, 8802, 2930, 99, 8661, 10115, 230, 2129, 246, 11336, 2257, 1881, 245, 6723, 8897, 225, 9254, 11, 6723, 11038, 233, 5233, 243, 1546, 50, 34213, 26020, 14245, 1543, 51528], "temperature": 0, "avg_logprob": -0.07737071812152863, "compression_ratio": 1.144, "no_speech_prob": 1.755056690255774e-11}, {"id": 7, "seek": 8462, "start": 84.62, "end": 95.92, "text": "阶段0很关键,不从摘金句开始,而是先读清整本书的骨架,这一步决定了后面提取的质量上限。", "tokens": [50365, 10034, 114, 28427, 15, 4563, 28053, 23049, 106, 11, 1960, 35630, 34783, 246, 19117, 34592, 45213, 11, 11070, 1541, 10108, 5233, 119, 21784, 27662, 8802, 2930, 99, 1546, 165, 16838, 7360, 114, 11, 5562, 2257, 31429, 5676, 111, 12088, 2289, 13547, 8833, 20949, 29436, 1546, 18464, 101, 26748, 5708, 43446, 1543, 50930], "temperature": 0, "avg_logprob": -0.06276329221396611, "compression_ratio": 1.0786026200873362, "no_speech_prob": 1.8250219044335658e-11}, {"id": 8, "seek": 8462, "start": 96.58000000000001, "end": 105.34, "text": "阶段1让五个Agent从五个维度并行扫描全文,各自独立工作,互不干扰,避免单线阅读时的视角遗漏。", "tokens": [50963, 10034, 114, 28427, 16, 33650, 21001, 7549, 32, 6930, 35630, 21001, 7549, 10115, 112, 13127, 3509, 114, 8082, 3416, 104, 11673, 237, 11319, 17174, 11, 17516, 9722, 18637, 105, 24409, 41315, 11, 1369, 240, 1960, 26111, 3416, 108, 11, 3330, 123, 2347, 235, 47446, 16853, 123, 10034, 227, 5233, 119, 15729, 1546, 40656, 29389, 3330, 245, 14065, 237, 1543, 51401], "temperature": 0, "avg_logprob": -0.06276329221396611, "compression_ratio": 1.0786026200873362, "no_speech_prob": 1.8250219044335658e-11}, {"id": 9, "seek": 10534, "start": 105.34, "end": 115.64, "text": "阶段1.5是三重验证筛选,每个候选单元必须过三关,没过就淘汰,宁缺无赖。", "tokens": [50365, 10034, 114, 28427, 16, 13, 20, 1541, 10960, 12624, 49657, 234, 5233, 223, 7973, 249, 2215, 231, 11, 23664, 7549, 11238, 2215, 231, 47446, 14812, 28531, 10178, 119, 16866, 10960, 28053, 11, 10062, 16866, 3111, 14176, 246, 162, 2442, 11, 2415, 223, 38109, 118, 31364, 5266, 244, 1543, 50880], "temperature": 0, "avg_logprob": -0.07065573046284337, "compression_ratio": 1.1327800829875518, "no_speech_prob": 1.5568511757546588e-11}, {"id": 10, "seek": 10534, "start": 116.04, "end": 120.48, "text": "一本书通常50到100个候选,最后只留10到25个。", "tokens": [50900, 2257, 8802, 2930, 99, 19550, 11279, 2803, 4511, 6879, 7549, 11238, 2215, 231, 11, 8661, 13547, 14003, 24456, 3279, 4511, 6074, 7549, 1543, 51122], "temperature": 0, "avg_logprob": -0.07065573046284337, "compression_ratio": 1.1327800829875518, "no_speech_prob": 1.5568511757546588e-11}, {"id": 11, "seek": 10534, "start": 121.14, "end": 128.7, "text": "阶段2构造Skill,最核心的是设计触发条件,没有触发条件,Agent根本不知道什么时候该调用它。", "tokens": [51155, 10034, 114, 28427, 17, 7360, 226, 37583, 50, 34213, 11, 8661, 8987, 116, 7945, 24620, 7422, 122, 7422, 94, 6758, 99, 28926, 48837, 20485, 11, 17944, 6758, 99, 28926, 48837, 20485, 11, 32, 6930, 31337, 8802, 17572, 10440, 29111, 44646, 8897, 225, 9254, 11284, 1543, 51533], "temperature": 0, "avg_logprob": -0.07065573046284337, "compression_ratio": 1.1327800829875518, "no_speech_prob": 1.5568511757546588e-11}, {"id": 12, "seek": 12870, "start": 128.7, "end": 139.48, "text": "阶段4座链接,把Skill之间的关系连成知识网络,这样遇到复杂问题,Agent能选一组Skill而不是单个。", "tokens": [50365, 10034, 114, 28427, 19, 6346, 100, 165, 241, 122, 14468, 11, 16075, 50, 34213, 9574, 31685, 1546, 28053, 25368, 3316, 252, 11336, 6498, 5233, 228, 16469, 239, 10115, 250, 11, 21209, 3330, 229, 4511, 1787, 235, 4422, 224, 34069, 11, 32, 6930, 8225, 2215, 231, 2257, 10115, 226, 50, 34213, 11070, 7296, 47446, 7549, 1543, 50904], "temperature": 0, "avg_logprob": -0.061451290929040246, "compression_ratio": 1.156, "no_speech_prob": 1.6572224495470245e-11}, {"id": 13, "seek": 12870, "start": 140.11999999999998, "end": 150.82, "text": "阶段5是压力测试,用诱饵测试故意给不该触发的场景,看Skill能不能忍住不激活,再用真实问题验证它给的是不是可落地的步骤。", "tokens": [50936, 10034, 114, 28427, 20, 1541, 5014, 233, 13486, 11038, 233, 5233, 243, 11, 9254, 5233, 109, 41520, 113, 11038, 233, 5233, 243, 43045, 9042, 23197, 1960, 44646, 6758, 99, 28926, 1546, 50255, 50218, 11, 4200, 50, 34213, 8225, 28590, 4263, 235, 21632, 1960, 26373, 222, 25956, 11, 8623, 9254, 6303, 24726, 34069, 49657, 234, 5233, 223, 11284, 23197, 1546, 23034, 4429, 30848, 10928, 1546, 31429, 49657, 97, 1543, 51471], "temperature": 0, "avg_logprob": -0.061451290929040246, "compression_ratio": 1.156, "no_speech_prob": 1.6572224495470245e-11}, {"id": 14, "seek": 15082, "start": 150.82, "end": 164.6, "text": "征流完的产物是一套结构清晰的Skill集合,有Rimi说明,有每个Skill的独立文件,有记录关系的Index,还有待测试用力的Test目录。", "tokens": [50365, 2172, 223, 27854, 14128, 1546, 1369, 100, 23516, 1541, 2257, 1881, 245, 45641, 7360, 226, 21784, 5094, 108, 1546, 50, 34213, 26020, 14245, 11, 2412, 49, 10121, 8090, 11100, 11, 2412, 23664, 7549, 50, 34213, 1546, 18637, 105, 24409, 17174, 20485, 11, 2412, 34756, 7391, 243, 28053, 25368, 1546, 21790, 3121, 11, 35091, 18390, 11038, 233, 5233, 243, 9254, 13486, 1546, 51, 377, 11386, 7391, 243, 1543, 51054], "temperature": 0, "avg_logprob": -0.1157084511172387, "compression_ratio": 1.1076923076923078, "no_speech_prob": 1.812298054681971e-11}, {"id": 15, "seek": 15082, "start": 165.06, "end": 171.04, "text": "而且测试用力兼容Darwin Scale,可以自动进化,持续优化,分数只升不降。", "tokens": [51077, 22942, 11038, 233, 5233, 243, 9254, 13486, 2347, 120, 25750, 35, 289, 9136, 42999, 11, 6723, 9722, 34961, 36700, 23756, 11, 17694, 10115, 255, 7384, 246, 23756, 11, 6627, 33188, 14003, 41670, 1960, 47421, 1543, 51376], "temperature": 0, "avg_logprob": -0.1157084511172387, "compression_ratio": 1.1076923076923078, "no_speech_prob": 1.812298054681971e-11}, {"id": 16, "seek": 15082, "start": 173.57999999999998, "end": 176.35999999999999, "text": "有人会问,这和Rack有什么区别?", "tokens": [51503, 45820, 12949, 22064, 11, 5562, 12565, 49, 501, 2412, 10440, 9937, 118, 18453, 30, 51642], "temperature": 0, "avg_logprob": -0.1157084511172387, "compression_ratio": 1.1076923076923078, "no_speech_prob": 1.812298054681971e-11}, {"id": 17, "seek": 17636, "start": 176.36, "end": 186.44000000000003, "text": "Rack解决的是知识管理,让你能查到书里有什么,知识经流解决的是知识运用,让Agent在对的时刻主动拿出对的框架。", "tokens": [50365, 49, 501, 17278, 5676, 111, 24620, 6498, 5233, 228, 23131, 13876, 11, 33650, 2166, 8225, 42623, 4511, 2930, 99, 15759, 2412, 10440, 11, 6498, 5233, 228, 30276, 27854, 17278, 5676, 111, 24620, 6498, 5233, 228, 3316, 238, 9254, 11, 33650, 32, 6930, 3581, 8713, 1546, 15729, 45500, 13557, 34961, 24351, 7781, 8713, 1546, 21416, 228, 7360, 114, 1543, 50869], "temperature": 0, "avg_logprob": -0.09882366502439821, "compression_ratio": 1.2024539877300613, "no_speech_prob": 1.5380715798207767e-11}, {"id": 18, "seek": 17636, "start": 187.08, "end": 189.94000000000003, "text": "当你不知道该问什么的时候,Rack帮不了你。", "tokens": [50901, 16233, 2166, 17572, 44646, 22064, 10440, 49873, 11, 49, 501, 4845, 106, 47225, 2166, 1543, 51044], "temperature": 0, "avg_logprob": -0.09882366502439821, "compression_ratio": 1.2024539877300613, "no_speech_prob": 1.5380715798207767e-11}, {"id": 19, "seek": 17636, "start": 190.54000000000002, "end": 196.26000000000002, "text": "他也吸收了Carpacy的LLM wiki思路,但目标不同,两者并不忽斥。", "tokens": [51074, 5000, 6404, 46431, 18681, 2289, 34, 6529, 2551, 1546, 24010, 44, 261, 9850, 8870, 24658, 11, 8395, 11386, 162, 3921, 47123, 11, 36257, 12444, 3509, 114, 1960, 4263, 121, 4307, 98, 1543, 51360], "temperature": 0, "avg_logprob": -0.09882366502439821, "compression_ratio": 1.2024539877300613, "no_speech_prob": 1.5380715798207767e-11}, {"id": 20, "seek": 17636, "start": 199.20000000000002, "end": 206.34, "text": "Tanger Scale第二版增加了视频蒸留,先用Video Downloader把视频下载,提取音频,转写成文字。", "tokens": [51507, 51, 3176, 42999, 19693, 42096, 24228, 252, 9990, 2289, 40656, 39752, 42356, 116, 24456, 11, 10108, 9254, 46287, 32282, 260, 16075, 40656, 39752, 4438, 17819, 121, 11, 20949, 29436, 18034, 39752, 11, 17819, 105, 5676, 247, 11336, 17174, 22381, 1543, 51864], "temperature": 0, "avg_logprob": -0.09882366502439821, "compression_ratio": 1.2024539877300613, "no_speech_prob": 1.5380715798207767e-11}, {"id": 21, "seek": 20636, "start": 206.36, "end": 207.9, "text": "再进入六阶段Sub。", "tokens": [50365, 8623, 36700, 14028, 30566, 10034, 114, 28427, 39582, 1543, 50442], "temperature": 0, "avg_logprob": -0.1446645332105232, "compression_ratio": 1.08, "no_speech_prob": 1.7017862813384355e-11}, {"id": 22, "seek": 20636, "start": 208.5, "end": 212.98000000000002, "text": "支持YouTube和B站,长视频,推荐用ASR API转写。", "tokens": [50472, 35488, 3223, 41173, 12565, 33, 34155, 11, 32271, 40656, 39752, 11, 33597, 31409, 238, 9254, 3160, 49, 9362, 17819, 105, 5676, 247, 1543, 50696], "temperature": 0, "avg_logprob": -0.1446645332105232, "compression_ratio": 1.08, "no_speech_prob": 1.7017862813384355e-11}, {"id": 23, "seek": 20636, "start": 213.64000000000001, "end": 221.54000000000002, "text": "多个同主题视频还能合并蒸留,自动取重,而视频获取和文本蒸留职责分离,各自演进。", "tokens": [50729, 6392, 7549, 13089, 13557, 30716, 40656, 39752, 14852, 8225, 14245, 3509, 114, 42356, 116, 24456, 11, 9722, 34961, 29436, 12624, 11, 11070, 40656, 39752, 31127, 115, 29436, 12565, 17174, 8802, 42356, 116, 24456, 8171, 234, 18464, 96, 6627, 18614, 119, 11, 17516, 9722, 31382, 36700, 1543, 51124], "temperature": 0, "avg_logprob": -0.1446645332105232, "compression_ratio": 1.08, "no_speech_prob": 1.7017862813384355e-11}, {"id": 24, "seek": 20636, "start": 224.08, "end": 225.78000000000003, "text": "什么材料适合蒸留?", "tokens": [51251, 10440, 4422, 238, 33404, 2215, 224, 14245, 42356, 116, 24456, 30, 51336], "temperature": 0, "avg_logprob": -0.1446645332105232, "compression_ratio": 1.08, "no_speech_prob": 1.7017862813384355e-11}, {"id": 25, "seek": 22578, "start": 225.78, "end": 237.16, "text": "方法论密度高的书最适合,五星,访谈和课程视频也不错,长视频播客可用,金句散文类就一般了,小说叙事文学基本不适合。", "tokens": [50365, 9249, 11148, 7422, 118, 47838, 13127, 12979, 1546, 2930, 99, 8661, 2215, 224, 14245, 11, 21001, 20682, 11, 7422, 123, 8897, 230, 12565, 5233, 122, 29649, 40656, 39752, 6404, 1960, 29900, 11, 32271, 40656, 39752, 49993, 32316, 4429, 9254, 11, 19117, 34592, 7017, 96, 17174, 22113, 119, 3111, 2257, 49640, 2289, 11, 7322, 8090, 2129, 247, 6973, 17174, 29618, 37946, 1960, 2215, 224, 14245, 1543, 50934], "temperature": 0, "avg_logprob": -0.06223413372827955, "compression_ratio": 1.1931330472103003, "no_speech_prob": 1.732680145194454e-11}, {"id": 26, "seek": 22578, "start": 237.82, "end": 245.68, "text": "还有一个潜质条件,蒸留前最好自己先读过一遍,蒸留是阅读后的结构化工具,不是替代阅读。", "tokens": [50967, 35091, 20182, 35622, 250, 18464, 101, 48837, 20485, 11, 42356, 116, 24456, 8945, 8661, 2131, 17645, 10108, 5233, 119, 16866, 2257, 3330, 235, 11, 42356, 116, 24456, 1541, 10034, 227, 5233, 119, 13547, 1546, 45641, 7360, 226, 23756, 23323, 39806, 11, 7296, 9531, 123, 19105, 10034, 227, 5233, 119, 1543, 51360], "temperature": 0, "avg_logprob": -0.06223413372827955, "compression_ratio": 1.1931330472103003, "no_speech_prob": 1.732680145194454e-11}, {"id": 27, "seek": 24568, "start": 245.68, "end": 259.1, "text": "知识经流是Token消耗密集型的,主要花在多Agent并行提取和验证上,建议用轻量模型做提取,强模型做验证和构造。", "tokens": [50365, 6498, 5233, 228, 30276, 27854, 1541, 51, 8406, 28837, 4450, 245, 47838, 26020, 39823, 1546, 11, 13557, 4275, 20127, 3581, 6392, 32, 6930, 3509, 114, 8082, 20949, 29436, 12565, 49657, 234, 5233, 223, 5708, 11, 34157, 7422, 106, 9254, 17819, 119, 26748, 41908, 39823, 10907, 20949, 29436, 11, 5702, 118, 41908, 39823, 10907, 49657, 234, 5233, 223, 12565, 7360, 226, 37583, 1543, 51036], "temperature": 0, "avg_logprob": -0.06444469962533064, "compression_ratio": 1.1235521235521235, "no_speech_prob": 2.018993305874517e-11}, {"id": 28, "seek": 24568, "start": 259.76, "end": 269.74, "text": "好在产物可以直接分享复用,把GitHub仓库地址给Agent,它就能自动安装使用,社区里同一本书不必每个人重复蒸留。", "tokens": [51069, 2131, 3581, 1369, 100, 23516, 6723, 43297, 30855, 1787, 235, 9254, 11, 16075, 38, 270, 21150, 1550, 241, 6346, 241, 10928, 14872, 222, 23197, 32, 6930, 11, 11284, 3111, 8225, 9722, 34961, 16206, 8083, 227, 22982, 9254, 11, 27658, 9937, 118, 15759, 13089, 2257, 8802, 2930, 99, 1960, 28531, 23664, 7549, 4035, 12624, 1787, 235, 42356, 116, 24456, 1543, 51568], "temperature": 0, "avg_logprob": -0.06444469962533064, "compression_ratio": 1.1235521235521235, "no_speech_prob": 2.018993305874517e-11}, {"id": 29, "seek": 26974, "start": 269.74, "end": 284.90000000000003, "text": "最后提醒几个误区,训练过的书也要蒸留,因为小众和新书,AI大概率没读过,而且蒸留的价值在触发条件,蒸留完还是要读书,它是补充不是替代。", "tokens": [50365, 8661, 13547, 20949, 40796, 6336, 254, 7549, 5233, 107, 9937, 118, 11, 7422, 255, 10115, 225, 16866, 1546, 2930, 99, 6404, 4275, 42356, 116, 24456, 11, 34627, 7322, 7384, 245, 12565, 12560, 2930, 99, 11, 48698, 32044, 44866, 10062, 5233, 119, 16866, 11, 22942, 42356, 116, 24456, 1546, 1550, 115, 40242, 3581, 6758, 99, 28926, 48837, 20485, 11, 42356, 116, 24456, 14128, 45726, 4275, 5233, 119, 2930, 99, 11, 11284, 1541, 9890, 98, 2347, 227, 7296, 9531, 123, 19105, 1543, 51123], "temperature": 0, "avg_logprob": -0.06439382855485125, "compression_ratio": 1.2225913621262459, "no_speech_prob": 1.521894416045555e-11}, {"id": 30, "seek": 26974, "start": 285.22, "end": 292.04, "text": "AI给建议不等于能直接执行,决策还是人的责任,覆盖也不是越广越好,边界要控住。", "tokens": [51139, 48698, 23197, 34157, 7422, 106, 1960, 10187, 37732, 8225, 43297, 3416, 100, 8082, 11, 5676, 111, 7973, 244, 45726, 4035, 1546, 18464, 96, 26443, 11, 2888, 228, 5419, 244, 6404, 7296, 25761, 3509, 123, 25761, 2131, 11, 40503, 21185, 4275, 48707, 21632, 1543, 51480], "temperature": 0, "avg_logprob": -0.06439382855485125, "compression_ratio": 1.2225913621262459, "no_speech_prob": 1.521894416045555e-11}, {"id": 31, "seek": 26974, "start": 292.38, "end": 298.42, "text": "以温达的AI入门课为例,26个视频就能蒸留成一套Skill集合。", "tokens": [51497, 3588, 9592, 102, 9830, 122, 1546, 48698, 14028, 8259, 101, 5233, 122, 13992, 17797, 11, 10880, 7549, 40656, 39752, 3111, 8225, 42356, 116, 24456, 11336, 2257, 1881, 245, 50, 34213, 26020, 14245, 1543, 51799], "temperature": 0, "avg_logprob": -0.06439382855485125, "compression_ratio": 1.2225913621262459, "no_speech_prob": 1.521894416045555e-11}, {"id": 32, "seek": 29842, "start": 298.42, "end": 306.72, "text": "知识经流是Skill的一种生产方式和Seg封装并行,产物可以在同一个Agent框架下混合使用。", "tokens": [50365, 6498, 5233, 228, 30276, 27854, 1541, 50, 34213, 1546, 2257, 39810, 8244, 1369, 100, 9249, 27584, 12565, 50, 1146, 1530, 223, 8083, 227, 3509, 114, 8082, 11, 1369, 100, 23516, 6723, 3581, 13089, 20182, 32, 6930, 21416, 228, 7360, 114, 4438, 48640, 14245, 22982, 9254, 1543, 50780], "temperature": 0, "avg_logprob": -0.09474493532764669, "compression_ratio": 0.9112903225806451, "no_speech_prob": 1.369621603630744e-11}], "language": "zh"}