{ "text": "大家好欢迎来到灵姐说AI\n这期视频啊\n给大家介绍一个\n大家可能比较熟悉\n听得很多\n但是未必深刻理解\n知道怎么去应用的一个\n模型的训练方法以及概念：蒸馏\n我希望通过这期视频\n你能够真正的理解蒸馏的本质\n它有哪些类型\n并且如何应用在我们\n实际的生产力的场景中\n我们先从最近闹得沸沸扬扬的\nAnthropic发布的蒸馏报告说起啊\n这个报告他们发布在9月10号\n他声称7家中国的AI大模型的实验室\n这些公司啊\n对他们进行了深度的蒸馏\n实际上Anthropic这家公司啊\n对于蒸馏这个事情非常的重视\n也非常的紧张\n他们在2月份的时候就\n披露了相关的数据\n当时他们专门点名了\n三家中国的AI模型公司\nDeepseek、月之暗面、Minimax\n进行了1,600多万次的交互蒸馏\n这一次涉及的公司更多\n覆盖的账户更多\n涉及的交互次数更多\n而且非常的深度\n都是工业化级别的蒸馏\n其中最其中最夸张的\n规模最大的\n大家猜猜是哪家公司\n是阿里千问\n根据Anthropic的报告说啊\n在26年的5-7月之间\n他观察到了阿里千问超过\n1.51亿次的Claude的数据交换和蒸馏\n也就是说\n它的峰值每天达到了300万次\n其余的几家公司包括了\n月之暗面、Deepseek、智谱、小米、商汤和Minimax\nMinimax\n而且这一次Anthropic的\n蒸馏报告特别指出啊\n这些中国的AI实验室不是简单的说\n拿Claude的回答做训练数据那么简单\n他们是建立了工业化的流水线\n去大规模的获取claude\n模型的推理轨迹\nagent能力\n编程能力\n数据分析能力和长程任务的能力\n再把这些能力蒸馏\n进自己的模型里面\n因为这些AI实验室啊\n他们要的不仅是答案\n更想要整个的推理过程\n而Claude方面\n它在模型的构建过程中\n它会刻意的去隐藏这些COT的推理链\n所以这个攻防的过程啊\n也变得特别有意思\n根据Anthropic描述他们的攻击方式啊\n我觉得也很有意思\n其实也非常具有参考价值\n这些模型公司啊\n他们想尽办法把Claude\n隐藏的这个reasoning trace\n这个推理的过程给挖出来\n比如说有人就会把这个\n会话伪装成一个debug的过程\n让它交出推理链\n或者说有人尝试说让Claude把\n之前他做的这个working memory\n提一个任务说\n你把这个内容翻译成\n只包含片假名的日语\n诱骗它把这个推理过程交出来\n而且更夸张的是啊\n有的AI实验室\n它直接用了1.2万种不同\n的请求方式去实验\n怎么把它的隐藏的\n推理过程给套路出来\n它的目的就是测试\n看看哪一种提示词\n哪一种路径\n能够把能够把Claude的\n推理路径给套出来\n有可能大部分的\n提示词方式都可能失败\n但是只要有几种成功了\n就可以用工业化的方式把它放大\n作为蒸馏的一种有效的路径\n所以这实际上是\n一个很有意思的攻防\n模型厂商呢\n他说我不给你原始的COT\n我只给你总结的thinking\n但蒸馏方就一直在研究\n怎么通过上下文翻译\nReplay的方式等等\n把这个推理的过程给恢复过来\n然后再把这个reasoning推理\n的过程变成训练数据\n你看看这有不有一点像AI\n时代的能力逆向工程\n实际上啊\n这个报告本身披露出了\n非常多类似这些AI模型公司\n他们蒸馏的方式啊\n而这些蒸馏的类型啊\n我会给大家在后面去给做总结\n并且啊总结出的方式我们\n可以进一步的拓展\n把它用于我们日常的生产力环节\n好准备好了吗\n接下来我就带着大家\n全面的去理解蒸馏\n蒸馏啊不要把它仅仅的理解为\n一种模型训练的技术\n我认为蒸馏的本质是\n一次昂贵的智能过程\n它实际上是在压缩成下一次更便宜、更快、更稳定\n一次更便宜更快更稳定\n甚至不需要还原老师的\n能力的这么一个过程\nAnthropic它给我们看的\n实际上是蒸馏中的一种\n就是模型蒸馏模型\n所以我会先跟大家去讲这个\n模型蒸馏模型有哪些过程\n有哪些类型\n而沿着这个定义往外推\n我们可以看到更多的东西\n除了模型蒸馏模型之外\n模型还可以蒸馏用户\n产品也能够蒸馏模型\nagent能蒸馏工作流\n企业能蒸馏员工\n甚至是人用AI来蒸馏自己\n而这些就是我们广阔的应用场景\n回到Anthropic这个报告来看啊\n它里面讲了非常多的细节\n是有助于我们去\n理解模型是怎么蒸馏\n传统的我们对蒸馏的理解\n实际上是特别简单直接的\n就是教师模型根据输入给出输出\n然后学生模型学习\n整个的信息交换的过程\n学习这些exchange的过程\n但是大家注意到这里面的\n环节会涉及到inputs和OUTPUTS\n在这个报告里面\n它实际上把这个inputs和outputs怎么来\n做了非常详细的解析和案例的披露\n也就是这些中国的AI实验室\n他们是怎么干的\n这些东西啊\n非常的有启发\n可以分为7个类型\n我总结下来\n一个个跟大家去讲\n第一种类型啊最传统\n我总结为叫出题蒸馏\n也可以叫做答案蒸馏吧\n就是学生给老师出题\n让老师回答\n形成一个高质量的答案\n你可以准备一个任务集\n这个任务集呢\n可以是coding数学agent数据分析\n或者长程任务等等都可以\n然后把这个问题给到Claude\n让Claude来回答\n形成高质量的答案或者推理\n然后把这个答案抄下来\n放进你的data set里面\n接着通过SFT监督微调\n或者是通过RL强化学习\n来训练自己的模型\n在这个报告里面啊\nAnthropic声称阿里千问的蒸馏\n模式就是最典型的例子\n它主要是针对Opus 4.6和4.7的推理轨迹\n它的整个的蒸馏任务是集中在agent\n软件工程\n还有一些长程任务上面\n并且它把整个的\nTranscript转成了SFT的数据\n这种呢就是考试考卷式的蒸馏\n这是最传统的蒸馏\n也是最容易理解的\n老师做1万道题\n学生直接拿老师给的答案进行训练\n第二种类型我总结为推理蒸馏\n就是学生不仅是要答案\n还要老师的解题过程\n这个答案确实有价值\n但是更值钱的是教师\n模型它是怎么拆解问题\n怎么判断\n怎么试错\n调用工具验证\n得出最终答案的过程\n要知其然\n更要知其所以然\n实际上它不仅是要老师授之以鱼\n更要授之以渔\n这个方式也是anthropic\n这次最在意的一种方式\n刚刚我提到了\n这些攻击者测试啊\n里面花了1.2万种方式\n通过翻译\ndebug replay等等方式\n试图去恢复Claude模型的推理\n的过程和找到其中的线索\n以后啊这个COT推理的过程会\n被藏的越来越深\n那么蒸馏怎么办呢\n其实就是不断的攻防攻防\n互相套路\n越狱嘛反正是一直去试\n而这些前沿模型也会越来越不\n愿意把原始的这个COT的链路啊\n公布出来\n展示出来\n以后可能会有一个很火的角色\n比如说像AI狱警这样\n第三个类型我总结为Judge蒸馏\n也就是让前沿模型\n让老师来当评委\n比如说你让一个普通模型\n比如说国产模型\n让它对一个问题生成a b c d四个答案\n这个时候啊\n你可以不用前沿模型来\n生成一个优质或者标准答案\n而是让这个前沿模型Claude\n来对这四个模型的答案进行打分\n他来评分 1-10分\n他打几分\n哪个答案好\n哪个答案不好\n为什么好\n为什么答错\n让他给出评价\n这时候链路就变成了什么\n普通模型生成答案\n然后Claude作为评委进行评分\n然后对所有的答案进行排序\n接着可以进行强化学习\n然后进行模型训练\n让普通的模型变强\n这个在Anthropic的这个报告里面\n关于智谱的这个描述\n和归因部分也有体现\n智谱呢就用这个Claude的模型\n让它生成一些评委评分\n的一些结果的一个标准\n然后还用它来做\n一些数据清洗的角色\n并且让它提供一些解决方案\n这个时候就不是纯粹的\n给答案对比这件事情了\n它让这个前沿模型在它的整个\n模型训练中扮演了更多的角色\n这个前沿的模型不仅\n是当老师给解题过程\n而是它也可以是出题的老师\n也可以是阅卷评分的老师\n它还可以是数据的清洗员\n还可以是安全的检查员\n还可以是最后的强化\n学习的这个评估员\n如果是经常看我节目\n的老的粉丝和观众\n一定感受到我平时在\n节目中经常用这个方法\n比如说我经常用多个\n模型来处理同一个任务\n然后我再用一个前沿模型\n比如说用GPT的Codex来对\n这些结果进行评分\n然后评价这几个模型\n的能力和最后的答案\n然后给出我的评测结论\n这本质上不就是使用judge的方式\n让这个前沿模型扮演\n一个评委的角色吗\n你看看从蒸馏的角度来\n理解我节目中的评测的方法\n是不是又升维了\n第四种蒸馏方式啊\n第四种蒸馏方式啊我总结叫做Replay蒸馏\n什么意思\n就是不用自己出题\n直接拿真实用户的问题\n重新让老师再做一遍\n你知道玩这个路子的\n是哪家模型公司吗\n是小米的模型MIMO他们家玩的\n就是假设啊\n他们家公司已经有100万的\n真实开发者的这个conversation\n这个对话\n这个实际上是非常宝贵的数据啊\n大家不要觉得自己很多\n地方可以免费用一些模型啊\n实际上都是这些模型公司的养料\n这里面会包含一些真的bug\n真的项目真的IDE\n真的文件\n真的用户意图\n真的失败案例\n原来在小米的模型里面\n它就会有原始的用户\n任务和小米给出的答案\n那么重新Replay一遍呢\n就是把同一个任务给到Claude\n然后再获得Claude的答案\n然后就立刻得到了什么呢\n首先是一个task\n用户的真实任务\n然后是student\nMIMO的答案\n然后是teacher\nClaude的答案\n接下来就可以干什么\n就可以比较\n可以judge\n可以回写\n可以SFT可以RL了\n小米把自己代码相关\n的coding session给到claude\n然后呢Claude还可以被用于\n重构它的开发环境\n再清洗对话\n然后再生成回答\n再生成响应评价的答案\n这可比人工去造Benchmark厉害多了\n经常我给大家演示的\n考虑到隐私啊\n我都是造的这些数据\n而这些数据都是用户的真实意图\n而用户本身就会成为源源不断\n的Benchmark的生产者和生成者\n第五种蒸馏方式啊我不说\n可能很多人还蒙在鼓里\n我叫做shadow teacher\n影子老师\n这是一个水加茅台的故事\n怎么说呢\n就是本来你只是要了一杯水\n你花了一杯水的钱\n结果这个水里面掺了茅台\n怎么理解啊\n就是用户以为自己用的是学生模型\n结果\n后台偷偷的让老师回答\n然后用户被告知的\n是学生模型的回答\n有不有毁你之前的认知\n就是你给他一个问题\n然后后台呢判断\n诶这个问题非常有价值啊\n这条数据很值钱啊\n然后呢\n他就把这个问题的请求转给了Claude\n让Claude来回答\n然后把Claude的答案推送给用户\n与此同时\n他在后台把这个提示词\n加上Claude的回答\n加上他的推理过程\n全部都被写入了他的训练的数据集\n根据Anthropic的报告啊\n月之暗面和Deepseek都在这样干\n那怎么发现的呢\n因为Claude它在清洗的时候发现\n这里面出现了非常多涉及\n企业和涉及一些特殊行业\n领域的一些机密的数据\n出现在了模型的对话里面\n而这些模型的对话本不\n应该出现在海外模型里面的\n这里面就有两个问题\n本来你花了一个更便宜的钱\n却买到了一个更贵的服务\n另外一方面呢\n就是本来你觉得你用的是国产模型\n不涉及到一些安全性的问题\n结果模型悄悄地转包到了海外模型\n把你的一些企业内部的信息\n泄露到了不应该泄露的地方\n涉及到信息安全的问题了\n但是要知道啊\n这样的方式本身对国产\n模型的训练是非常有帮助的\n因为很多这样的问题\n它是在特定的\n比如说一个国产的IDE里面去产生的\n它本来是它原始的Harness的架构\n那在这个Harness里面\n它本身是自带任务\n还带上下文\n又带文件\n带工具调用\n还带模型的尝试\n带错误带用户反馈\n然后现在再返回一个最终的结果\n它的信息密度是非常高的\n比单纯的那种聊天记录\n的信息密度高非常多\n所以通过这样的方式转包\n获得了前沿模型的回答\n然后作为训练集\n对于模型的成长是非常有帮助的\n所以你也会感知到为什么几家\n大的公司都在拼命推他们的IDE\n推他们的agent\n第六种蒸馏方式啊\n第六种蒸馏方式啊我叫做中转站蒸馏\n哎很多人没听错啊\n你是不是经常发现很多中转站\n它使用一些前沿模型的价格\n比你直接到官方那里使用要便宜\n为什么\n这个商业逻辑想清楚了没有啊\n他赚的可能根本不是\n什么TOKEN间的差价\n而是这个数据的产业链\n他最后啊会把这个对话的\n过程直接卖给其他的AI实验室\n就是自己不用产生数据\n直接去中转站把\n上下文把数据买过来\n买别人的Claude对话\n把这些对话拿来训练自己的模型\n因为在这些中转站上面啊\n天然就有提示词\n上下文和Claude的回答\n如果说你去保存\n成千上百的整个的会话过程\n那么就会形成非常\n有价值的数据资产\n甚至你知道的一些中转站\n它就是大厂下水自己干的\n只不过外面包了一个壳而已\n第七种方式啊\n第七种方式啊我定义为Pipeline蒸馏\n就是数据工厂蒸馏\n怎么理解啊\n原来呢我们蒸馏的方式是让这个\n教师模型教会学生模型\n这种方式啊\n非常的隐秘\n但是很高级\n现在他直接让这个教师构建出学校\n然后这个学校就可以\n源源不断的产生教师\n培养更多的学生了\n就是让这个老师自己出题\n自己回答\n自己批改\n你已经不需要给一个具体的题库了\n你直接给Claude一个具体的领域\n让它基于这个领域来进行自我出题\n让它出10万道题\n然后自己来回答\n让学生模型来完成整个任务\n然后Claude继续对这个任务的\n结果进行批改审阅评分\n然后可以纠错\n筛选出优质的回答和过程\n通过SFT和RL进行训练\n这个时候\n前沿的模型的角色不仅是老师\n他是整个学校\n出题老师是他\n任课老师是他\n批卷老师是他\n教研组组长是他\n教材的编写也是他\n他帮你设计的是整个流程\n他来构建整个框架\n上面我讲的7种蒸馏方式啊\n就是我从Anthropic的这个\n报告里面抽象总结出来的\n模型蒸馏模型的方式啊\n从某种意义上面来讲啊\n我觉得蒸馏也是AI时代技术\n扩散的一种非常重要的方式\n以前啊我们训练模型啊\n是人类写数据AI学\n或者AI生成数据AI学\n现在呢通过蒸馏的方式\nAI模型之间啊\n它们互相学习的程度已经到\n了一个非常深的一个程度\n在Anthropic的角度啊\n他确实觉得应该反对这种蒸馏\n而且他会找出各种\n方式来对抗这种蒸馏\n但是站在模型研究的角度\n蒸馏本身又是一种非常正常\n非常重要的技术\n连Anthropic自己都不得不承认这一点\n所以啊这一次\nClaude发现自己的智商正在被轻易\n的工业化的批量复制的时候\n又不禁引起我们反思\nAI行业它真正的护城河在哪里\n而原来我们在聊天沟通中所\n喜欢看到的这个COT的推理链\n是不是就从用户的一个体验功能\n变成了模型护城河的\n核心资产和安全边界了呢\n这个问题啊\n我觉得很复杂\n但是关于蒸馏这个事情啊\n我还想再深入拓展一下\n我们不要只停留在模型蒸馏模型\n实际上它有更大的\n应用价值和技术机制\n我这里讲一个暴论啊\n叫做一切皆可蒸馏\n等会我讲完蒸馏的整个拓展\n你再回过头来看这句话\n你会觉得特别的正确\n而且特别的有用\n有实用价值\n我们先看第一层拓展\n叫做用户的蒸馏\n你想过没有\n我们和AI对话交互的整个过程\n实际上就是模型蒸馏人的过程\n大家现在用Codex啊额度完全不够\n额度完全不够\n但是你想过没有\n你和GPT的聊天的额度和Codex\n执行任务的额度是分开的\n而且你使用GPT聊天的这个额度啊\n是很少达到上限的\n就是他给你的额度非常多\n为什么大部分的模型公司都会给你\n开放免费的和AI交互的这个额度啊\n实际上是这些模型公司在蒸馏你\n举个例子啊\n你想得到某个答案\n解决某个问题\n你会跟AI可能对话20轮\n最后拿到你想要的结果\n在这20轮对话里面\n你会做什么呢\n你会示范\n告诉他什么是对的\n给他举例子\n你会取舍\n告诉他a要b不要\n你会纠偏\n告诉他这个方向不对\n而这个过程\n其实你就是给了他非常多的输入\n最后得到满意的答案\n这些东西都是给到AI模型的语料\n而下一次优化训练之后\n模型就可以很快的识别你的意图\n它不需要你再过20轮\n它通过一到两轮的对话\n就可以实现满意的答案\n这也是为什么顶尖的大模型\n用户多的大模型\n它会有更好更多的数据集\n来提升它本身的模型能力\n大家现在用GPT的模型\n大家现在用GPT的模型从GPT4到GPT4o\n大家是不是有明显的感知\n模型越来越聪明\n需要的聊天的次数和轮次越来越少\n它能越来越知道你的心中所想\n怎么出来的\n就是通过模型蒸馏人\n的过程来逐步实现的\n我们以为我们在使用AI实际上\n从训练系统的视角来看\n人的每一次纠错\n实际上都可能成为什么是\n最好答案的一个监督信号\n既然模型能够蒸馏人\n那么我们是不是能进一步拓展\n模型公司也能够蒸馏它的员工呢\n比如说大厂\n像前段时间爆出来Meta\n它会记录它所有员工\n的键盘的各种操作\n来更大范围的炼化员工的能力\n去实现对人的能力的蒸馏\n这个实际上也是模型\n对人蒸馏的一种方式\n那我们更进一步\n我们可以主动的去实现这个过程\n比如说我有大量的节目制作\n的过程语料纠偏等等内容\n都在我的本机\n在我的新版的GPT的对话里面\n还有很多本地的材料\n还存在我本地的Obsidian里面\n那在这个里面我就可以实现让模型\n对我的所有的语料过程进行蒸馏\n不仅是我这样的自媒体博主\n你日常的工作如果都实现了\n在线化和颗粒度的对齐和AI\n模型的对齐和纠偏的话\n你都可以尝试用前沿模型\n对你本身进行自我蒸馏\n让你的整体的能力再上一个台阶\n举个例子啊\n大家现在看到的是我Codex的对话框\n我正在对我自己的生活和工作场景\n特别是我视频制作的过程进行蒸馏\n这里面就会涉及到我的选题和脚本\n我的封面和标题\n客户的方案等等等等\n每一个环节都是可以\n蒸馏和抽象出来的\n就以选题和脚本为例\n那么AI会给我输出一个脚本\n我会返回给他告诉他\n跟他对齐颗粒度\n在这个脚本里面\n哪些内容我觉得不顺滑\n哪些部分你讲得太少\n哪些地方你可以加一些金句等等\n这个实际上就是我\n给他的投喂和纠偏\n那他就可以去对比他输出的答案\n和我给的一个调整示范的内容\n通过这样的方式\n原来我可能要聊三四轮四五轮\n到后面我可能只需要聊一两轮\n就可以完成我的选题和脚本的确认\n通过这样的方式\n我可以沉淀出来我自己的skill\n或者是通过某种工具的调用\n去让整个的过程更加的丝滑\n这某种程度上也是一种蒸馏\n因为这里的蒸馏它\n不只限于一期节目\n我的100多期\n200多期节目\n整个的制作过程如果都是\n在本地在Codex完成了整个过程\n再加上我本地的知识库obsidian\n那我的蒸馏过程就能够把\n我的整体能力去加强放大\n而这里只是其中的一个环节\n我还可以拓展到我的封面和标题\n之前我可能要不停的跟\n他讲我需要什么风格\n那形成一定的能力沉淀之后\n他可以很快的通过一轮\n对话就完成我想要的封面\n在这里的蒸馏示范里面呢\n我就选择了3期节目\n我告诉他具体的节目内容和索引\n因为这三期节目我是在整个的\n对话框中给了他明确的调整链\n它是有明确的我的调整\n和AI输出的前后对比的\n基于这个样本\n它给我做了第一轮的蒸馏\n你看已经完成了\n它这里是蒸馏的结果\n这里呢\n它会有明确的修改链和蒸馏发现\n它会识别我改了哪些\n然后在这里啊\n它是帮我直接沉淀了一个方法包\n那这个方法包是可以复用的\n相当于他把我的工作能力习惯\n判断这些有价值的东西抽象出来了\n你想想是不是\n这个事情在公司里面的员工的操作\n其实也是这个逻辑\n这个员工他要审核pass\n或者不通过\n那他判断的依据标准是什么\n就可以通过这样的方式蒸馏出来\n他为什么这么判断\n而蒸馏出来的其实\n就是人的能力和智慧\n我们再往前思考一步啊\n我们还可以做垂直专业领域的蒸馏\n而且越是垂直\n越是专业的领域啊\n小模型对大模型的蒸馏\n的替代的表现就会越好\n举个例子啊\n假如你有一家咨询公司\n你这个公司呢\n本身就有很多的典型任务\n再加上比如说10年的文档\n再加上专家的经验\n这个时候你可以做什么呢\n你就可以把你的企业\n的历史数据拿出来\n再加上Claude它的前沿的能力\n让Claude加上企业数据\n去完成200个任务\n这200个任务最好是有代表性的\n能够覆盖这个垂直\n领域的大部分内容\nClaude完成之后呢\n它就会形成一个答案和回复\n然后呢在这个咨询公司里会有专家\n就让这些专家基于克劳德答案进行纠偏\n克劳德答案进行纠偏\n而这些专家的纠偏\n就是非常有价值的输入\nClaude就可以根据这些内容去总结\n什么情况下应该这样做\n什么情况下不能这么做\n什么是对的\n什么是错的\n什么是优秀答案\n什么是理想答案\n然后在这个领域\n你就形成了你的\n一个比较黄金的data set\n形成了一个数据集\n形成了知识库\n然后你就可以用国产模型小\n模型或者一些agent来跑了\n这个也是一种非常\n深度的新型的蒸馏\n就是你蒸馏的不纯粹\n的是Claude本身的能力\n你蒸馏的是\n企业数据加上Claude能力\n再加上人类的专家判断\n三者融合之后的能力\n而这个价值本身\n它比Claude直接蒸馏的意义是更大的\n它是在垂直领域非常专业的\n像这些应用啊\n就可以用到很多具体的领域\n包括写作设计法律咨询\n企业等等内部的专业领域\n其实在这个领域啊\n会有非常多的现实案例\n可以实现这样的智能蒸馏\n我给大家讲两个具体的场景\n相信对大家有启发\n你会知道怎么去做蒸馏\n怎么去提升自己的生产力\n第一个场景\n是一个设计师的场景\n怎么让AI蒸馏出这个\n设计师的审美和洞察\n那首先呢\n我给这个AI100张这个设计师经手的\n确认的认可的100张设计图\n然后让AI来蒸馏这些图\n的生成的逻辑和方式\n他们做的好的地方\n为什么要这样做等等\n蒸馏完成之后\n然后按照这个逻辑\n来生成一系列的图\n然后把生成好的图\n再丢给这个设计师\n让设计师来挑选确认\n并且最好是标注一下\n为什么这张图ok\n为什么这张图不行\n那人确认的这个挑选结果\n的反馈本身就是一种蒸馏对不对\n然后AI继续优化逻辑\n继续生成新的图\n然后这个逻辑又不断的上升了\n它生图的整个的逻辑又升级了\n又不断的有人反馈\n那这个人的反馈\n人的纠偏\n越到后面会变得越来越少\n你会发现在这个过程\n人AI它实际上是在双向蒸馏的\n有了这样的蒸馏思路之后啊\n甚至我们都能够基于\n结果来反向的蒸馏\n就以大家现在看到的这个文创\n产品协同的设计方案为例啊\n这是我给到AI的一个完整的方案\n我把这个PPT给到它\n我让它给我反向的抽象\n蒸馏出这个文创产品方案\n它形成的生产路径是什么\n我可以以果为因来反向蒸馏过程\n是不是就和蒸馏的逆向\n工程的过程是一样的\n这里是我蒸馏的初步的结果\n我实际上是可以继续的迭代去做的\n然后再加一下一些人的反馈\n那么以后我就可以\n通过这样的逆向工程\n把它蒸馏出来整个的生产过程\n然后再反向的推回去\n我可以基于它的推导\n帮我实际的生成验证\n然后我再不断的反馈和迭代\n最后这个整个的方案的生产过程\n是不是就可以被我推导\n然后蒸馏出来了\n再说一个比较实用的现实的场景\n比如电商公司要做投放\n你就可以让AI来生成\n一系列的投放计划\n然后这个投放计划你可以\n交给专门的投手去审核\n这个投手就会判断\n哪些计划是可行的\n哪些计划是不可行的\n对于公司来说\n你可以要求这些投手\n标注它可行不可行的原因\n合理不合理的原因\n人做出反馈\n标注好之后再给到AI去\n调整这些投放计划而且\n这些AI整合出来的投放计划\n最后投出去\n它会有数据反馈的\n最后投出来的效果好不好\n人还可以去做评价\n去说明对应的原因和反馈\n然后AI就可以根据人的意见\n以及投放的数据反馈再\n进一步生成新的投放组合\n在这个过程中\n投手的意见相当于是专家的意见\n是不是就被蒸馏了\nAI对投放这个事情的理解\n是不是也被蒸馏了\n而在这个过程中\n可以不断的用一些高等级的模型\n对标准答案的回答进行沉淀\n告诉哪些是好的答案\n哪些是坏的答案\n它的整个的data set就这样沉淀下来了\n这样就可以不断的\n无限循环和loop下去\n实际上这个过程再\n拓散一下是什么呢\n很多高级的模型\n也是这么训练出来的\n我们再发散一步\n是不是可以继续做AI来\n蒸馏AI组织的协作方式\n在未来agent可能越来越成熟了\n那么一个很重要的问题可能\n是怎么把一个优秀的agent的\n协作方式把它蒸馏出来\n比如说在一个协作体系里面\n它有计划的agent\n有研究的agent\n有写代码的agent\n有复盘的agent\n有做测试的agent\n还有审阅的agent\n把整个的团队方式运行10万次\n然后去发现什么任务它是\n需要planner计划这个角色的\n什么时候需要reviewer复盘者介入\n什么时候应该让贵的模型介入\n什么时候可以选择\n性价比的模型就可以了\n最后蒸馏出来的是一种agent组织\n协同的一种方式规则和框架\n再回过头来看我们开始的那句话\n蒸馏一切\n是不是挺对的\n我们可以把蒸馏一切啊\n分为5个层级\n整理成一张非常漂亮的图谱\n这是一条非常优美的演化\n从蒸馏答案获得大模型的结果\n再到蒸馏思考获得推理链\n再接着可以蒸馏行动\n再深入一点\n我们可以蒸馏垂直的工作方法\n形成skill\n形成skill、SOP、prompt、memory\n进一步的\n我们对对象\n对人和组织进行蒸馏\n形成我们的agent system\n蒸馏的本质\n实际上是把一次昂贵的智能过程\n变成下一次相对廉价可以复用的\n能力所以啊\n我们万事万物皆可蒸馏\nClaude的智能可以蒸馏\nCodex的computer use的执行过程可以蒸馏\n一个顶级的专家的认知可以蒸馏\n一套组织能力可以蒸馏\n我觉得这才是蒸馏让普通人\n能够产生无限想象力的地方\n蒸馏你搞懂了吗\n还有哪些概念或者训练\n方法你希望去深入的了解\n希望跟我探讨的\n欢迎在评论区跟我留言\n如果觉得我的这个视频有趣有料\n喜欢我的视频\n也欢迎订阅我的频道\n灵姐说AI\n打开你的小铃铛\n我们下期节目再见啦拜拜", "language": "zh", "duration": 2090.0, "segments": [ { "id": 0, "start": 0.1, "end": 2.233, "text": "大家好欢迎来到灵姐说AI", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 1, "start": 2.533, "end": 3.4, "text": "这期视频啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 2, "start": 3.4, "end": 4.6, "text": "给大家介绍一个", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 3, "start": 4.633, "end": 6.2, "text": "大家可能比较熟悉", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 4, "start": 6.2, "end": 7.2, "text": "听得很多", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 5, "start": 7.3, "end": 9.733, "text": "但是未必深刻理解", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 6, "start": 9.733, "end": 11.866, "text": "知道怎么去应用的一个", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 7, "start": 11.866, "end": 14.733, "text": "模型的训练方法以及概念：蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "模型的训练方法以及概念蒸馏", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 8, "start": 14.766, "end": 16.466, "text": "我希望通过这期视频", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 9, "start": 16.533, "end": 20.166, "text": "你能够真正的理解蒸馏的本质", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 10, "start": 20.433, "end": 21.6, "text": "它有哪些类型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 11, "start": 21.633, "end": 23.666, "text": "并且如何应用在我们", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 12, "start": 23.7, "end": 26.2, "text": "实际的生产力的场景中", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 13, "start": 26.233, "end": 29.7, "text": "我们先从最近闹得沸沸扬扬的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 14, "start": 29.766, "end": 32.933, "text": "Anthropic发布的蒸馏报告说起啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 15, "start": 33.0, "end": 35.633, "text": "这个报告他们发布在9月10号", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 16, "start": 35.633, "end": 40.266, "text": "他声称7家中国的AI大模型的实验室", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 17, "start": 40.666, "end": 41.533, "text": "这些公司啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 18, "start": 41.566, "end": 44.1, "text": "对他们进行了深度的蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 19, "start": 44.333, "end": 46.5, "text": "实际上Anthropic这家公司啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 20, "start": 46.566, "end": 48.8, "text": "对于蒸馏这个事情非常的重视", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 21, "start": 48.833, "end": 49.933, "text": "也非常的紧张", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 22, "start": 50.066, "end": 51.733, "text": "他们在2月份的时候就", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 23, "start": 51.766, "end": 53.2, "text": "披露了相关的数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 24, "start": 53.233, "end": 55.5, "text": "当时他们专门点名了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 25, "start": 55.533, "end": 57.733, "text": "三家中国的AI模型公司", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 26, "start": 57.9, "end": 60.266, "text": "Deepseek、月之暗面、Minimax", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "Deepseek月之暗面Minimax", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 27, "start": 60.5, "end": 64.533, "text": "进行了1,600多万次的交互蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 28, "start": 64.666, "end": 67.266, "text": "这一次涉及的公司更多", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 29, "start": 67.4, "end": 69.2, "text": "覆盖的账户更多", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 30, "start": 69.233, "end": 71.433, "text": "涉及的交互次数更多", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 31, "start": 71.433, "end": 73.033, "text": "而且非常的深度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 32, "start": 73.2, "end": 75.566, "text": "都是工业化级别的蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 33, "start": 75.833, "end": 77.333, "text": "其中最其中最夸张的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 34, "start": 77.333, "end": 78.333, "text": "规模最大的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 35, "start": 78.333, "end": 80.266, "text": "大家猜猜是哪家公司", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 36, "start": 80.466, "end": 82.166, "text": "是阿里千问", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 37, "start": 82.166, "end": 84.066, "text": "根据Anthropic的报告说啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 38, "start": 84.066, "end": 86.733, "text": "在26年的5-7月之间", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 39, "start": 86.866, "end": 90.066, "text": "他观察到了阿里千问超过", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 40, "start": 90.066, "end": 94.233, "text": "1.51亿次的Claude的数据交换和蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 41, "start": 94.6, "end": 95.433, "text": "也就是说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 42, "start": 95.433, "end": 98.833, "text": "它的峰值每天达到了300万次", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 43, "start": 98.866, "end": 100.766, "text": "其余的几家公司包括了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 44, "start": 100.766, "end": 104.333, "text": "月之暗面、Deepseek、智谱、小米、商汤和Minimax", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "月之暗面Deepseek智谱小米商汤和", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 45, "start": 104.366, "end": 105.0, "text": "Minimax", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 46, "start": 105.033, "end": 106.566, "text": "而且这一次Anthropic的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 47, "start": 106.566, "end": 108.233, "text": "蒸馏报告特别指出啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 48, "start": 108.366, "end": 111.4, "text": "这些中国的AI实验室不是简单的说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 49, "start": 111.433, "end": 114.433, "text": "拿Claude的回答做训练数据那么简单", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 50, "start": 114.633, "end": 117.666, "text": "他们是建立了工业化的流水线", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 51, "start": 117.966, "end": 120.166, "text": "去大规模的获取claude", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 52, "start": 120.233, "end": 121.833, "text": "模型的推理轨迹", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 53, "start": 122.1, "end": 122.966, "text": "agent能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 54, "start": 123.033, "end": 123.866, "text": "编程能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 55, "start": 123.9, "end": 127.366, "text": "数据分析能力和长程任务的能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 56, "start": 127.366, "end": 129.566, "text": "再把这些能力蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 57, "start": 129.566, "end": 131.266, "text": "进自己的模型里面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 58, "start": 131.266, "end": 133.033, "text": "因为这些AI实验室啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 59, "start": 133.033, "end": 134.6, "text": "他们要的不仅是答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 60, "start": 134.8, "end": 136.966, "text": "更想要整个的推理过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 61, "start": 137.033, "end": 137.966, "text": "而Claude方面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 62, "start": 137.966, "end": 139.8, "text": "它在模型的构建过程中", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 63, "start": 139.9, "end": 143.3, "text": "它会刻意的去隐藏这些COT的推理链", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 64, "start": 143.333, "end": 145.133, "text": "所以这个攻防的过程啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 65, "start": 145.133, "end": 146.633, "text": "也变得特别有意思", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 66, "start": 146.633, "end": 149.733, "text": "根据Anthropic描述他们的攻击方式啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 67, "start": 149.733, "end": 150.8, "text": "我觉得也很有意思", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 68, "start": 150.8, "end": 153.4, "text": "其实也非常具有参考价值", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 69, "start": 153.4, "end": 154.7, "text": "这些模型公司啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 70, "start": 154.733, "end": 157.166, "text": "他们想尽办法把Claude", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 71, "start": 157.3, "end": 159.366, "text": "隐藏的这个reasoning trace", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 72, "start": 159.5, "end": 161.733, "text": "这个推理的过程给挖出来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 73, "start": 161.766, "end": 163.366, "text": "比如说有人就会把这个", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 74, "start": 163.366, "end": 166.5, "text": "会话伪装成一个debug的过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 75, "start": 166.633, "end": 168.1, "text": "让它交出推理链", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 76, "start": 168.1, "end": 171.2, "text": "或者说有人尝试说让Claude把", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 77, "start": 171.2, "end": 173.433, "text": "之前他做的这个working memory", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 78, "start": 173.5, "end": 174.666, "text": "提一个任务说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 79, "start": 174.666, "end": 176.633, "text": "你把这个内容翻译成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 80, "start": 176.633, "end": 179.066, "text": "只包含片假名的日语", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 81, "start": 179.4, "end": 182.566, "text": "诱骗它把这个推理过程交出来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 82, "start": 182.633, "end": 184.066, "text": "而且更夸张的是啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 83, "start": 184.066, "end": 185.7, "text": "有的AI实验室", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 84, "start": 186.033, "end": 189.433, "text": "它直接用了1.2万种不同", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 85, "start": 189.466, "end": 191.633, "text": "的请求方式去实验", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 86, "start": 191.9, "end": 193.9, "text": "怎么把它的隐藏的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 87, "start": 193.9, "end": 196.066, "text": "推理过程给套路出来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 88, "start": 196.066, "end": 197.833, "text": "它的目的就是测试", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 89, "start": 197.966, "end": 199.966, "text": "看看哪一种提示词", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 90, "start": 200.166, "end": 201.333, "text": "哪一种路径", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 91, "start": 201.333, "end": 202.766, "text": "能够把能够把Claude的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 92, "start": 202.8, "end": 204.366, "text": "推理路径给套出来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 93, "start": 204.533, "end": 206.066, "text": "有可能大部分的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 94, "start": 206.1, "end": 208.433, "text": "提示词方式都可能失败", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 95, "start": 208.466, "end": 210.733, "text": "但是只要有几种成功了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 96, "start": 210.766, "end": 214.033, "text": "就可以用工业化的方式把它放大", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 97, "start": 214.1, "end": 216.4, "text": "作为蒸馏的一种有效的路径", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 98, "start": 216.433, "end": 217.433, "text": "所以这实际上是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 99, "start": 217.433, "end": 219.333, "text": "一个很有意思的攻防", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 100, "start": 219.633, "end": 220.6, "text": "模型厂商呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 101, "start": 220.633, "end": 222.8, "text": "他说我不给你原始的COT", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 102, "start": 222.866, "end": 224.766, "text": "我只给你总结的thinking", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 103, "start": 224.8, "end": 227.166, "text": "但蒸馏方就一直在研究", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 104, "start": 227.433, "end": 229.933, "text": "怎么通过上下文翻译", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 105, "start": 230.1, "end": 231.833, "text": "Replay的方式等等", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 106, "start": 231.9, "end": 234.733, "text": "把这个推理的过程给恢复过来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 107, "start": 234.766, "end": 237.133, "text": "然后再把这个reasoning推理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 108, "start": 237.166, "end": 239.366, "text": "的过程变成训练数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 109, "start": 239.466, "end": 242.033, "text": "你看看这有不有一点像AI", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 110, "start": 242.066, "end": 244.733, "text": "时代的能力逆向工程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 111, "start": 244.766, "end": 245.2, "text": "实际上啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 112, "start": 245.2, "end": 247.633, "text": "这个报告本身披露出了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 113, "start": 247.633, "end": 251.633, "text": "非常多类似这些AI模型公司", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 114, "start": 251.666, "end": 253.133, "text": "他们蒸馏的方式啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 115, "start": 253.2, "end": 255.0, "text": "而这些蒸馏的类型啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 116, "start": 255.033, "end": 257.733, "text": "我会给大家在后面去给做总结", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 117, "start": 258.1, "end": 260.033, "text": "并且啊总结出的方式我们", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 118, "start": 260.033, "end": 261.4, "text": "可以进一步的拓展", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 119, "start": 261.533, "end": 264.433, "text": "把它用于我们日常的生产力环节", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 120, "start": 264.466, "end": 265.6, "text": "好准备好了吗", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 121, "start": 265.766, "end": 267.166, "text": "接下来我就带着大家", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 122, "start": 267.233, "end": 269.233, "text": "全面的去理解蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 123, "start": 269.6, "end": 271.766, "text": "蒸馏啊不要把它仅仅的理解为", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 124, "start": 271.8, "end": 273.7, "text": "一种模型训练的技术", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 125, "start": 273.833, "end": 275.966, "text": "我认为蒸馏的本质是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 126, "start": 276.0, "end": 278.7, "text": "一次昂贵的智能过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 127, "start": 278.733, "end": 281.266, "text": "它实际上是在压缩成下一次更便宜、更快、更稳定", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "它实际上是在压缩成下", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 128, "start": 281.266, "end": 284.533, "text": "一次更便宜更快更稳定", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 129, "start": 284.7, "end": 287.6, "text": "甚至不需要还原老师的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 130, "start": 287.6, "end": 289.333, "text": "能力的这么一个过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 131, "start": 289.366, "end": 290.9, "text": "Anthropic它给我们看的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 132, "start": 290.9, "end": 293.0, "text": "实际上是蒸馏中的一种", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 133, "start": 293.1, "end": 295.7, "text": "就是模型蒸馏模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 134, "start": 295.833, "end": 297.9, "text": "所以我会先跟大家去讲这个", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 135, "start": 297.933, "end": 300.733, "text": "模型蒸馏模型有哪些过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 136, "start": 300.733, "end": 301.733, "text": "有哪些类型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 137, "start": 302.1, "end": 304.8, "text": "而沿着这个定义往外推", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 138, "start": 304.833, "end": 306.9, "text": "我们可以看到更多的东西", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 139, "start": 307.133, "end": 309.133, "text": "除了模型蒸馏模型之外", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 140, "start": 309.266, "end": 311.4, "text": "模型还可以蒸馏用户", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 141, "start": 311.6, "end": 313.933, "text": "产品也能够蒸馏模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 142, "start": 314.1, "end": 315.9, "text": "agent能蒸馏工作流", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 143, "start": 316.033, "end": 318.166, "text": "企业能蒸馏员工", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 144, "start": 318.3, "end": 321.766, "text": "甚至是人用AI来蒸馏自己", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 145, "start": 321.833, "end": 324.833, "text": "而这些就是我们广阔的应用场景", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 146, "start": 324.866, "end": 327.6, "text": "回到Anthropic这个报告来看啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 147, "start": 327.8, "end": 329.833, "text": "它里面讲了非常多的细节", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 148, "start": 329.866, "end": 331.166, "text": "是有助于我们去", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 149, "start": 331.166, "end": 333.6, "text": "理解模型是怎么蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 150, "start": 333.666, "end": 335.466, "text": "传统的我们对蒸馏的理解", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 151, "start": 335.5, "end": 337.5, "text": "实际上是特别简单直接的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 152, "start": 337.533, "end": 341.8, "text": "就是教师模型根据输入给出输出", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 153, "start": 342.0, "end": 344.766, "text": "然后学生模型学习", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 154, "start": 344.833, "end": 346.966, "text": "整个的信息交换的过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 155, "start": 347.033, "end": 349.0, "text": "学习这些exchange的过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 156, "start": 349.033, "end": 350.566, "text": "但是大家注意到这里面的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 157, "start": 350.6, "end": 353.233, "text": "环节会涉及到inputs和OUTPUTS", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 158, "start": 353.233, "end": 354.333, "text": "在这个报告里面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 159, "start": 354.366, "end": 358.3, "text": "它实际上把这个inputs和outputs怎么来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 160, "start": 358.4, "end": 362.133, "text": "做了非常详细的解析和案例的披露", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 161, "start": 362.133, "end": 365.233, "text": "也就是这些中国的AI实验室", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 162, "start": 365.266, "end": 366.366, "text": "他们是怎么干的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 163, "start": 366.533, "end": 367.2, "text": "这些东西啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 164, "start": 367.233, "end": 368.233, "text": "非常的有启发", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 165, "start": 368.366, "end": 370.0, "text": "可以分为7个类型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 166, "start": 370.0, "end": 370.866, "text": "我总结下来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 167, "start": 370.866, "end": 372.4, "text": "一个个跟大家去讲", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 168, "start": 372.466, "end": 374.133, "text": "第一种类型啊最传统", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 169, "start": 374.166, "end": 376.566, "text": "我总结为叫出题蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "我总结为叫做出题蒸馏", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 170, "start": 376.766, "end": 378.9, "text": "也可以叫做答案蒸馏吧", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 171, "start": 378.966, "end": 380.933, "text": "就是学生给老师出题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 172, "start": 380.933, "end": 382.266, "text": "让老师回答", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 173, "start": 382.266, "end": 384.466, "text": "形成一个高质量的答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 174, "start": 384.8, "end": 386.633, "text": "你可以准备一个任务集", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 175, "start": 386.8, "end": 387.833, "text": "这个任务集呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 176, "start": 387.833, "end": 390.966, "text": "可以是coding数学agent数据分析", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 177, "start": 391.0, "end": 392.933, "text": "或者长程任务等等都可以", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 178, "start": 393.0, "end": 395.033, "text": "然后把这个问题给到Claude", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 179, "start": 395.133, "end": 396.5, "text": "让Claude来回答", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 180, "start": 396.533, "end": 399.4, "text": "形成高质量的答案或者推理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 181, "start": 399.7, "end": 401.433, "text": "然后把这个答案抄下来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 182, "start": 401.433, "end": 403.533, "text": "放进你的data set里面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 183, "start": 403.566, "end": 406.4, "text": "接着通过SFT监督微调", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 184, "start": 406.4, "end": 409.4, "text": "或者是通过RL强化学习", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 185, "start": 409.433, "end": 411.2, "text": "来训练自己的模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 186, "start": 411.3, "end": 412.566, "text": "在这个报告里面啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 187, "start": 412.7, "end": 415.8, "text": "Anthropic声称阿里千问的蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 188, "start": 415.8, "end": 418.066, "text": "模式就是最典型的例子", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 189, "start": 418.533, "end": 422.6, "text": "它主要是针对Opus 4.6和4.7的推理轨迹", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 190, "start": 422.833, "end": 425.733, "text": "它的整个的蒸馏任务是集中在agent", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 191, "start": 425.933, "end": 426.966, "text": "软件工程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 192, "start": 427.033, "end": 429.266, "text": "还有一些长程任务上面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 193, "start": 429.466, "end": 431.166, "text": "并且它把整个的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 194, "start": 431.166, "end": 434.5, "text": "Transcript转成了SFT的数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 195, "start": 434.533, "end": 437.766, "text": "这种呢就是考试考卷式的蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 196, "start": 437.866, "end": 439.533, "text": "这是最传统的蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 197, "start": 439.533, "end": 441.033, "text": "也是最容易理解的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 198, "start": 441.033, "end": 442.566, "text": "老师做1万道题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 199, "start": 442.6, "end": 446.033, "text": "学生直接拿老师给的答案进行训练", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 200, "start": 446.066, "end": 448.7, "text": "第二种类型我总结为推理蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 201, "start": 448.766, "end": 450.8, "text": "就是学生不仅是要答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 202, "start": 450.933, "end": 453.0, "text": "还要老师的解题过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 203, "start": 453.033, "end": 454.866, "text": "这个答案确实有价值", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 204, "start": 454.866, "end": 457.033, "text": "但是更值钱的是教师", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 205, "start": 457.033, "end": 458.733, "text": "模型它是怎么拆解问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 206, "start": 458.766, "end": 459.6, "text": "怎么判断", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 207, "start": 459.733, "end": 460.6, "text": "怎么试错", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 208, "start": 460.6, "end": 462.133, "text": "调用工具验证", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 209, "start": 462.233, "end": 464.233, "text": "得出最终答案的过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 210, "start": 464.266, "end": 465.2, "text": "要知其然", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 211, "start": 465.266, "end": 467.1, "text": "更要知其所以然", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 212, "start": 467.4, "end": 470.066, "text": "实际上它不仅是要老师授之以鱼", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 213, "start": 470.233, "end": 471.9, "text": "更要授之以渔", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 214, "start": 472.166, "end": 474.033, "text": "这个方式也是anthropic", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 215, "start": 474.066, "end": 476.133, "text": "这次最在意的一种方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 216, "start": 476.3, "end": 477.433, "text": "刚刚我提到了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 217, "start": 477.633, "end": 479.3, "text": "这些攻击者测试啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 218, "start": 479.366, "end": 481.966, "text": "里面花了1.2万种方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 219, "start": 481.966, "end": 482.866, "text": "通过翻译", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 220, "start": 483.066, "end": 485.3, "text": "debug replay等等方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 221, "start": 485.366, "end": 489.333, "text": "试图去恢复Claude模型的推理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 222, "start": 489.333, "end": 491.466, "text": "的过程和找到其中的线索", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 223, "start": 491.5, "end": 493.866, "text": "以后啊这个COT推理的过程会", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 224, "start": 493.866, "end": 495.9, "text": "被藏的越来越深", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 225, "start": 495.933, "end": 497.266, "text": "那么蒸馏怎么办呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 226, "start": 497.266, "end": 500.033, "text": "其实就是不断的攻防攻防", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 227, "start": 500.066, "end": 501.033, "text": "互相套路", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 228, "start": 501.066, "end": 502.9, "text": "越狱嘛反正是一直去试", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 229, "start": 503.066, "end": 505.5, "text": "而这些前沿模型也会越来越不", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 230, "start": 505.533, "end": 508.466, "text": "愿意把原始的这个COT的链路啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 231, "start": 508.6, "end": 509.266, "text": "公布出来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 232, "start": 509.3, "end": 510.0, "text": "展示出来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 233, "start": 510.066, "end": 512.5, "text": "以后可能会有一个很火的角色", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 234, "start": 512.5, "end": 514.8, "text": "比如说像AI狱警这样", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 235, "start": 514.866, "end": 517.466, "text": "第三个类型我总结为Judge蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "第三个类型我总结为judge蒸馏", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 236, "start": 517.666, "end": 519.8, "text": "也就是让前沿模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 237, "start": 519.8, "end": 521.5, "text": "让老师来当评委", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 238, "start": 521.666, "end": 523.6, "text": "比如说你让一个普通模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 239, "start": 523.6, "end": 524.866, "text": "比如说国产模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 240, "start": 525.066, "end": 529.0, "text": "让它对一个问题生成a b c d四个答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 241, "start": 529.033, "end": 529.666, "text": "这个时候啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 242, "start": 529.666, "end": 531.333, "text": "你可以不用前沿模型来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 243, "start": 531.333, "end": 534.033, "text": "生成一个优质或者标准答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 244, "start": 534.166, "end": 537.333, "text": "而是让这个前沿模型Claude", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 245, "start": 537.766, "end": 541.633, "text": "来对这四个模型的答案进行打分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 246, "start": 541.833, "end": 543.366, "text": "他来评分 1-10分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 247, "start": 543.4, "end": 544.2, "text": "他打几分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 248, "start": 544.366, "end": 545.7, "text": "哪个答案好", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 249, "start": 545.766, "end": 546.866, "text": "哪个答案不好", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 250, "start": 547.0, "end": 547.766, "text": "为什么好", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 251, "start": 547.766, "end": 548.766, "text": "为什么答错", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 252, "start": 548.8, "end": 550.4, "text": "让他给出评价", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 253, "start": 550.433, "end": 552.366, "text": "这时候链路就变成了什么", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 254, "start": 552.566, "end": 554.333, "text": "普通模型生成答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 255, "start": 554.466, "end": 557.366, "text": "然后Claude作为评委进行评分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 256, "start": 557.533, "end": 560.133, "text": "然后对所有的答案进行排序", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 257, "start": 560.533, "end": 563.1, "text": "接着可以进行强化学习", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 258, "start": 563.3, "end": 565.2, "text": "然后进行模型训练", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 259, "start": 565.2, "end": 567.333, "text": "让普通的模型变强", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 260, "start": 567.433, "end": 570.1, "text": "这个在Anthropic的这个报告里面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 261, "start": 570.133, "end": 571.8, "text": "关于智谱的这个描述", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 262, "start": 571.833, "end": 573.6, "text": "和归因部分也有体现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 263, "start": 573.866, "end": 576.0, "text": "智谱呢就用这个Claude的模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 264, "start": 576.066, "end": 578.933, "text": "让它生成一些评委评分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 265, "start": 578.933, "end": 581.333, "text": "的一些结果的一个标准", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 266, "start": 581.6, "end": 582.733, "text": "然后还用它来做", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 267, "start": 582.766, "end": 584.666, "text": "一些数据清洗的角色", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 268, "start": 584.666, "end": 586.833, "text": "并且让它提供一些解决方案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 269, "start": 586.933, "end": 588.533, "text": "这个时候就不是纯粹的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 270, "start": 588.566, "end": 590.566, "text": "给答案对比这件事情了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 271, "start": 590.7, "end": 593.8, "text": "它让这个前沿模型在它的整个", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 272, "start": 593.833, "end": 596.833, "text": "模型训练中扮演了更多的角色", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 273, "start": 596.866, "end": 598.466, "text": "这个前沿的模型不仅", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 274, "start": 598.466, "end": 600.8, "text": "是当老师给解题过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 275, "start": 600.833, "end": 603.166, "text": "而是它也可以是出题的老师", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 276, "start": 603.333, "end": 606.066, "text": "也可以是阅卷评分的老师", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 277, "start": 606.3, "end": 608.9, "text": "它还可以是数据的清洗员", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 278, "start": 609.0, "end": 611.3, "text": "还可以是安全的检查员", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 279, "start": 611.333, "end": 613.333, "text": "还可以是最后的强化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 280, "start": 613.333, "end": 614.9, "text": "学习的这个评估员", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 281, "start": 614.933, "end": 616.566, "text": "如果是经常看我节目", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 282, "start": 616.566, "end": 618.266, "text": "的老的粉丝和观众", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 283, "start": 618.266, "end": 620.266, "text": "一定感受到我平时在", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 284, "start": 620.3, "end": 622.1, "text": "节目中经常用这个方法", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 285, "start": 622.266, "end": 623.933, "text": "比如说我经常用多个", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 286, "start": 623.966, "end": 626.633, "text": "模型来处理同一个任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 287, "start": 626.933, "end": 628.9, "text": "然后我再用一个前沿模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 288, "start": 628.933, "end": 631.633, "text": "比如说用GPT的Codex来对", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 289, "start": 631.633, "end": 633.533, "text": "这些结果进行评分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 290, "start": 633.6, "end": 635.966, "text": "然后评价这几个模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 291, "start": 635.966, "end": 637.666, "text": "的能力和最后的答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 292, "start": 637.833, "end": 639.9, "text": "然后给出我的评测结论", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 293, "start": 640.133, "end": 643.266, "text": "这本质上不就是使用judge的方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 294, "start": 643.366, "end": 645.366, "text": "让这个前沿模型扮演", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 295, "start": 645.366, "end": 647.266, "text": "一个评委的角色吗", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 296, "start": 647.3, "end": 649.566, "text": "你看看从蒸馏的角度来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 297, "start": 649.566, "end": 652.433, "text": "理解我节目中的评测的方法", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 298, "start": 652.5, "end": 654.266, "text": "是不是又升维了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 299, "start": 654.3, "end": 655.633, "text": "第四种蒸馏方式啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 300, "start": 655.633, "end": 658.166, "text": "第四种蒸馏方式啊我总结叫做Replay蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "我总结叫做Replay蒸馏", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 301, "start": 658.466, "end": 659.1, "text": "什么意思", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 302, "start": 659.1, "end": 661.166, "text": "就是不用自己出题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 303, "start": 661.333, "end": 664.0, "text": "直接拿真实用户的问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 304, "start": 664.2, "end": 667.1, "text": "重新让老师再做一遍", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 305, "start": 667.133, "end": 668.466, "text": "你知道玩这个路子的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 306, "start": 668.466, "end": 670.033, "text": "是哪家模型公司吗", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 307, "start": 670.233, "end": 673.766, "text": "是小米的模型MIMO他们家玩的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 308, "start": 674.033, "end": 675.166, "text": "就是假设啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 309, "start": 675.4, "end": 677.833, "text": "他们家公司已经有100万的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 310, "start": 677.866, "end": 680.466, "text": "真实开发者的这个conversation", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 311, "start": 680.666, "end": 681.366, "text": "这个对话", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 312, "start": 681.666, "end": 684.633, "text": "这个实际上是非常宝贵的数据啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 313, "start": 684.666, "end": 687.266, "text": "大家不要觉得自己很多", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 314, "start": 687.266, "end": 689.133, "text": "地方可以免费用一些模型啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 315, "start": 689.133, "end": 692.566, "text": "实际上都是这些模型公司的养料", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 316, "start": 692.866, "end": 695.3, "text": "这里面会包含一些真的bug", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 317, "start": 695.566, "end": 697.466, "text": "真的项目真的IDE", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 318, "start": 697.466, "end": 698.333, "text": "真的文件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 319, "start": 698.466, "end": 699.9, "text": "真的用户意图", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 320, "start": 700.066, "end": 701.766, "text": "真的失败案例", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 321, "start": 701.8, "end": 703.8, "text": "原来在小米的模型里面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 322, "start": 703.9, "end": 705.666, "text": "它就会有原始的用户", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 323, "start": 705.666, "end": 708.1, "text": "任务和小米给出的答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 324, "start": 708.2, "end": 710.0, "text": "那么重新Replay一遍呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 325, "start": 710.0, "end": 712.733, "text": "就是把同一个任务给到Claude", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 326, "start": 712.966, "end": 715.166, "text": "然后再获得Claude的答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 327, "start": 715.2, "end": 716.966, "text": "然后就立刻得到了什么呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 328, "start": 717.066, "end": 718.266, "text": "首先是一个task", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 329, "start": 718.7, "end": 720.066, "text": "用户的真实任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 330, "start": 720.066, "end": 721.433, "text": "然后是student", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 331, "start": 721.8, "end": 722.766, "text": "MIMO的答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 332, "start": 722.933, "end": 724.033, "text": "然后是teacher", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 333, "start": 724.2, "end": 725.066, "text": "Claude的答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 334, "start": 725.166, "end": 726.533, "text": "接下来就可以干什么", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 335, "start": 726.6, "end": 727.666, "text": "就可以比较", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 336, "start": 727.733, "end": 728.4, "text": "可以judge", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 337, "start": 728.6, "end": 729.5, "text": "可以回写", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 338, "start": 729.9, "end": 732.633, "text": "可以SFT可以RL了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 339, "start": 732.666, "end": 734.233, "text": "小米把自己代码相关", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 340, "start": 734.233, "end": 736.633, "text": "的coding session给到claude", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 341, "start": 736.733, "end": 739.066, "text": "然后呢Claude还可以被用于", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 342, "start": 739.066, "end": 741.333, "text": "重构它的开发环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 343, "start": 741.333, "end": 742.533, "text": "再清洗对话", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 344, "start": 742.566, "end": 744.666, "text": "然后再生成回答", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 345, "start": 744.733, "end": 747.2, "text": "再生成响应评价的答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 346, "start": 747.2, "end": 750.866, "text": "这可比人工去造Benchmark厉害多了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 347, "start": 750.966, "end": 752.4, "text": "经常我给大家演示的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 348, "start": 752.4, "end": 753.533, "text": "考虑到隐私啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 349, "start": 753.566, "end": 755.566, "text": "我都是造的这些数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 350, "start": 755.666, "end": 759.066, "text": "而这些数据都是用户的真实意图", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 351, "start": 759.4, "end": 762.966, "text": "而用户本身就会成为源源不断", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 352, "start": 762.966, "end": 766.033, "text": "的Benchmark的生产者和生成者", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 353, "start": 766.066, "end": 767.9, "text": "第五种蒸馏方式啊我不说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 354, "start": 767.9, "end": 770.2, "text": "可能很多人还蒙在鼓里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 355, "start": 770.433, "end": 772.2, "text": "我叫做shadow teacher", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 356, "start": 772.4, "end": 773.633, "text": "影子老师", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 357, "start": 773.633, "end": 776.233, "text": "这是一个水加茅台的故事", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 358, "start": 776.333, "end": 777.0, "text": "怎么说呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 359, "start": 777.066, "end": 779.1, "text": "就是本来你只是要了一杯水", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 360, "start": 779.133, "end": 780.666, "text": "你花了一杯水的钱", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 361, "start": 780.766, "end": 783.466, "text": "结果这个水里面掺了茅台", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 362, "start": 783.633, "end": 784.533, "text": "怎么理解啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 363, "start": 784.7, "end": 788.9, "text": "就是用户以为自己用的是学生模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 364, "start": 788.933, "end": 789.6, "text": "结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 365, "start": 790.0, "end": 793.266, "text": "后台偷偷的让老师回答", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 366, "start": 793.266, "end": 795.133, "text": "然后用户被告知的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 367, "start": 795.133, "end": 797.1, "text": "是学生模型的回答", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 368, "start": 797.333, "end": 799.5, "text": "有不有毁你之前的认知", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 369, "start": 799.533, "end": 800.9, "text": "就是你给他一个问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 370, "start": 801.066, "end": 802.9, "text": "然后后台呢判断", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 371, "start": 803.033, "end": 805.6, "text": "诶这个问题非常有价值啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 372, "start": 805.633, "end": 807.166, "text": "这条数据很值钱啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 373, "start": 807.433, "end": 808.1, "text": "然后呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 374, "start": 808.2, "end": 812.566, "text": "他就把这个问题的请求转给了Claude", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 375, "start": 812.733, "end": 814.166, "text": "让Claude来回答", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 376, "start": 814.333, "end": 817.533, "text": "然后把Claude的答案推送给用户", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 377, "start": 817.566, "end": 818.333, "text": "与此同时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 378, "start": 818.333, "end": 820.6, "text": "他在后台把这个提示词", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 379, "start": 820.833, "end": 822.6, "text": "加上Claude的回答", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 380, "start": 822.733, "end": 824.566, "text": "加上他的推理过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 381, "start": 824.6, "end": 828.433, "text": "全部都被写入了他的训练的数据集", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 382, "start": 828.466, "end": 830.366, "text": "根据Anthropic的报告啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 383, "start": 830.633, "end": 833.8, "text": "月之暗面和Deepseek都在这样干", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 384, "start": 833.8, "end": 835.033, "text": "那怎么发现的呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 385, "start": 835.066, "end": 838.5, "text": "因为Claude它在清洗的时候发现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 386, "start": 838.566, "end": 841.633, "text": "这里面出现了非常多涉及", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 387, "start": 841.666, "end": 844.766, "text": "企业和涉及一些特殊行业", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 388, "start": 844.8, "end": 847.566, "text": "领域的一些机密的数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 389, "start": 847.7, "end": 850.3, "text": "出现在了模型的对话里面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 390, "start": 850.466, "end": 852.733, "text": "而这些模型的对话本不", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 391, "start": 852.766, "end": 856.3, "text": "应该出现在海外模型里面的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 392, "start": 856.333, "end": 857.933, "text": "这里面就有两个问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 393, "start": 857.966, "end": 860.4, "text": "本来你花了一个更便宜的钱", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 394, "start": 860.533, "end": 862.766, "text": "却买到了一个更贵的服务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 395, "start": 862.966, "end": 864.1, "text": "另外一方面呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 396, "start": 864.1, "end": 867.133, "text": "就是本来你觉得你用的是国产模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 397, "start": 867.166, "end": 869.366, "text": "不涉及到一些安全性的问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 398, "start": 869.7, "end": 873.366, "text": "结果模型悄悄地转包到了海外模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 399, "start": 873.533, "end": 876.466, "text": "把你的一些企业内部的信息", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 400, "start": 876.833, "end": 879.4, "text": "泄露到了不应该泄露的地方", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 401, "start": 879.633, "end": 882.1, "text": "涉及到信息安全的问题了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 402, "start": 882.133, "end": 883.066, "text": "但是要知道啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 403, "start": 883.1, "end": 885.866, "text": "这样的方式本身对国产", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 404, "start": 885.9, "end": 888.666, "text": "模型的训练是非常有帮助的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 405, "start": 888.9, "end": 891.333, "text": "因为很多这样的问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 406, "start": 891.533, "end": 893.2, "text": "它是在特定的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 407, "start": 893.233, "end": 896.533, "text": "比如说一个国产的IDE里面去产生的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 408, "start": 896.633, "end": 899.466, "text": "它本来是它原始的Harness的架构", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 409, "start": 899.533, "end": 901.066, "text": "那在这个Harness里面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 410, "start": 901.1, "end": 903.4, "text": "它本身是自带任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 411, "start": 903.433, "end": 904.5, "text": "还带上下文", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 412, "start": 904.5, "end": 905.366, "text": "又带文件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 413, "start": 905.4, "end": 906.233, "text": "带工具调用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 414, "start": 906.233, "end": 907.8, "text": "还带模型的尝试", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 415, "start": 908.066, "end": 909.866, "text": "带错误带用户反馈", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 416, "start": 909.966, "end": 913.266, "text": "然后现在再返回一个最终的结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 417, "start": 913.6, "end": 916.266, "text": "它的信息密度是非常高的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 418, "start": 916.3, "end": 918.666, "text": "比单纯的那种聊天记录", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 419, "start": 918.666, "end": 920.533, "text": "的信息密度高非常多", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 420, "start": 920.633, "end": 922.833, "text": "所以通过这样的方式转包", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 421, "start": 922.833, "end": 925.066, "text": "获得了前沿模型的回答", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 422, "start": 925.1, "end": 926.633, "text": "然后作为训练集", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 423, "start": 926.733, "end": 930.266, "text": "对于模型的成长是非常有帮助的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 424, "start": 930.3, "end": 933.033, "text": "所以你也会感知到为什么几家", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 425, "start": 933.033, "end": 936.133, "text": "大的公司都在拼命推他们的IDE", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 426, "start": 936.3, "end": 937.566, "text": "推他们的agent", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 427, "start": 937.566, "end": 938.7, "text": "第六种蒸馏方式啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 428, "start": 938.733, "end": 940.9, "text": "第六种蒸馏方式啊我叫做中转站蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "我叫做中转站蒸馏", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 429, "start": 941.266, "end": 942.866, "text": "哎很多人没听错啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 430, "start": 942.933, "end": 945.466, "text": "你是不是经常发现很多中转站", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 431, "start": 945.7, "end": 948.566, "text": "它使用一些前沿模型的价格", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 432, "start": 948.633, "end": 951.933, "text": "比你直接到官方那里使用要便宜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 433, "start": 952.2, "end": 952.866, "text": "为什么", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 434, "start": 953.066, "end": 955.4, "text": "这个商业逻辑想清楚了没有啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 435, "start": 955.433, "end": 957.333, "text": "他赚的可能根本不是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 436, "start": 957.333, "end": 959.5, "text": "什么TOKEN间的差价", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 437, "start": 959.766, "end": 962.566, "text": "而是这个数据的产业链", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 438, "start": 962.933, "end": 965.366, "text": "他最后啊会把这个对话的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 439, "start": 965.366, "end": 969.133, "text": "过程直接卖给其他的AI实验室", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 440, "start": 969.466, "end": 972.1, "text": "就是自己不用产生数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 441, "start": 972.333, "end": 974.166, "text": "直接去中转站把", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 442, "start": 974.166, "end": 976.266, "text": "上下文把数据买过来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 443, "start": 976.366, "end": 978.366, "text": "买别人的Claude对话", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 444, "start": 978.533, "end": 982.066, "text": "把这些对话拿来训练自己的模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 445, "start": 982.066, "end": 983.833, "text": "因为在这些中转站上面啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 446, "start": 983.833, "end": 985.633, "text": "天然就有提示词", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 447, "start": 985.666, "end": 987.766, "text": "上下文和Claude的回答", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 448, "start": 987.933, "end": 989.733, "text": "如果说你去保存", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 449, "start": 989.9, "end": 992.8, "text": "成千上百的整个的会话过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 450, "start": 992.866, "end": 994.866, "text": "那么就会形成非常", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 451, "start": 994.866, "end": 997.0, "text": "有价值的数据资产", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 452, "start": 997.4, "end": 1000.5, "text": "甚至你知道的一些中转站", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 453, "start": 1000.7, "end": 1004.466, "text": "它就是大厂下水自己干的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 454, "start": 1004.466, "end": 1007.333, "text": "只不过外面包了一个壳而已", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 455, "start": 1007.366, "end": 1008.433, "text": "第七种方式啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 456, "start": 1008.433, "end": 1010.633, "text": "第七种方式啊我定义为Pipeline蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "我定义为pipeline蒸馏", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 457, "start": 1010.8, "end": 1013.133, "text": "就是数据工厂蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 458, "start": 1013.2, "end": 1014.033, "text": "怎么理解啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 459, "start": 1014.233, "end": 1016.433, "text": "原来呢我们蒸馏的方式是让这个", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 460, "start": 1016.466, "end": 1019.066, "text": "教师模型教会学生模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 461, "start": 1019.366, "end": 1020.233, "text": "这种方式啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 462, "start": 1020.233, "end": 1021.033, "text": "非常的隐秘", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 463, "start": 1021.066, "end": 1022.1, "text": "但是很高级", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 464, "start": 1022.233, "end": 1026.166, "text": "现在他直接让这个教师构建出学校", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 465, "start": 1026.433, "end": 1028.2, "text": "然后这个学校就可以", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 466, "start": 1028.2, "end": 1030.633, "text": "源源不断的产生教师", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 467, "start": 1030.8, "end": 1033.133, "text": "培养更多的学生了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 468, "start": 1033.333, "end": 1036.166, "text": "就是让这个老师自己出题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 469, "start": 1036.166, "end": 1037.266, "text": "自己回答", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 470, "start": 1037.3, "end": 1038.4, "text": "自己批改", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 471, "start": 1038.466, "end": 1041.633, "text": "你已经不需要给一个具体的题库了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 472, "start": 1041.633, "end": 1044.466, "text": "你直接给Claude一个具体的领域", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 473, "start": 1044.7, "end": 1048.733, "text": "让它基于这个领域来进行自我出题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 474, "start": 1049.0, "end": 1050.666, "text": "让它出10万道题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 475, "start": 1050.766, "end": 1052.533, "text": "然后自己来回答", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 476, "start": 1052.633, "end": 1055.233, "text": "让学生模型来完成整个任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 477, "start": 1055.433, "end": 1058.0, "text": "然后Claude继续对这个任务的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 478, "start": 1058.033, "end": 1060.833, "text": "结果进行批改审阅评分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 479, "start": 1060.833, "end": 1062.2, "text": "然后可以纠错", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 480, "start": 1062.233, "end": 1064.933, "text": "筛选出优质的回答和过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 481, "start": 1065.2, "end": 1068.466, "text": "通过SFT和RL进行训练", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 482, "start": 1068.5, "end": 1069.266, "text": "这个时候", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 483, "start": 1069.366, "end": 1072.133, "text": "前沿的模型的角色不仅是老师", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 484, "start": 1072.366, "end": 1073.866, "text": "他是整个学校", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 485, "start": 1074.033, "end": 1075.366, "text": "出题老师是他", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 486, "start": 1075.433, "end": 1076.633, "text": "任课老师是他", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 487, "start": 1076.866, "end": 1078.233, "text": "批卷老师是他", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 488, "start": 1078.433, "end": 1080.133, "text": "教研组组长是他", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 489, "start": 1080.3, "end": 1082.4, "text": "教材的编写也是他", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 490, "start": 1082.433, "end": 1085.233, "text": "他帮你设计的是整个流程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 491, "start": 1085.4, "end": 1087.766, "text": "他来构建整个框架", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 492, "start": 1087.8, "end": 1090.066, "text": "上面我讲的7种蒸馏方式啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 493, "start": 1090.1, "end": 1091.9, "text": "就是我从Anthropic的这个", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 494, "start": 1091.933, "end": 1094.566, "text": "报告里面抽象总结出来的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 495, "start": 1094.6, "end": 1096.8, "text": "模型蒸馏模型的方式啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 496, "start": 1097.133, "end": 1099.3, "text": "从某种意义上面来讲啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 497, "start": 1099.466, "end": 1103.8, "text": "我觉得蒸馏也是AI时代技术", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 498, "start": 1103.9, "end": 1107.033, "text": "扩散的一种非常重要的方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 499, "start": 1107.066, "end": 1108.433, "text": "以前啊我们训练模型啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 500, "start": 1108.433, "end": 1110.333, "text": "是人类写数据AI学", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 501, "start": 1110.366, "end": 1112.666, "text": "或者AI生成数据AI学", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 502, "start": 1112.833, "end": 1114.833, "text": "现在呢通过蒸馏的方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 503, "start": 1115.033, "end": 1116.366, "text": "AI模型之间啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 504, "start": 1116.366, "end": 1118.866, "text": "它们互相学习的程度已经到", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 505, "start": 1118.866, "end": 1121.3, "text": "了一个非常深的一个程度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 506, "start": 1121.366, "end": 1122.8, "text": "在Anthropic的角度啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 507, "start": 1122.8, "end": 1125.366, "text": "他确实觉得应该反对这种蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 508, "start": 1125.366, "end": 1127.533, "text": "而且他会找出各种", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 509, "start": 1127.566, "end": 1129.833, "text": "方式来对抗这种蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 510, "start": 1129.833, "end": 1132.133, "text": "但是站在模型研究的角度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 511, "start": 1132.4, "end": 1135.666, "text": "蒸馏本身又是一种非常正常", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 512, "start": 1135.866, "end": 1137.333, "text": "非常重要的技术", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 513, "start": 1137.366, "end": 1140.9, "text": "连Anthropic自己都不得不承认这一点", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 514, "start": 1140.9, "end": 1141.9, "text": "所以啊这一次", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 515, "start": 1141.9, "end": 1145.9, "text": "Claude发现自己的智商正在被轻易", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 516, "start": 1145.933, "end": 1148.666, "text": "的工业化的批量复制的时候", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 517, "start": 1148.7, "end": 1150.3, "text": "又不禁引起我们反思", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 518, "start": 1150.533, "end": 1154.133, "text": "AI行业它真正的护城河在哪里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 519, "start": 1154.166, "end": 1156.8, "text": "而原来我们在聊天沟通中所", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 520, "start": 1156.8, "end": 1159.866, "text": "喜欢看到的这个COT的推理链", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 521, "start": 1160.166, "end": 1163.166, "text": "是不是就从用户的一个体验功能", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 522, "start": 1163.333, "end": 1166.0, "text": "变成了模型护城河的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 523, "start": 1166.0, "end": 1168.666, "text": "核心资产和安全边界了呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 524, "start": 1168.7, "end": 1169.3, "text": "这个问题啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 525, "start": 1169.3, "end": 1170.3, "text": "我觉得很复杂", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 526, "start": 1170.3, "end": 1172.166, "text": "但是关于蒸馏这个事情啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 527, "start": 1172.2, "end": 1174.766, "text": "我还想再深入拓展一下", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 528, "start": 1174.766, "end": 1178.5, "text": "我们不要只停留在模型蒸馏模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 529, "start": 1178.5, "end": 1180.3, "text": "实际上它有更大的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 530, "start": 1180.333, "end": 1182.433, "text": "应用价值和技术机制", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 531, "start": 1182.466, "end": 1183.9, "text": "我这里讲一个暴论啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 532, "start": 1183.933, "end": 1186.433, "text": "叫做一切皆可蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 533, "start": 1186.533, "end": 1189.0, "text": "等会我讲完蒸馏的整个拓展", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 534, "start": 1189.0, "end": 1191.0, "text": "你再回过头来看这句话", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 535, "start": 1191.133, "end": 1193.166, "text": "你会觉得特别的正确", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 536, "start": 1193.2, "end": 1194.966, "text": "而且特别的有用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 537, "start": 1195.0, "end": 1196.133, "text": "有实用价值", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 538, "start": 1196.533, "end": 1198.6, "text": "我们先看第一层拓展", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 539, "start": 1198.766, "end": 1200.833, "text": "叫做用户的蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 540, "start": 1201.133, "end": 1202.1, "text": "你想过没有", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 541, "start": 1202.1, "end": 1205.833, "text": "我们和AI对话交互的整个过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 542, "start": 1206.066, "end": 1209.566, "text": "实际上就是模型蒸馏人的过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 543, "start": 1209.933, "end": 1211.6, "text": "大家现在用Codex啊额度完全不够", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "大家现在用Codex啊", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 544, "start": 1211.633, "end": 1212.966, "text": "额度完全不够", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 545, "start": 1213.033, "end": 1214.433, "text": "但是你想过没有", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 546, "start": 1214.466, "end": 1218.166, "text": "你和GPT的聊天的额度和Codex", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 547, "start": 1218.266, "end": 1220.533, "text": "执行任务的额度是分开的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 548, "start": 1220.833, "end": 1223.733, "text": "而且你使用GPT聊天的这个额度啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 549, "start": 1223.766, "end": 1226.2, "text": "是很少达到上限的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 550, "start": 1226.2, "end": 1228.4, "text": "就是他给你的额度非常多", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 551, "start": 1228.6, "end": 1232.533, "text": "为什么大部分的模型公司都会给你", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 552, "start": 1232.533, "end": 1236.3, "text": "开放免费的和AI交互的这个额度啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 553, "start": 1236.433, "end": 1240.733, "text": "实际上是这些模型公司在蒸馏你", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 554, "start": 1240.766, "end": 1241.6, "text": "举个例子啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 555, "start": 1241.6, "end": 1243.6, "text": "你想得到某个答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 556, "start": 1243.866, "end": 1244.966, "text": "解决某个问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 557, "start": 1244.966, "end": 1247.366, "text": "你会跟AI可能对话20轮", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 558, "start": 1247.433, "end": 1249.633, "text": "最后拿到你想要的结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 559, "start": 1249.966, "end": 1251.5, "text": "在这20轮对话里面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 560, "start": 1251.5, "end": 1252.566, "text": "你会做什么呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 561, "start": 1252.6, "end": 1253.5, "text": "你会示范", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 562, "start": 1253.633, "end": 1254.866, "text": "告诉他什么是对的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 563, "start": 1254.866, "end": 1255.9, "text": "给他举例子", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 564, "start": 1255.933, "end": 1256.8, "text": "你会取舍", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 565, "start": 1256.9, "end": 1258.966, "text": "告诉他a要b不要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 566, "start": 1258.966, "end": 1260.066, "text": "你会纠偏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 567, "start": 1260.1, "end": 1262.1, "text": "告诉他这个方向不对", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 568, "start": 1262.233, "end": 1263.866, "text": "而这个过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 569, "start": 1264.1, "end": 1266.966, "text": "其实你就是给了他非常多的输入", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 570, "start": 1267.066, "end": 1269.266, "text": "最后得到满意的答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 571, "start": 1269.433, "end": 1273.633, "text": "这些东西都是给到AI模型的语料", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 572, "start": 1273.7, "end": 1276.233, "text": "而下一次优化训练之后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 573, "start": 1276.366, "end": 1279.566, "text": "模型就可以很快的识别你的意图", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 574, "start": 1279.7, "end": 1281.233, "text": "它不需要你再过20轮", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 575, "start": 1281.433, "end": 1283.9, "text": "它通过一到两轮的对话", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 576, "start": 1283.9, "end": 1286.466, "text": "就可以实现满意的答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 577, "start": 1286.5, "end": 1288.766, "text": "这也是为什么顶尖的大模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 578, "start": 1288.933, "end": 1290.3, "text": "用户多的大模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 579, "start": 1290.333, "end": 1292.8, "text": "它会有更好更多的数据集", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 580, "start": 1293.133, "end": 1295.566, "text": "来提升它本身的模型能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 581, "start": 1295.8, "end": 1297.833, "text": "大家现在用GPT的模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 582, "start": 1297.933, "end": 1299.466, "text": "大家现在用GPT的模型从GPT4到GPT4o", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "从GPT5到GPT6", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 583, "start": 1299.5, "end": 1301.566, "text": "大家是不是有明显的感知", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 584, "start": 1301.833, "end": 1303.666, "text": "模型越来越聪明", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 585, "start": 1303.8, "end": 1307.766, "text": "需要的聊天的次数和轮次越来越少", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 586, "start": 1307.8, "end": 1310.633, "text": "它能越来越知道你的心中所想", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 587, "start": 1310.633, "end": 1311.533, "text": "怎么出来的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 588, "start": 1311.766, "end": 1314.133, "text": "就是通过模型蒸馏人", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 589, "start": 1314.133, "end": 1316.7, "text": "的过程来逐步实现的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 590, "start": 1316.933, "end": 1319.566, "text": "我们以为我们在使用AI实际上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 591, "start": 1319.6, "end": 1322.366, "text": "从训练系统的视角来看", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 592, "start": 1322.5, "end": 1324.666, "text": "人的每一次纠错", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 593, "start": 1324.733, "end": 1327.533, "text": "实际上都可能成为什么是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 594, "start": 1327.533, "end": 1330.633, "text": "最好答案的一个监督信号", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 595, "start": 1330.666, "end": 1332.6, "text": "既然模型能够蒸馏人", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 596, "start": 1332.633, "end": 1334.6, "text": "那么我们是不是能进一步拓展", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 597, "start": 1334.833, "end": 1338.566, "text": "模型公司也能够蒸馏它的员工呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 598, "start": 1338.566, "end": 1339.566, "text": "比如说大厂", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 599, "start": 1339.566, "end": 1341.6, "text": "像前段时间爆出来Meta", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 600, "start": 1341.633, "end": 1343.766, "text": "它会记录它所有员工", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 601, "start": 1343.766, "end": 1345.833, "text": "的键盘的各种操作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 602, "start": 1346.166, "end": 1349.533, "text": "来更大范围的炼化员工的能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 603, "start": 1349.6, "end": 1352.566, "text": "去实现对人的能力的蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 604, "start": 1352.6, "end": 1354.566, "text": "这个实际上也是模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 605, "start": 1354.566, "end": 1356.566, "text": "对人蒸馏的一种方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 606, "start": 1356.666, "end": 1358.2, "text": "那我们更进一步", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 607, "start": 1358.2, "end": 1360.566, "text": "我们可以主动的去实现这个过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 608, "start": 1360.6, "end": 1363.3, "text": "比如说我有大量的节目制作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 609, "start": 1363.333, "end": 1366.033, "text": "的过程语料纠偏等等内容", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 610, "start": 1366.3, "end": 1367.733, "text": "都在我的本机", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 611, "start": 1367.8, "end": 1370.533, "text": "在我的新版的GPT的对话里面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 612, "start": 1370.6, "end": 1372.0, "text": "还有很多本地的材料", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 613, "start": 1372.033, "end": 1374.466, "text": "还存在我本地的Obsidian里面", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "还存在我本地的obsidian里面", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 614, "start": 1374.466, "end": 1378.466, "text": "那在这个里面我就可以实现让模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 615, "start": 1378.5, "end": 1382.033, "text": "对我的所有的语料过程进行蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 616, "start": 1382.033, "end": 1384.233, "text": "不仅是我这样的自媒体博主", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 617, "start": 1384.233, "end": 1386.7, "text": "你日常的工作如果都实现了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 618, "start": 1386.7, "end": 1389.9, "text": "在线化和颗粒度的对齐和AI", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 619, "start": 1389.933, "end": 1391.866, "text": "模型的对齐和纠偏的话", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 620, "start": 1391.966, "end": 1394.633, "text": "你都可以尝试用前沿模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 621, "start": 1394.633, "end": 1397.1, "text": "对你本身进行自我蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 622, "start": 1397.333, "end": 1400.233, "text": "让你的整体的能力再上一个台阶", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 623, "start": 1400.266, "end": 1400.833, "text": "举个例子啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 624, "start": 1400.833, "end": 1403.566, "text": "大家现在看到的是我Codex的对话框", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 625, "start": 1403.833, "end": 1407.233, "text": "我正在对我自己的生活和工作场景", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 626, "start": 1407.433, "end": 1411.0, "text": "特别是我视频制作的过程进行蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 627, "start": 1411.366, "end": 1413.933, "text": "这里面就会涉及到我的选题和脚本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 628, "start": 1414.0, "end": 1415.833, "text": "我的封面和标题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 629, "start": 1415.833, "end": 1417.266, "text": "客户的方案等等等等", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 630, "start": 1417.433, "end": 1419.5, "text": "每一个环节都是可以", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 631, "start": 1419.533, "end": 1421.166, "text": "蒸馏和抽象出来的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 632, "start": 1421.4, "end": 1423.533, "text": "就以选题和脚本为例", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 633, "start": 1423.6, "end": 1426.033, "text": "那么AI会给我输出一个脚本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 634, "start": 1426.033, "end": 1427.766, "text": "我会返回给他告诉他", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 635, "start": 1427.866, "end": 1429.166, "text": "跟他对齐颗粒度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 636, "start": 1429.166, "end": 1430.3, "text": "在这个脚本里面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 637, "start": 1430.333, "end": 1431.933, "text": "哪些内容我觉得不顺滑", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 638, "start": 1432.133, "end": 1433.666, "text": "哪些部分你讲得太少", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 639, "start": 1433.7, "end": 1436.433, "text": "哪些地方你可以加一些金句等等", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 640, "start": 1436.7, "end": 1437.966, "text": "这个实际上就是我", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 641, "start": 1438.0, "end": 1440.066, "text": "给他的投喂和纠偏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 642, "start": 1440.1, "end": 1442.833, "text": "那他就可以去对比他输出的答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 643, "start": 1442.9, "end": 1446.3, "text": "和我给的一个调整示范的内容", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 644, "start": 1446.566, "end": 1447.933, "text": "通过这样的方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 645, "start": 1448.033, "end": 1450.866, "text": "原来我可能要聊三四轮四五轮", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 646, "start": 1450.966, "end": 1453.933, "text": "到后面我可能只需要聊一两轮", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 647, "start": 1454.066, "end": 1457.9, "text": "就可以完成我的选题和脚本的确认", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 648, "start": 1457.933, "end": 1458.9, "text": "通过这样的方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 649, "start": 1458.9, "end": 1461.566, "text": "我可以沉淀出来我自己的skill", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 650, "start": 1461.8, "end": 1464.6, "text": "或者是通过某种工具的调用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 651, "start": 1464.633, "end": 1467.433, "text": "去让整个的过程更加的丝滑", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 652, "start": 1467.433, "end": 1470.6, "text": "这某种程度上也是一种蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 653, "start": 1470.966, "end": 1472.066, "text": "因为这里的蒸馏它", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 654, "start": 1472.1, "end": 1473.866, "text": "不只限于一期节目", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 655, "start": 1474.033, "end": 1475.133, "text": "我的100多期", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 656, "start": 1475.166, "end": 1476.433, "text": "200多期节目", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 657, "start": 1476.433, "end": 1478.7, "text": "整个的制作过程如果都是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 658, "start": 1478.966, "end": 1482.066, "text": "在本地在Codex完成了整个过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 659, "start": 1482.266, "end": 1484.9, "text": "再加上我本地的知识库obsidian", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 660, "start": 1484.966, "end": 1488.066, "text": "那我的蒸馏过程就能够把", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 661, "start": 1488.066, "end": 1490.766, "text": "我的整体能力去加强放大", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 662, "start": 1490.833, "end": 1493.2, "text": "而这里只是其中的一个环节", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 663, "start": 1493.3, "end": 1496.0, "text": "我还可以拓展到我的封面和标题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 664, "start": 1496.1, "end": 1497.466, "text": "之前我可能要不停的跟", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 665, "start": 1497.5, "end": 1498.933, "text": "他讲我需要什么风格", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 666, "start": 1499.0, "end": 1502.1, "text": "那形成一定的能力沉淀之后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 667, "start": 1502.2, "end": 1504.533, "text": "他可以很快的通过一轮", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 668, "start": 1504.566, "end": 1507.266, "text": "对话就完成我想要的封面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 669, "start": 1507.3, "end": 1509.033, "text": "在这里的蒸馏示范里面呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 670, "start": 1509.033, "end": 1510.7, "text": "我就选择了3期节目", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 671, "start": 1510.7, "end": 1513.3, "text": "我告诉他具体的节目内容和索引", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 672, "start": 1513.433, "end": 1515.966, "text": "因为这三期节目我是在整个的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 673, "start": 1515.966, "end": 1518.833, "text": "对话框中给了他明确的调整链", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 674, "start": 1518.833, "end": 1520.9, "text": "它是有明确的我的调整", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 675, "start": 1520.9, "end": 1523.466, "text": "和AI输出的前后对比的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 676, "start": 1523.466, "end": 1524.5, "text": "基于这个样本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 677, "start": 1524.5, "end": 1526.733, "text": "它给我做了第一轮的蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 678, "start": 1526.766, "end": 1528.566, "text": "你看已经完成了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 679, "start": 1528.7, "end": 1530.333, "text": "它这里是蒸馏的结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 680, "start": 1530.4, "end": 1530.9, "text": "这里呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 681, "start": 1530.933, "end": 1534.233, "text": "它会有明确的修改链和蒸馏发现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 682, "start": 1534.266, "end": 1536.7, "text": "它会识别我改了哪些", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 683, "start": 1536.733, "end": 1537.566, "text": "然后在这里啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 684, "start": 1537.566, "end": 1540.666, "text": "它是帮我直接沉淀了一个方法包", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 685, "start": 1540.733, "end": 1543.366, "text": "那这个方法包是可以复用的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 686, "start": 1543.633, "end": 1546.733, "text": "相当于他把我的工作能力习惯", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 687, "start": 1546.8, "end": 1550.233, "text": "判断这些有价值的东西抽象出来了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 688, "start": 1550.233, "end": 1551.333, "text": "你想想是不是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 689, "start": 1551.366, "end": 1554.466, "text": "这个事情在公司里面的员工的操作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 690, "start": 1554.466, "end": 1555.8, "text": "其实也是这个逻辑", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 691, "start": 1555.933, "end": 1558.0, "text": "这个员工他要审核pass", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 692, "start": 1558.433, "end": 1559.4, "text": "或者不通过", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 693, "start": 1559.433, "end": 1561.933, "text": "那他判断的依据标准是什么", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 694, "start": 1562.033, "end": 1565.766, "text": "就可以通过这样的方式蒸馏出来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 695, "start": 1565.866, "end": 1568.2, "text": "他为什么这么判断", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 696, "start": 1568.233, "end": 1570.4, "text": "而蒸馏出来的其实", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 697, "start": 1570.4, "end": 1573.233, "text": "就是人的能力和智慧", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 698, "start": 1573.266, "end": 1574.966, "text": "我们再往前思考一步啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 699, "start": 1575.066, "end": 1579.133, "text": "我们还可以做垂直专业领域的蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 700, "start": 1579.166, "end": 1580.733, "text": "而且越是垂直", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 701, "start": 1580.966, "end": 1582.733, "text": "越是专业的领域啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 702, "start": 1582.833, "end": 1585.233, "text": "小模型对大模型的蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 703, "start": 1585.233, "end": 1587.6, "text": "的替代的表现就会越好", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 704, "start": 1587.866, "end": 1588.366, "text": "举个例子啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 705, "start": 1588.4, "end": 1590.1, "text": "假如你有一家咨询公司", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 706, "start": 1590.3, "end": 1591.333, "text": "你这个公司呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 707, "start": 1591.366, "end": 1594.033, "text": "本身就有很多的典型任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 708, "start": 1594.133, "end": 1596.3, "text": "再加上比如说10年的文档", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 709, "start": 1596.5, "end": 1598.333, "text": "再加上专家的经验", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 710, "start": 1598.433, "end": 1599.766, "text": "这个时候你可以做什么呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 711, "start": 1599.833, "end": 1601.533, "text": "你就可以把你的企业", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 712, "start": 1601.533, "end": 1603.066, "text": "的历史数据拿出来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 713, "start": 1603.3, "end": 1606.066, "text": "再加上Claude它的前沿的能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 714, "start": 1606.266, "end": 1608.2, "text": "让Claude加上企业数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 715, "start": 1608.2, "end": 1610.433, "text": "去完成200个任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 716, "start": 1610.433, "end": 1613.166, "text": "这200个任务最好是有代表性的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 717, "start": 1613.2, "end": 1615.2, "text": "能够覆盖这个垂直", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 718, "start": 1615.233, "end": 1616.666, "text": "领域的大部分内容", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 719, "start": 1616.666, "end": 1617.833, "text": "Claude完成之后呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 720, "start": 1617.833, "end": 1620.066, "text": "它就会形成一个答案和回复", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 721, "start": 1620.066, "end": 1622.566, "text": "然后呢在这个咨询公司里会有专家", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 722, "start": 1622.666, "end": 1624.733, "text": "就让这些专家基于克劳德答案进行纠偏", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "就让这些专家基于", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 723, "start": 1624.766, "end": 1626.933, "text": "克劳德答案进行纠偏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 724, "start": 1626.933, "end": 1628.866, "text": "而这些专家的纠偏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 725, "start": 1629.0, "end": 1631.433, "text": "就是非常有价值的输入", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 726, "start": 1631.8, "end": 1634.833, "text": "Claude就可以根据这些内容去总结", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 727, "start": 1634.933, "end": 1636.5, "text": "什么情况下应该这样做", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 728, "start": 1636.533, "end": 1638.3, "text": "什么情况下不能这么做", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 729, "start": 1638.433, "end": 1639.233, "text": "什么是对的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 730, "start": 1639.266, "end": 1640.066, "text": "什么是错的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 731, "start": 1640.066, "end": 1641.766, "text": "什么是优秀答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 732, "start": 1641.766, "end": 1643.3, "text": "什么是理想答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 733, "start": 1643.366, "end": 1644.8, "text": "然后在这个领域", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 734, "start": 1644.866, "end": 1646.4, "text": "你就形成了你的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 735, "start": 1646.4, "end": 1648.766, "text": "一个比较黄金的data set", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 736, "start": 1648.9, "end": 1650.433, "text": "形成了一个数据集", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 737, "start": 1650.7, "end": 1651.733, "text": "形成了知识库", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 738, "start": 1651.766, "end": 1653.966, "text": "然后你就可以用国产模型小", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 739, "start": 1654.0, "end": 1656.566, "text": "模型或者一些agent来跑了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 740, "start": 1656.633, "end": 1658.633, "text": "这个也是一种非常", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 741, "start": 1658.633, "end": 1660.333, "text": "深度的新型的蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 742, "start": 1660.366, "end": 1662.2, "text": "就是你蒸馏的不纯粹", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 743, "start": 1662.2, "end": 1664.533, "text": "的是Claude本身的能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 744, "start": 1664.6, "end": 1665.966, "text": "你蒸馏的是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 745, "start": 1666.266, "end": 1668.566, "text": "企业数据加上Claude能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 746, "start": 1668.6, "end": 1671.4, "text": "再加上人类的专家判断", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 747, "start": 1671.4, "end": 1674.0, "text": "三者融合之后的能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 748, "start": 1674.1, "end": 1676.333, "text": "而这个价值本身", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 749, "start": 1676.433, "end": 1680.1, "text": "它比Claude直接蒸馏的意义是更大的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 750, "start": 1680.233, "end": 1683.233, "text": "它是在垂直领域非常专业的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 751, "start": 1683.266, "end": 1684.2, "text": "像这些应用啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 752, "start": 1684.2, "end": 1686.233, "text": "就可以用到很多具体的领域", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 753, "start": 1686.433, "end": 1689.6, "text": "包括写作设计法律咨询", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 754, "start": 1689.6, "end": 1692.4, "text": "企业等等内部的专业领域", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 755, "start": 1692.4, "end": 1693.433, "text": "其实在这个领域啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 756, "start": 1693.433, "end": 1695.266, "text": "会有非常多的现实案例", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 757, "start": 1695.266, "end": 1697.4, "text": "可以实现这样的智能蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 758, "start": 1697.433, "end": 1699.8, "text": "我给大家讲两个具体的场景", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 759, "start": 1699.8, "end": 1701.066, "text": "相信对大家有启发", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 760, "start": 1701.066, "end": 1702.966, "text": "你会知道怎么去做蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 761, "start": 1703.0, "end": 1704.7, "text": "怎么去提升自己的生产力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 762, "start": 1704.866, "end": 1705.766, "text": "第一个场景", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 763, "start": 1705.8, "end": 1707.6, "text": "是一个设计师的场景", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 764, "start": 1707.733, "end": 1710.6, "text": "怎么让AI蒸馏出这个", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 765, "start": 1710.6, "end": 1712.566, "text": "设计师的审美和洞察", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 766, "start": 1712.733, "end": 1713.9, "text": "那首先呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 767, "start": 1713.933, "end": 1718.1, "text": "我给这个AI100张这个设计师经手的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 768, "start": 1718.1, "end": 1720.866, "text": "确认的认可的100张设计图", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 769, "start": 1721.033, "end": 1724.366, "text": "然后让AI来蒸馏这些图", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 770, "start": 1724.366, "end": 1727.166, "text": "的生成的逻辑和方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 771, "start": 1727.3, "end": 1728.7, "text": "他们做的好的地方", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 772, "start": 1728.733, "end": 1730.3, "text": "为什么要这样做等等", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 773, "start": 1730.333, "end": 1731.433, "text": "蒸馏完成之后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 774, "start": 1731.6, "end": 1733.6, "text": "然后按照这个逻辑", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 775, "start": 1733.8, "end": 1735.933, "text": "来生成一系列的图", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 776, "start": 1736.133, "end": 1737.766, "text": "然后把生成好的图", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 777, "start": 1737.9, "end": 1739.866, "text": "再丢给这个设计师", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 778, "start": 1740.0, "end": 1742.3, "text": "让设计师来挑选确认", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 779, "start": 1742.4, "end": 1743.933, "text": "并且最好是标注一下", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 780, "start": 1744.066, "end": 1745.366, "text": "为什么这张图ok", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 781, "start": 1745.366, "end": 1746.733, "text": "为什么这张图不行", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 782, "start": 1746.766, "end": 1749.766, "text": "那人确认的这个挑选结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 783, "start": 1749.8, "end": 1754.566, "text": "的反馈本身就是一种蒸馏对不对", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 784, "start": 1754.6, "end": 1757.433, "text": "然后AI继续优化逻辑", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 785, "start": 1757.533, "end": 1759.5, "text": "继续生成新的图", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 786, "start": 1759.633, "end": 1762.566, "text": "然后这个逻辑又不断的上升了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 787, "start": 1762.633, "end": 1765.766, "text": "它生图的整个的逻辑又升级了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 788, "start": 1765.8, "end": 1767.666, "text": "又不断的有人反馈", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 789, "start": 1767.666, "end": 1769.166, "text": "那这个人的反馈", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 790, "start": 1769.166, "end": 1770.2, "text": "人的纠偏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 791, "start": 1770.4, "end": 1773.2, "text": "越到后面会变得越来越少", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 792, "start": 1773.2, "end": 1775.6, "text": "你会发现在这个过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 793, "start": 1775.833, "end": 1779.966, "text": "人AI它实际上是在双向蒸馏的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 794, "start": 1779.966, "end": 1781.866, "text": "有了这样的蒸馏思路之后啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 795, "start": 1781.966, "end": 1783.433, "text": "甚至我们都能够基于", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 796, "start": 1783.466, "end": 1785.333, "text": "结果来反向的蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 797, "start": 1785.5, "end": 1788.266, "text": "就以大家现在看到的这个文创", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 798, "start": 1788.3, "end": 1790.433, "text": "产品协同的设计方案为例啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 799, "start": 1790.8, "end": 1794.866, "text": "这是我给到AI的一个完整的方案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 800, "start": 1794.866, "end": 1796.233, "text": "我把这个PPT给到它", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 801, "start": 1796.466, "end": 1799.533, "text": "我让它给我反向的抽象", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 802, "start": 1799.766, "end": 1803.333, "text": "蒸馏出这个文创产品方案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 803, "start": 1803.366, "end": 1806.433, "text": "它形成的生产路径是什么", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 804, "start": 1806.8, "end": 1810.733, "text": "我可以以果为因来反向蒸馏过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 805, "start": 1810.733, "end": 1813.666, "text": "是不是就和蒸馏的逆向", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 806, "start": 1813.7, "end": 1815.966, "text": "工程的过程是一样的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 807, "start": 1816.066, "end": 1818.433, "text": "这里是我蒸馏的初步的结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 808, "start": 1818.6, "end": 1821.066, "text": "我实际上是可以继续的迭代去做的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 809, "start": 1821.133, "end": 1823.566, "text": "然后再加一下一些人的反馈", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 810, "start": 1823.6, "end": 1825.1, "text": "那么以后我就可以", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 811, "start": 1825.133, "end": 1826.933, "text": "通过这样的逆向工程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 812, "start": 1827.0, "end": 1829.1, "text": "把它蒸馏出来整个的生产过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 813, "start": 1829.2, "end": 1831.666, "text": "然后再反向的推回去", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 814, "start": 1831.733, "end": 1833.366, "text": "我可以基于它的推导", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 815, "start": 1833.5, "end": 1835.666, "text": "帮我实际的生成验证", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 816, "start": 1835.8, "end": 1838.266, "text": "然后我再不断的反馈和迭代", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 817, "start": 1838.333, "end": 1841.666, "text": "最后这个整个的方案的生产过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 818, "start": 1841.733, "end": 1843.966, "text": "是不是就可以被我推导", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 819, "start": 1844.2, "end": 1845.833, "text": "然后蒸馏出来了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 820, "start": 1845.9, "end": 1848.633, "text": "再说一个比较实用的现实的场景", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 821, "start": 1848.666, "end": 1850.866, "text": "比如电商公司要做投放", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 822, "start": 1851.033, "end": 1853.466, "text": "你就可以让AI来生成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 823, "start": 1853.466, "end": 1855.133, "text": "一系列的投放计划", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 824, "start": 1855.466, "end": 1857.133, "text": "然后这个投放计划你可以", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 825, "start": 1857.166, "end": 1859.633, "text": "交给专门的投手去审核", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 826, "start": 1859.866, "end": 1861.333, "text": "这个投手就会判断", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 827, "start": 1861.333, "end": 1863.066, "text": "哪些计划是可行的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 828, "start": 1863.133, "end": 1864.966, "text": "哪些计划是不可行的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 829, "start": 1865.5, "end": 1866.5, "text": "对于公司来说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 830, "start": 1866.533, "end": 1868.5, "text": "你可以要求这些投手", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 831, "start": 1868.5, "end": 1871.266, "text": "标注它可行不可行的原因", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 832, "start": 1871.433, "end": 1873.1, "text": "合理不合理的原因", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 833, "start": 1873.133, "end": 1874.2, "text": "人做出反馈", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 834, "start": 1874.233, "end": 1877.0, "text": "标注好之后再给到AI去", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 835, "start": 1877.033, "end": 1879.733, "text": "调整这些投放计划而且", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 836, "start": 1880.066, "end": 1882.933, "text": "这些AI整合出来的投放计划", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 837, "start": 1882.966, "end": 1884.066, "text": "最后投出去", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 838, "start": 1884.2, "end": 1886.633, "text": "它会有数据反馈的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 839, "start": 1886.633, "end": 1888.933, "text": "最后投出来的效果好不好", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 840, "start": 1889.0, "end": 1891.366, "text": "人还可以去做评价", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 841, "start": 1891.433, "end": 1894.133, "text": "去说明对应的原因和反馈", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 842, "start": 1894.2, "end": 1897.033, "text": "然后AI就可以根据人的意见", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 843, "start": 1897.2, "end": 1900.066, "text": "以及投放的数据反馈再", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 844, "start": 1900.066, "end": 1903.033, "text": "进一步生成新的投放组合", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 845, "start": 1903.266, "end": 1904.4, "text": "在这个过程中", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 846, "start": 1904.7, "end": 1908.1, "text": "投手的意见相当于是专家的意见", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 847, "start": 1908.3, "end": 1910.4, "text": "是不是就被蒸馏了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 848, "start": 1910.566, "end": 1914.1, "text": "AI对投放这个事情的理解", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 849, "start": 1914.366, "end": 1916.4, "text": "是不是也被蒸馏了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 850, "start": 1916.433, "end": 1917.533, "text": "而在这个过程中", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 851, "start": 1917.533, "end": 1921.033, "text": "可以不断的用一些高等级的模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 852, "start": 1921.133, "end": 1924.6, "text": "对标准答案的回答进行沉淀", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 853, "start": 1924.866, "end": 1927.1, "text": "告诉哪些是好的答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 854, "start": 1927.233, "end": 1928.633, "text": "哪些是坏的答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 855, "start": 1928.8, "end": 1932.4, "text": "它的整个的data set就这样沉淀下来了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 856, "start": 1932.433, "end": 1934.2, "text": "这样就可以不断的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 857, "start": 1934.2, "end": 1936.566, "text": "无限循环和loop下去", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 858, "start": 1936.833, "end": 1939.6, "text": "实际上这个过程再", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 859, "start": 1939.6, "end": 1941.266, "text": "拓散一下是什么呢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 860, "start": 1941.3, "end": 1943.266, "text": "很多高级的模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 861, "start": 1943.466, "end": 1945.666, "text": "也是这么训练出来的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 862, "start": 1945.666, "end": 1947.2, "text": "我们再发散一步", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 863, "start": 1947.233, "end": 1950.166, "text": "是不是可以继续做AI来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 864, "start": 1950.2, "end": 1953.533, "text": "蒸馏AI组织的协作方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 865, "start": 1953.566, "end": 1956.9, "text": "在未来agent可能越来越成熟了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 866, "start": 1956.9, "end": 1959.333, "text": "那么一个很重要的问题可能", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 867, "start": 1959.333, "end": 1962.666, "text": "是怎么把一个优秀的agent的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 868, "start": 1962.666, "end": 1965.033, "text": "协作方式把它蒸馏出来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 869, "start": 1965.033, "end": 1967.566, "text": "比如说在一个协作体系里面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 870, "start": 1967.633, "end": 1968.9, "text": "它有计划的agent", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 871, "start": 1969.266, "end": 1970.666, "text": "有研究的agent", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 872, "start": 1970.933, "end": 1972.4, "text": "有写代码的agent", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 873, "start": 1972.666, "end": 1974.433, "text": "有复盘的agent", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 874, "start": 1974.633, "end": 1975.933, "text": "有做测试的agent", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 875, "start": 1976.0, "end": 1977.4, "text": "还有审阅的agent", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 876, "start": 1977.833, "end": 1981.466, "text": "把整个的团队方式运行10万次", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 877, "start": 1981.6, "end": 1984.466, "text": "然后去发现什么任务它是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 878, "start": 1984.5, "end": 1987.3, "text": "需要planner计划这个角色的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 879, "start": 1987.333, "end": 1990.566, "text": "什么时候需要reviewer复盘者介入", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 880, "start": 1991.1, "end": 1993.766, "text": "什么时候应该让贵的模型介入", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 881, "start": 1993.9, "end": 1995.3, "text": "什么时候可以选择", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 882, "start": 1995.366, "end": 1997.366, "text": "性价比的模型就可以了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 883, "start": 1997.4, "end": 2001.066, "text": "最后蒸馏出来的是一种agent组织", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 884, "start": 2001.1, "end": 2004.5, "text": "协同的一种方式规则和框架", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 885, "start": 2004.533, "end": 2006.9, "text": "再回过头来看我们开始的那句话", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 886, "start": 2007.033, "end": 2007.966, "text": "蒸馏一切", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 887, "start": 2008.1, "end": 2009.466, "text": "是不是挺对的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 888, "start": 2009.566, "end": 2011.2, "text": "我们可以把蒸馏一切啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 889, "start": 2011.2, "end": 2012.6, "text": "分为5个层级", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 890, "start": 2012.933, "end": 2016.033, "text": "整理成一张非常漂亮的图谱", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 891, "start": 2016.033, "end": 2018.8, "text": "这是一条非常优美的演化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 892, "start": 2018.966, "end": 2022.066, "text": "从蒸馏答案获得大模型的结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 893, "start": 2022.2, "end": 2025.2, "text": "再到蒸馏思考获得推理链", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 894, "start": 2025.2, "end": 2027.5, "text": "再接着可以蒸馏行动", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 895, "start": 2027.566, "end": 2028.433, "text": "再深入一点", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 896, "start": 2028.433, "end": 2031.766, "text": "我们可以蒸馏垂直的工作方法", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 897, "start": 2031.766, "end": 2032.633, "text": "形成skill", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 898, "start": 2032.933, "end": 2034.366, "text": "形成skill、SOP、prompt、memory", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "SOP prompt memory", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 899, "start": 2034.466, "end": 2035.166, "text": "进一步的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 900, "start": 2035.2, "end": 2036.4, "text": "我们对对象", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 901, "start": 2036.433, "end": 2039.333, "text": "对人和组织进行蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 902, "start": 2039.533, "end": 2041.5, "text": "形成我们的agent system", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 903, "start": 2041.566, "end": 2042.633, "text": "蒸馏的本质", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 904, "start": 2042.633, "end": 2046.1, "text": "实际上是把一次昂贵的智能过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 905, "start": 2046.2, "end": 2049.6, "text": "变成下一次相对廉价可以复用的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 906, "start": 2049.633, "end": 2050.766, "text": "能力所以啊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 907, "start": 2050.8, "end": 2053.8, "text": "我们万事万物皆可蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 908, "start": 2053.833, "end": 2055.533, "text": "Claude的智能可以蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 909, "start": 2055.766, "end": 2059.133, "text": "Codex的computer use的执行过程可以蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 910, "start": 2059.166, "end": 2061.966, "text": "一个顶级的专家的认知可以蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 911, "start": 2062.033, "end": 2064.633, "text": "一套组织能力可以蒸馏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 912, "start": 2064.633, "end": 2067.733, "text": "我觉得这才是蒸馏让普通人", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 913, "start": 2067.733, "end": 2070.6, "text": "能够产生无限想象力的地方", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 914, "start": 2070.7, "end": 2072.1, "text": "蒸馏你搞懂了吗", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 915, "start": 2072.266, "end": 2074.066, "text": "还有哪些概念或者训练", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 916, "start": 2074.066, "end": 2076.133, "text": "方法你希望去深入的了解", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 917, "start": 2076.233, "end": 2077.466, "text": "希望跟我探讨的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 918, "start": 2077.466, "end": 2079.433, "text": "欢迎在评论区跟我留言", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 919, "start": 2079.466, "end": 2082.1, "text": "如果觉得我的这个视频有趣有料", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 920, "start": 2082.166, "end": 2083.033, "text": "喜欢我的视频", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 921, "start": 2083.066, "end": 2084.733, "text": "也欢迎订阅我的频道", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 922, "start": 2084.833, "end": 2085.933, "text": "灵姐说AI", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 923, "start": 2085.933, "end": 2087.1, "text": "打开你的小铃铛", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 924, "start": 2087.133, "end": 2089.4, "text": "我们下期节目再见啦拜拜", "source": "video_source_subtitle", "authority": "video_source_subtitle" } ], "transcript_selection": { "schema": "videonote.transcript_selection.v1", "status": "SELECTED", "requested_language": "zh", "duration": 2090.0, "reference_path": "/Users/sagawa/AI_WORK/video_notes/url/20260922_100639__一条视频讲透_蒸馏_Claude为什么怕蒸馏_普通人如何自我_智能蒸馏/reference_subtitle.vtt", "reference_kind": "manual", "reference_language": "zh", "selected_source": "video_source_subtitle", "source_caption_kind": "manual", "source_caption_processing": "verbatim_direct", "youtube_overlap_clamps": [], "general_high_confidence_repairs": [], "decision_reason": "complete_source_subtitle_fast_path_no_whisper", "whisper_rejected": false, "source_profile": { "segment_count": 925, "duration_coverage": 0.958339 } }, "presegmentation_correction": { "applied": true, "correction_count": 14, "deterministic_fix_count": 0 } }