{ "text": "离开你 谁还把我当小孩教\n最近国内顶尖高校上海交大\n在GitHub上开源了一套动手学AI的实操教程\n这套教程没有任何水分 全是硬核干货\n完整学下来就能在自己电脑上部署专属于自己的大模型\n不用依赖外网 隐私安全更有保障\n该教程由张卓胜教授主导\n凝聚了多位业内专家的智慧结晶\n非常适合小白学习\n从最简单的API调用\n一步步深入到模型部署与微调\n甚至连模型安全防御\n以及自动化智能体开发都交给你了\n这套教程属于是\n直接把知识味到你嘴里\n知道大家懒得去找\n我已经把配套资源和学习路线整理好了\n照着学就行\n留下学习直接暴走\n劝你别再盲目自学AI开发了\n看了几百个教程\n收藏了几十个热门工作流\n最后自己想做个自动化工具还是跑不通\n为什么\n因为大模型这个东西\n它的思维逻辑\n跟我们传统的软件开发是完全相反的\n以前是一加一肯定等于二\n但现在大模型给你的\n是一个大体上不错\n但是偶尔会瞎编的一个概率\n如果你搞不懂怎么去规避这种不确定性\n学再多的框架也是在做无用功\n想要做出一款真正能自动干活\n甚至能拿去变现的AI产品\n其实关键就在于四项能力\n这四项能力也直接决定了\n你做出来的到底是个不好用的玩具\n还是能真正落地的商业产品\n它们分别是任务拆解能力\n工具调用能力\n评测和可观测性\n以及最硬核的生产环境能力\n接下来我们一个一个讲清楚\n我们做大模型应用开发\n接手一个新项目的时候\n第一件事是要干什么\n其实不是急着去写代码\n也不是去挑什么酷炫的框架\n我们最先要做的\n就是判断这个业务到底有多复杂\n然后给它配一个最合适的技术架构\n你看我们平时做项目\n最容易犯两个错\n一种叫做技术降维\n就是不管业务有多复杂\n我们都只想用一段prompt解决\n结果模型疯狂产生幻觉\n另一种叫做过度设计\n明明是个很简单的事情\n非要套一个很繁琐的智能体框架\n这两种做法在企业里其实都是要踩坑的\n那为了精准匹配\n我们一般能把业务复杂度分成了三档\n第一档是那种低复杂度\n不需要记住状态的任务\n比如你想要模型帮你写一封桌报邮件\n或者让他帮忙看一下这段代码里面有没有拼写错误\n这种一次性的交互\n我们直接用Damp Prompt方案就行了\n这就好比我们在路上问路\n电铁站怎么走\n对方给你指个方向\n这事就结了\n不需要复杂的逻辑单次解决\n但是如果我们的任务变复杂了\n到了第二档\n我们需要在整个过程中\n保持很高的一个可控性\n比如说我们要让模型\n帮我们审批合同\n或者去跑一个标准的财务报表\n这个时候呢\nDamp Prompt肯定顶不住\n我们需要使用Workflow工作流方案\n这就像我们去银行里面办业务\n先取号填表柜台办理最后签字\n每一步怎么走\n节点之间怎么传递信息\n都是定死的显性的\n这种流逝的步骤设计\n最适合用来跑标准的SOP\n出来的结果非常稳定\n那再往上走\n到了第三档\n就是高复杂度\n长周期\n需要多角色分工的混沌任务了\n比如我们想要大模型\n去写一个完整的软件模块\n这个里面设计的需求就太多了\n我们必须用多智能体协同的方案\n那在这个架构下\n我们需要引入\n人机协同的安全红线\n比如智能体写完了代码\n在准备部署上线的那一步\n必须停下来\n等待人工审核确认\n确认过了才能继续\n同时\n各个智能体之间\n还有自己的独立上下文\n不能够把信息给混在一起\n不然模型自己就糊涂了\n那么面对这种高复杂度的\n多智能体任务\n我们在工程上\n到底该怎么去编排\n怎么设计\n才能够保证它输出的质量呢\n其实现在行业里\n有一个挺成熟的思路\n是分布循环\n加上格力子智能体\n听起来是有点专业\n我们拆开来聊\n其实非常简单\n先说分布循环\n就是把任务交给模型之后\n它一上来又直接开始写\n结果写出来的东西完全不能用\n所以我们需要在框架上来限制它的行为\n比如我们可以采用一个探索规划执行的一个循环\n第一步是探索\n这能体现去读取相关的文档或者搜索代码库\n所以在这个阶段我们只给它只读的权限\n它只能看不能改\n第二步是规划\n它根据看懂的信息做一个具体的执行计划\n那么在计划做好之后\n我们可以加入人工确认\n也就是人机协同\n人类觉得计划可行再放行\n第三步才是执行\n他开始去写文件\n运行命令\n这个时候他才能拿到完整的操作权限\n执行完之后\n如果发现报错了\n他会自动回到第一步重新探索\n这样一套循环下来\n任务就会被框得非常稳\n那么除了这个循环\n还有一个更关键的点\n叫做上下文隔离子智能体\n你看如果一个智能体要做的事情太多\n它的上下文里就会塞满各种各样信息\n这就好比我们自己开一家公司\n你不能让一个人既当销售又当财务\n还要当程序员对吧\n信息一多他肯定会记错\n所以呢我们要设立岗位\n比如我们有一个主智能体\n他手底下呢有三个上下文隔离的子智能体\n一个专门找资料的搜索智能体\n一个做计划的规划智能体\n一个负责写代码的执行智能体\n最关键的是\n他们每个人只拥有自己专属的上下文\n写代码的不知道财务数据\n找资料的也不需要管代码里的bug\n他们之间只通过主智能体进行最标准最干净的数据传递\n通过这种各干各的相互隔离的设计\n每个子智能体面对的信息量都会极大减少\n任务被彻底拆干净了\n模型的准确率自然就上去了\nOK 聊完了业务怎么拆解\n接下来我们讲第二个核心能力\n工具调用的能力\n其实大模型项目在线上翻车\n绝大部分问题都不是出在模型不够聪明上\n而是我们的工具设计出了问题\n我们很多刚入行做应用开发的朋友\n在让大模型去修改文件或者运行命令的时候\n特别喜欢给他一个全能的Mesh工具\n也就是我们今天常说的命令行\n大模型想要读文件\n改代码\n搜索关键词\n全都通过这一个命令行\n去敲各种指令来解决\n他们觉得这样模型很自由很省事对吧\n但实际上在工业级开发里\n这是一个非常危险的反面模式\n为什么这么说呢\n你想想\n如果大模型手里拿的是一个\n可以为所欲为的万能命令行\n它能干的事情就太多了\n一旦它某一次产生了幻觉\n或者拼错了一个字符\n就非常容易拼接出一个\n把系统文件彻底弄坏的错误指令\n而且作为开发者\n你怎么去审查它的权限\n你根本没有办法\n给它做精细化的安全授权\n这就像是你给新来的实习生\n配了一把万能钥匙\n能拔开公司所有的门\n这个安全隐患应该是非常大的\n所以我们更提倡一种\n单一职责工具模式\n简单来说\n就是把这些大而全的命令行\n拆成一个个只盖一键小式的专业工具\n读文件的工具就是read\n改文件的呢就是edit\n查找关键词的工具是grab\n写文件的工具呢是write\n这样设计的好处显而易见\n每一个工具都有非常严格的参数规范\n也就是schema\n大模型在调用的时候\n边界非常明确\n而且我们还可以给不同的工具分配独立的权限\n接口的确定性变高了\n模型的理解偏差自然就小了\n那这时候新的问题就来了\n随着我们的项目越做越大\n大模型需要用的工具越来越多\n我们该怎么去管理和扩展这些工具呢\n另外如果大模型确实需要跑一些像删除文件\n还有推送代法这些高风险的命令\n我们又该怎么控制它的安全风险呢\n在实际工程落地中\n我们一般要遵循两个原则\n第一个原则叫渐进式工具扩展\n什么意思呢\n就是我们一开始\n千万别一上来就把几十个上百个工具\n全都喂给大木型\n工具给的太多\n大木型在做选择的时候\n非常容易挑花眼\n最后选错工具\n刚开始的时候\n我们本地只给他提供\n一二十个最常用最基础的工具\n像读写修改就完全够了\n如果后面我们发现需要接入一些外部系统\n比如要读公司的数据库\n要调用外部的第三方服务\n那我们再加上MCP\n也就是模型上下文协议\n标准化的把这些外部工具给加载进来\n如果再往后需要调用远程云端的主机\n或远程的数据库\n我们再介入远程工具\n一步一步按需加载\n这样模型的工具调用准确率才不会崩\n那么第二个原则\n是必须在底层建立一个安全杀箱\n做好命令的风险分级\n大模型它毕竟是模型\n它有时候可能会失控\n所以我们必须给它调用的所有命令定下规则\n分出安全等级\n比如像跑个测试或者看一下文件差异\n这些完全是安全的命令\n我们可以设置为自动执行\n不需要打扰用户\n但如果是像强推代码这样的操作\n这就有一定的风险了\n系统在跑这个命令之前\n必须暂停下来\n弹出来要用户确认一下\n必须等我们人肉确认了\n他才能跑\n而如果是像这种\n会煽光系统的毁灭性指令\n不用问我们在沙箱底层\n直接硬变码彻底拉黑\n不给他任何运行的机会\n你看我们只有在底层\n把这一套分级拦截的沙箱机制打好了\n才能真正放心的把工具\n交到大模型手里去跑对吧\n那接下来我们讲第三个很容易被忽略\n但是却决定了项目能不能真正交付的能力\n评测和可观测性\n平时我们自己调模型的时候\n经常会想\n我测了几次感觉效果还挺不错的\n但是企业级最怕的就是感觉这两个字\n你改了一个prompt或者换了一个模型\n你感觉变好了\n但可能在某个你没测到的业务角落\n它其实退化了\n这在工程上叫做静默退化\n你根本不知道它什么时候\n在什么地方变差了\n所以我们必须告别主观体感\n建立一套体系化的量化评测流程\n这就像去医院体检\n不能只靠医生问你感觉怎么样\n而是要抽血化验\n看各项指标的数值\n具体在开发里\n这个流程分为三步\n第一步我们要构建一个黄金数据集\n这个数据集要覆盖我们最核心\n最真实的业务场景\n比如包含50个或者100个经典的客户提问\n和对应的标准答案\n这就是我们的考试试卷\n第二步有了试卷\n我们得找个客观的乐卷老师来自动进行打分\n我们一般会引入一个强推力模型\n让大模型当裁判\n也就是业内常说的\nLM as a judge\n这样每次我们改的代码\n直接自动化跑一遍考试\n摆脱人工人肉测试的低效\n那第三步看成绩单\n这个成绩单不能只有一个总分\n而是要有一个多维度的指标体系\n比如我们要看幻觉率有没有上升\n模型给出的答案跟问题相不相关\n检索知识库的召回率高不高\n还有最实际的响应耗时有没有变慢\n通过这样的一套闭环评测\n我们每次迭代升级\n心里才是有底的对吧\n有了这一套评测流程\n只能说明我们的应用在模拟考试的时候\n表现还不错\n但如果项目真的上线了\n用户在使用过程中遇到了爆错\n或者突然输出了一堆乱码\n我们该怎么排查呢\n很多项目上线之后\n一旦爆错了\n开发者就懵了\n因为大模型应用就像个黑盒\n你根本不知道中间哪一步出了问题\n所以我们必须要做全链路可观测性设计\n每次推理每次工具调用\n它的整个生命周期都必须被完整的记录下来\n首先在输入阶段\n我们要捕获最原始的prompt和系统当时的状态\n接着在推理中间菜\n我们要监控模型的思维链\n看它是怎么一步步思考\n同时记录它消耗了多少token\n这直接关系到我们的账单成本\n然后在工具调用阶段\n大模型传了什么参数\n调用了什么接口\n接口响应花了多少时间\n也要记下来\n最后在输出阶段\n如果发生了异常\n比如说超时了呀\n或者是格式不对\n或者触发了我们的服务降级策略\n系统也要能够精准的捕捉\n把这些数据完整的记录下来之后\n最核心的好处\n就是我们可以进行一次\n故障复现和回放调试\n比如线上用户反馈说\n模型刚才瞎编了\n我们不需要去猜\n直接把那次调用的数据给导出来\n在本地重新跑一遍\n看看大模型在思维链的哪一步走偏了\n或者是调工具的时候\n传哪个参数传错了\n然后我们就可以精准优化prompt\n或者加上超时重视\n格式纠错这样的容错机制\n这才是真正的工程化调试思维\n好\n有了这个评测和可观测性\n我们的应用在开发阶段\n已经看起来很健康了\n但是如果要真正推上线上\n交给成千上万的用户去用\n这就到了最拉开差距的地方\n我们的第四个核心能力\n生产环境能力\n生产环境里有一个非常实际的痛点\n就是模型特别健忘\n而且绘画越长\n上下文塞的越满\n不仅模型响应变得越来越慢\n公司的token资费也会像流水一样花出去\n我们不能指望用户每次跟智能体聊天\n都把项目规则和历史对话\n复制粘贴一遍吧\n所以在生产及状态管理里\n我们首先要在代码库里\n放一个持久化的配置指令文件\n比如说cloud.md\n这个文件是随我们的代码库\n一起做版本控制的\n当绘画启动时\n系统会自动把项目级的规范\n测试命令自动加载进去\n模型一上线\n立刻就知道自己身处什么项目\n需要遵守什么规则\n总进一步呢\n我们要建立一个分层记忆机制\n对对上计算机有缓存\n有内存\n有硬盘\n第一层是永远加载的核心索引\n这部分一般控制在200行以内\n保持极低的响应延迟\n然后第二层是按需加载的相关文档和当前状态\n比如模型当前要修改某个代码文件\n我们就只要把这个文件和当前的锻点状态加载进来\n这里非常重要的一点\n是支持断点恢复\n万一服务中断了\n智能体重启之后\n能接着上次的进度干\n不用从头跑\n那第三层\n则是只支持检索的\n完整历史对话转录\n不用到的历史记录\n全部打包归档\n只有需要的时候才去搜索\n通过这种分层\n我们既能处理长周期的任务\n又能帮公司省下一大笔的投肯开销\n不过有了分层记忆\n如果用户一直跟智能体聊下去\n聊了上百轮\n当前的绘画窗口\n还是会面临一个爆满的风险\n这时候我们要怎么防止上下纹爆掉呢\n其实智能体在这一方面和我们人类很像\n人总不能一直不睡觉拼命公渡对吧\n所以我们可以给智能体引入一个\n记忆睡眠巩固机制\n在系统空闲的时候默默运行\n我们可以让一个后台智能体\n在用户不说话的闲时\n去帮我们整理零段的上下纹\n它会自动把过期的冲突的事实给修剪掉\n去虫之后合并成一个干净的核心索引\n这就像我们睡觉做梦的时候\n大脑会自动整理白天的记忆\n丢弃无用信息一样\n这是一种应用的自我进化\n同时在绘画进行中\n我们还要配合渐进式上下文压缩\n随着对话轮次的增加\n我们不是粗暴的直接截掉前面的对话\n而是施加不同强度的压缩\n近期对话保留最详细的明气\n中期对话做轻度总结\n而远期对话则进行一个极限折叠\n这样层层递进\n既保证了模型不会失忆\n又彻底解决了上下文窗口爆棚的问题\nOK\n在生产环境里\n除了记忆管理\n我们还要解决两个问题\n一个是系统的安全合规\n另一个是高昂的API障碍成本\n首先是系统安全\n我们不能全指望prompt去叮嘱大模型\n你不要违规\n格式千万要写对\n因为模型总有遗忘的时候\n我们的原则是\n必须把硬逻辑剥离到prompt\n用一个确定性的生命周期钩子。\n比如,在大模型调用外部工具之前,\n系统在底层强行的插入一个前置钩子,\n自动降业命令安不安全。\n工具跑完了之后,\n我们再强行插入一个后置钩子,\n自动去跑一次代码格式化和测试\n从再配置文件\n这些钩子是在系统底层用代码硬编码的\n不管大模型怎么产生幻觉\n这些安全红线它绝对越不过去\n然后就是商业化成本的控制\n如果每天接上万个用户请求\n所有请求都用最贵最强的大模型\n公司可能很快就撑不住了\n所以我们要设计一个混合模型路由机制\n当用户的请求进来的时候\n先过一个路由节点\n做动态难度分发\n如果是写系统架构这种高难度任务\n路由就把请求分发给大型颗粒模型\n那如果是问路打招呼\n或者是简单做一个格式化这种任务\n路由就自动分流给急速小模型\n或者直接命中我们的提示词缓存\n这样既兼顾了系统的高并发响应\n又帮我们省下了大量的API资费\n这才是合格的商业级架构\n而在高频的业务场景里\n我们最常碰到的落地项目\n其实还是企业知识库\n也就是RG系统\n但是在生产环境里\n一个合格的RG绝对不是单向的\n检索完就丢给模型\n它应该是一个能自我进化的闭环\n在生产环境里\n我们要打通混合检索和重排架构\n形成一个数据沉淀模型反母的飞轮\n第一步混合检索\n当用户提出问题的时候\n我们把向量检索和关键词检索\n融合在一起\n保证无论用户怎么提问\n我们都能精准定位底层的原始文档\n然后第二步重排提纯\n搜出来的文档可能有几十个\n那模型看不完也看不过来\n我们用专门的重排模型\n挑选出最精准的几个节点\n送进大幕型的上下文窗口\n接下来第三步\n就是结合业务逻辑\n生成最终的一个融合输出\n重点在第四步\n数据沉淀\n我们要补获用户的真实反馈\n记录高频的查询盲区\n那最后一步\n反补迭代\n根据这些反馈\n自动去更新\n修正和强化我们的企业知识库\n那么当这一套流程运转起来之后\n知识库里的盲区就会越来越少\n下一轮检索就会更准\n我们的应用也会越用越聪明\n好 聊到这里\n我们基本上把一个工业级的\n大模型应用开发工程师\n需要具备的能力全都梳理了一遍\n但最后我们来做一个复盘\n其实一个真正合格\n能够帮企业解决实际大模型落地的工程师\n它的成长路径其实是非常清晰的\n在业务架构上\n我们要从只会写单一prompt\n升级到能够驾驭多智能体编排\n在工具与协议上\n我们要从硬编码施API\n升级到运用标准的AMCP协议\n和单一职责工具\n在生产机状态上\n我们要从粗暴的内存阶段\n升级到分层记忆和动态上下纹压缩\n那在系统级整合上\n我们要从基础的检索\n升级到刚才讲的RG飞轮和模型路由\n然后在质量可观测性上\n升级到全列路追踪和量化评测", "language": "zh", "duration": 1073.806813, "segments": [ { "start": 0.12, "end": 1.7, "text": "离开你 谁还把我当小孩教", "chunk": 0, "language": "zh", "avg_logprob": -0.06179458194308811, "no_speech_prob": 2.722098241259019e-11, "compression_ratio": 1.2699228791773778 }, { "start": 1.7, "end": 3.52, "text": "最近国内顶尖高校上海交大", "chunk": 0, "language": "zh", "avg_logprob": -0.06179458194308811, "no_speech_prob": 2.722098241259019e-11, "compression_ratio": 1.2699228791773778 }, { "start": 3.52, "end": 6.44, "text": "在GitHub上开源了一套动手学AI的实操教程", "chunk": 0, "language": "zh", "avg_logprob": -0.06179458194308811, "no_speech_prob": 2.722098241259019e-11, "compression_ratio": 1.2699228791773778 }, { "start": 6.44, "end": 8.84, "text": "这套教程没有任何水分 全是硬核干货", "chunk": 0, "language": "zh", "avg_logprob": -0.06179458194308811, "no_speech_prob": 2.722098241259019e-11, "compression_ratio": 1.2699228791773778 }, { "start": 8.84, "end": 12.1, "text": "完整学下来就能在自己电脑上部署专属于自己的大模型", "chunk": 0, "language": "zh", "avg_logprob": -0.06179458194308811, "no_speech_prob": 2.722098241259019e-11, "compression_ratio": 1.2699228791773778 }, { "start": 12.1, "end": 14.26, "text": "不用依赖外网 隐私安全更有保障", "chunk": 0, "language": "zh", "avg_logprob": -0.06179458194308811, "no_speech_prob": 2.722098241259019e-11, "compression_ratio": 1.2699228791773778 }, { "start": 14.26, "end": 16.0, "text": "该教程由张卓胜教授主导", "chunk": 0, "language": "zh", "avg_logprob": -0.06179458194308811, "no_speech_prob": 2.722098241259019e-11, "compression_ratio": 1.2699228791773778 }, { "start": 16.0, "end": 17.94, "text": "凝聚了多位业内专家的智慧结晶", "chunk": 0, "language": "zh", "avg_logprob": -0.06179458194308811, "no_speech_prob": 2.722098241259019e-11, "compression_ratio": 1.2699228791773778 }, { "start": 17.94, "end": 19.22, "text": "非常适合小白学习", "chunk": 0, "language": "zh", "avg_logprob": -0.06179458194308811, "no_speech_prob": 2.722098241259019e-11, "compression_ratio": 1.2699228791773778 }, { "start": 19.22, "end": 20.66, "text": "从最简单的API调用", "chunk": 0, "language": "zh", "avg_logprob": -0.06179458194308811, "no_speech_prob": 2.722098241259019e-11, "compression_ratio": 1.2699228791773778 }, { "start": 20.66, "end": 22.64, "text": "一步步深入到模型部署与微调", "chunk": 0, "language": "zh", "avg_logprob": -0.06179458194308811, "no_speech_prob": 2.722098241259019e-11, "compression_ratio": 1.2699228791773778 }, { "start": 22.64, "end": 24.26, "text": "甚至连模型安全防御", "chunk": 0, "language": "zh", "avg_logprob": -0.06179458194308811, "no_speech_prob": 2.722098241259019e-11, "compression_ratio": 1.2699228791773778 }, { "start": 24.26, "end": 26.32, "text": "以及自动化智能体开发都交给你了", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 26.32, "end": 27.44, "text": "这套教程属于是", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 27.44, "end": 28.74, "text": "直接把知识味到你嘴里", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 28.74, "end": 29.76, "text": "知道大家懒得去找", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 29.76, "end": 31.92, "text": "我已经把配套资源和学习路线整理好了", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 31.92, "end": 32.7, "text": "照着学就行", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 32.7, "end": 34.62, "text": "留下学习直接暴走", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 34.62, "end": 36.78, "text": "劝你别再盲目自学AI开发了", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 36.78, "end": 38.08, "text": "看了几百个教程", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 38.08, "end": 39.68, "text": "收藏了几十个热门工作流", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 39.68, "end": 42.0, "text": "最后自己想做个自动化工具还是跑不通", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 42.0, "end": 42.44, "text": "为什么", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 42.44, "end": 44.36, "text": "因为大模型这个东西", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 44.36, "end": 45.42, "text": "它的思维逻辑", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 45.42, "end": 47.76, "text": "跟我们传统的软件开发是完全相反的", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 47.76, "end": 50.08, "text": "以前是一加一肯定等于二", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 50.08, "end": 51.7, "text": "但现在大模型给你的", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 51.7, "end": 52.88, "text": "是一个大体上不错", "chunk": 0, "language": "zh", "avg_logprob": -0.09819211186589422, "no_speech_prob": 1.984035678803675e-11, "compression_ratio": 1.3638676844783715 }, { "start": 52.88, "end": 54.8, "text": "但是偶尔会瞎编的一个概率", "chunk": 0, "language": "zh", "avg_logprob": -0.08211777589031469, "no_speech_prob": 1.6836946767351257e-11, "compression_ratio": 1.39247311827957 }, { "start": 54.8, "end": 57.56, "text": "如果你搞不懂怎么去规避这种不确定性", "chunk": 0, "language": "zh", "avg_logprob": -0.08211777589031469, "no_speech_prob": 1.6836946767351257e-11, "compression_ratio": 1.39247311827957 }, { "start": 57.56, "end": 59.62, "text": "学再多的框架也是在做无用功", "chunk": 0, "language": "zh", "avg_logprob": -0.08211777589031469, "no_speech_prob": 1.6836946767351257e-11, "compression_ratio": 1.39247311827957 }, { "start": 59.62, "end": 62.08, "text": "想要做出一款真正能自动干活", "chunk": 0, "language": "zh", "avg_logprob": -0.08211777589031469, "no_speech_prob": 1.6836946767351257e-11, "compression_ratio": 1.39247311827957 }, { "start": 62.08, "end": 63.82, "text": "甚至能拿去变现的AI产品", "chunk": 0, "language": "zh", "avg_logprob": -0.08211777589031469, "no_speech_prob": 1.6836946767351257e-11, "compression_ratio": 1.39247311827957 }, { "start": 63.82, "end": 65.64, "text": "其实关键就在于四项能力", "chunk": 0, "language": "zh", "avg_logprob": -0.08211777589031469, "no_speech_prob": 1.6836946767351257e-11, "compression_ratio": 1.39247311827957 }, { "start": 65.64, "end": 67.64, "text": "这四项能力也直接决定了", "chunk": 0, "language": "zh", "avg_logprob": -0.08211777589031469, "no_speech_prob": 1.6836946767351257e-11, "compression_ratio": 1.39247311827957 }, { "start": 67.64, "end": 69.42, "text": "你做出来的到底是个不好用的玩具", "chunk": 0, "language": "zh", "avg_logprob": -0.08211777589031469, "no_speech_prob": 1.6836946767351257e-11, "compression_ratio": 1.39247311827957 }, { "start": 69.42, "end": 71.38, "text": "还是能真正落地的商业产品", "chunk": 0, "language": "zh", "avg_logprob": -0.08211777589031469, "no_speech_prob": 1.6836946767351257e-11, "compression_ratio": 1.39247311827957 }, { "start": 71.38, "end": 73.68, "text": "它们分别是任务拆解能力", "chunk": 0, "language": "zh", "avg_logprob": -0.08211777589031469, "no_speech_prob": 1.6836946767351257e-11, "compression_ratio": 1.39247311827957 }, { "start": 73.68, "end": 74.78, "text": "工具调用能力", "chunk": 0, "language": "zh", "avg_logprob": -0.08211777589031469, "no_speech_prob": 1.6836946767351257e-11, "compression_ratio": 1.39247311827957 }, { "start": 74.78, "end": 76.26, "text": "评测和可观测性", "chunk": 0, "language": "zh", "avg_logprob": -0.08211777589031469, "no_speech_prob": 1.6836946767351257e-11, "compression_ratio": 1.39247311827957 }, { "start": 76.26, "end": 78.3, "text": "以及最硬核的生产环境能力", "chunk": 0, "language": "zh", "avg_logprob": -0.08211777589031469, "no_speech_prob": 1.6836946767351257e-11, "compression_ratio": 1.39247311827957 }, { "start": 78.3, "end": 80.16, "text": "接下来我们一个一个讲清楚", "chunk": 0, "language": "zh", "avg_logprob": -0.08211777589031469, "no_speech_prob": 1.6836946767351257e-11, "compression_ratio": 1.39247311827957 }, { "start": 80.16, "end": 82.3, "text": "我们做大模型应用开发", "chunk": 0, "language": "zh", "avg_logprob": -0.08211777589031469, "no_speech_prob": 1.6836946767351257e-11, "compression_ratio": 1.39247311827957 }, { "start": 82.3, "end": 83.5, "text": "接手一个新项目的时候", "chunk": 0, "language": "zh", "avg_logprob": -0.09014745994850441, "no_speech_prob": 1.9001640538807152e-11, "compression_ratio": 1.4228571428571428 }, { "start": 83.5, "end": 84.94, "text": "第一件事是要干什么", "chunk": 0, "language": "zh", "avg_logprob": -0.09014745994850441, "no_speech_prob": 1.9001640538807152e-11, "compression_ratio": 1.4228571428571428 }, { "start": 84.94, "end": 87.18, "text": "其实不是急着去写代码", "chunk": 0, "language": "zh", "avg_logprob": -0.09014745994850441, "no_speech_prob": 1.9001640538807152e-11, "compression_ratio": 1.4228571428571428 }, { "start": 87.18, "end": 89.06, "text": "也不是去挑什么酷炫的框架", "chunk": 0, "language": "zh", "avg_logprob": -0.09014745994850441, "no_speech_prob": 1.9001640538807152e-11, "compression_ratio": 1.4228571428571428 }, { "start": 89.06, "end": 90.44, "text": "我们最先要做的", "chunk": 0, "language": "zh", "avg_logprob": -0.09014745994850441, "no_speech_prob": 1.9001640538807152e-11, "compression_ratio": 1.4228571428571428 }, { "start": 90.44, "end": 92.34, "text": "就是判断这个业务到底有多复杂", "chunk": 0, "language": "zh", "avg_logprob": -0.09014745994850441, "no_speech_prob": 1.9001640538807152e-11, "compression_ratio": 1.4228571428571428 }, { "start": 92.34, "end": 94.44, "text": "然后给它配一个最合适的技术架构", "chunk": 0, "language": "zh", "avg_logprob": -0.09014745994850441, "no_speech_prob": 1.9001640538807152e-11, "compression_ratio": 1.4228571428571428 }, { "start": 94.44, "end": 96.18, "text": "你看我们平时做项目", "chunk": 0, "language": "zh", "avg_logprob": -0.09014745994850441, "no_speech_prob": 1.9001640538807152e-11, "compression_ratio": 1.4228571428571428 }, { "start": 96.18, "end": 97.3, "text": "最容易犯两个错", "chunk": 0, "language": "zh", "avg_logprob": -0.09014745994850441, "no_speech_prob": 1.9001640538807152e-11, "compression_ratio": 1.4228571428571428 }, { "start": 97.3, "end": 99.22, "text": "一种叫做技术降维", "chunk": 0, "language": "zh", "avg_logprob": -0.09014745994850441, "no_speech_prob": 1.9001640538807152e-11, "compression_ratio": 1.4228571428571428 }, { "start": 99.22, "end": 101.08, "text": "就是不管业务有多复杂", "chunk": 0, "language": "zh", "avg_logprob": -0.09014745994850441, "no_speech_prob": 1.9001640538807152e-11, "compression_ratio": 1.4228571428571428 }, { "start": 101.08, "end": 102.78, "text": "我们都只想用一段prompt解决", "chunk": 0, "language": "zh", "avg_logprob": -0.09014745994850441, "no_speech_prob": 1.9001640538807152e-11, "compression_ratio": 1.4228571428571428 }, { "start": 102.78, "end": 104.68, "text": "结果模型疯狂产生幻觉", "chunk": 0, "language": "zh", "avg_logprob": -0.09014745994850441, "no_speech_prob": 1.9001640538807152e-11, "compression_ratio": 1.4228571428571428 }, { "start": 104.68, "end": 106.94, "text": "另一种叫做过度设计", "chunk": 0, "language": "zh", "avg_logprob": -0.09014745994850441, "no_speech_prob": 1.9001640538807152e-11, "compression_ratio": 1.4228571428571428 }, { "start": 106.94, "end": 108.82, "text": "明明是个很简单的事情", "chunk": 0, "language": "zh", "avg_logprob": -0.09014745994850441, "no_speech_prob": 1.9001640538807152e-11, "compression_ratio": 1.4228571428571428 }, { "start": 108.82, "end": 110.92, "text": "非要套一个很繁琐的智能体框架", "chunk": 0, "language": "zh", "avg_logprob": -0.09014745994850441, "no_speech_prob": 1.9001640538807152e-11, "compression_ratio": 1.4228571428571428 }, { "start": 110.92, "end": 113.8, "text": "这两种做法在企业里其实都是要踩坑的", "chunk": 0, "language": "zh", "avg_logprob": -0.11034669664171007, "no_speech_prob": 1.7233995480703257e-11, "compression_ratio": 1.336734693877551 }, { "start": 113.8, "end": 115.3, "text": "那为了精准匹配", "chunk": 0, "language": "zh", "avg_logprob": -0.11034669664171007, "no_speech_prob": 1.7233995480703257e-11, "compression_ratio": 1.336734693877551 }, { "start": 115.3, "end": 117.72, "text": "我们一般能把业务复杂度分成了三档", "chunk": 0, "language": "zh", "avg_logprob": -0.11034669664171007, "no_speech_prob": 1.7233995480703257e-11, "compression_ratio": 1.336734693877551 }, { "start": 117.72, "end": 119.8, "text": "第一档是那种低复杂度", "chunk": 0, "language": "zh", "avg_logprob": -0.11034669664171007, "no_speech_prob": 1.7233995480703257e-11, "compression_ratio": 1.336734693877551 }, { "start": 119.8, "end": 121.36, "text": "不需要记住状态的任务", "chunk": 0, "language": "zh", "avg_logprob": -0.11034669664171007, "no_speech_prob": 1.7233995480703257e-11, "compression_ratio": 1.336734693877551 }, { "start": 121.36, "end": 123.94, "text": "比如你想要模型帮你写一封桌报邮件", "chunk": 0, "language": "zh", "avg_logprob": -0.11034669664171007, "no_speech_prob": 1.7233995480703257e-11, "compression_ratio": 1.336734693877551 }, { "start": 123.94, "end": 127.58, "text": "或者让他帮忙看一下这段代码里面有没有拼写错误", "chunk": 0, "language": "zh", "avg_logprob": -0.11034669664171007, "no_speech_prob": 1.7233995480703257e-11, "compression_ratio": 1.336734693877551 }, { "start": 127.58, "end": 128.68, "text": "这种一次性的交互", "chunk": 0, "language": "zh", "avg_logprob": -0.11034669664171007, "no_speech_prob": 1.7233995480703257e-11, "compression_ratio": 1.336734693877551 }, { "start": 128.68, "end": 130.54, "text": "我们直接用Damp Prompt方案就行了", "chunk": 0, "language": "zh", "avg_logprob": -0.11034669664171007, "no_speech_prob": 1.7233995480703257e-11, "compression_ratio": 1.336734693877551 }, { "start": 130.54, "end": 132.38, "text": "这就好比我们在路上问路", "chunk": 0, "language": "zh", "avg_logprob": -0.11034669664171007, "no_speech_prob": 1.7233995480703257e-11, "compression_ratio": 1.336734693877551 }, { "start": 132.38, "end": 133.48, "text": "电铁站怎么走", "chunk": 0, "language": "zh", "avg_logprob": -0.11034669664171007, "no_speech_prob": 1.7233995480703257e-11, "compression_ratio": 1.336734693877551 }, { "start": 133.48, "end": 134.84, "text": "对方给你指个方向", "chunk": 0, "language": "zh", "avg_logprob": -0.11034669664171007, "no_speech_prob": 1.7233995480703257e-11, "compression_ratio": 1.336734693877551 }, { "start": 134.84, "end": 135.58, "text": "这事就结了", "chunk": 0, "language": "zh", "avg_logprob": -0.11034669664171007, "no_speech_prob": 1.7233995480703257e-11, "compression_ratio": 1.336734693877551 }, { "start": 135.58, "end": 137.88, "text": "不需要复杂的逻辑单次解决", "chunk": 0, "language": "zh", "avg_logprob": -0.11034669664171007, "no_speech_prob": 1.7233995480703257e-11, "compression_ratio": 1.336734693877551 }, { "start": 137.88, "end": 140.7, "text": "但是如果我们的任务变复杂了", "chunk": 0, "language": "zh", "avg_logprob": -0.11034669664171007, "no_speech_prob": 1.7233995480703257e-11, "compression_ratio": 1.336734693877551 }, { "start": 140.7, "end": 141.46, "text": "到了第二档", "chunk": 0, "language": "zh", "avg_logprob": -0.14661186099670093, "no_speech_prob": 2.236957320767008e-11, "compression_ratio": 1.256267409470752 }, { "start": 141.46, "end": 143.14, "text": "我们需要在整个过程中", "chunk": 0, "language": "zh", "avg_logprob": -0.14661186099670093, "no_speech_prob": 2.236957320767008e-11, "compression_ratio": 1.256267409470752 }, { "start": 143.14, "end": 145.26, "text": "保持很高的一个可控性", "chunk": 0, "language": "zh", "avg_logprob": -0.14661186099670093, "no_speech_prob": 2.236957320767008e-11, "compression_ratio": 1.256267409470752 }, { "start": 145.26, "end": 147.18, "text": "比如说我们要让模型", "chunk": 0, "language": "zh", "avg_logprob": -0.14661186099670093, "no_speech_prob": 2.236957320767008e-11, "compression_ratio": 1.256267409470752 }, { "start": 147.18, "end": 148.66, "text": "帮我们审批合同", "chunk": 0, "language": "zh", "avg_logprob": -0.14661186099670093, "no_speech_prob": 2.236957320767008e-11, "compression_ratio": 1.256267409470752 }, { "start": 148.66, "end": 150.82, "text": "或者去跑一个标准的财务报表", "chunk": 0, "language": "zh", "avg_logprob": -0.14661186099670093, "no_speech_prob": 2.236957320767008e-11, "compression_ratio": 1.256267409470752 }, { "start": 150.82, "end": 152.14, "text": "这个时候呢", "chunk": 0, "language": "zh", "avg_logprob": -0.14661186099670093, "no_speech_prob": 2.236957320767008e-11, "compression_ratio": 1.256267409470752 }, { "start": 152.14, "end": 153.32, "text": "Damp Prompt肯定顶不住", "chunk": 0, "language": "zh", "avg_logprob": -0.14661186099670093, "no_speech_prob": 2.236957320767008e-11, "compression_ratio": 1.256267409470752 }, { "start": 153.32, "end": 155.86, "text": "我们需要使用Workflow工作流方案", "chunk": 0, "language": "zh", "avg_logprob": -0.14661186099670093, "no_speech_prob": 2.236957320767008e-11, "compression_ratio": 1.256267409470752 }, { "start": 155.86, "end": 158.08, "text": "这就像我们去银行里面办业务", "chunk": 0, "language": "zh", "avg_logprob": -0.14661186099670093, "no_speech_prob": 2.236957320767008e-11, "compression_ratio": 1.256267409470752 }, { "start": 158.08, "end": 161.18, "text": "先取号填表柜台办理最后签字", "chunk": 0, "language": "zh", "avg_logprob": -0.14661186099670093, "no_speech_prob": 2.236957320767008e-11, "compression_ratio": 1.256267409470752 }, { "start": 161.18, "end": 162.42, "text": "每一步怎么走", "chunk": 0, "language": "zh", "avg_logprob": -0.14661186099670093, "no_speech_prob": 2.236957320767008e-11, "compression_ratio": 1.256267409470752 }, { "start": 162.42, "end": 163.96, "text": "节点之间怎么传递信息", "chunk": 0, "language": "zh", "avg_logprob": -0.14661186099670093, "no_speech_prob": 2.236957320767008e-11, "compression_ratio": 1.256267409470752 }, { "start": 163.96, "end": 165.38, "text": "都是定死的显性的", "chunk": 0, "language": "zh", "avg_logprob": -0.14661186099670093, "no_speech_prob": 2.236957320767008e-11, "compression_ratio": 1.256267409470752 }, { "start": 165.38, "end": 167.42, "text": "这种流逝的步骤设计", "chunk": 0, "language": "zh", "avg_logprob": -0.14661186099670093, "no_speech_prob": 2.236957320767008e-11, "compression_ratio": 1.256267409470752 }, { "start": 167.42, "end": 169.26, "text": "最适合用来跑标准的SOP", "chunk": 0, "language": "zh", "avg_logprob": -0.14661186099670093, "no_speech_prob": 2.236957320767008e-11, "compression_ratio": 1.256267409470752 }, { "start": 169.26, "end": 170.84, "text": "出来的结果非常稳定", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 170.84, "end": 172.44, "text": "那再往上走", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 172.44, "end": 173.22, "text": "到了第三档", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 173.22, "end": 174.48, "text": "就是高复杂度", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 174.48, "end": 175.38, "text": "长周期", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 175.38, "end": 177.38, "text": "需要多角色分工的混沌任务了", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 177.38, "end": 179.26, "text": "比如我们想要大模型", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 179.26, "end": 180.74, "text": "去写一个完整的软件模块", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 180.74, "end": 182.8, "text": "这个里面设计的需求就太多了", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 182.8, "end": 185.66, "text": "我们必须用多智能体协同的方案", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 185.66, "end": 186.74, "text": "那在这个架构下", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 186.74, "end": 187.68, "text": "我们需要引入", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 187.68, "end": 189.26, "text": "人机协同的安全红线", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 189.26, "end": 191.2, "text": "比如智能体写完了代码", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 191.2, "end": 192.78, "text": "在准备部署上线的那一步", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 192.78, "end": 193.74, "text": "必须停下来", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 193.74, "end": 195.02, "text": "等待人工审核确认", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 195.02, "end": 196.36, "text": "确认过了才能继续", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 196.36, "end": 197.44, "text": "同时", "chunk": 0, "language": "zh", "avg_logprob": -0.08393542574480246, "no_speech_prob": 2.195171128205331e-11, "compression_ratio": 1.3832335329341316 }, { "start": 197.44, "end": 198.7, "text": "各个智能体之间", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 198.7, "end": 200.04, "text": "还有自己的独立上下文", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 200.04, "end": 201.62, "text": "不能够把信息给混在一起", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 201.62, "end": 203.34, "text": "不然模型自己就糊涂了", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 203.34, "end": 205.68, "text": "那么面对这种高复杂度的", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 205.68, "end": 206.5, "text": "多智能体任务", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 206.5, "end": 207.54, "text": "我们在工程上", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 207.54, "end": 208.66, "text": "到底该怎么去编排", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 208.66, "end": 209.22, "text": "怎么设计", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 209.22, "end": 211.04, "text": "才能够保证它输出的质量呢", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 211.04, "end": 212.88, "text": "其实现在行业里", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 212.88, "end": 214.08, "text": "有一个挺成熟的思路", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 214.08, "end": 215.42, "text": "是分布循环", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 215.42, "end": 216.66, "text": "加上格力子智能体", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 216.66, "end": 218.24, "text": "听起来是有点专业", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 218.24, "end": 219.3, "text": "我们拆开来聊", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 219.3, "end": 220.32, "text": "其实非常简单", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 220.32, "end": 221.54, "text": "先说分布循环", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 221.54, "end": 223.62, "text": "就是把任务交给模型之后", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 223.62, "end": 225.2, "text": "它一上来又直接开始写", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 225.2, "end": 227.16, "text": "结果写出来的东西完全不能用", "chunk": 0, "language": "zh", "avg_logprob": -0.08967247009277343, "no_speech_prob": 1.6837784638790154e-11, "compression_ratio": 1.3994638069705094 }, { "start": 227.44, "end": 229.96, "text": "所以我们需要在框架上来限制它的行为", "chunk": 0, "language": "zh", "avg_logprob": -0.08437998845027044, "no_speech_prob": 1.945393325597511e-11, "compression_ratio": 1.4485981308411215 }, { "start": 229.96, "end": 233.4, "text": "比如我们可以采用一个探索规划执行的一个循环", "chunk": 0, "language": "zh", "avg_logprob": -0.08437998845027044, "no_speech_prob": 1.945393325597511e-11, "compression_ratio": 1.4485981308411215 }, { "start": 233.4, "end": 234.96, "text": "第一步是探索", "chunk": 0, "language": "zh", "avg_logprob": -0.08437998845027044, "no_speech_prob": 1.945393325597511e-11, "compression_ratio": 1.4485981308411215 }, { "start": 234.96, "end": 237.98, "text": "这能体现去读取相关的文档或者搜索代码库", "chunk": 0, "language": "zh", "avg_logprob": -0.08437998845027044, "no_speech_prob": 1.945393325597511e-11, "compression_ratio": 1.4485981308411215 }, { "start": 237.98, "end": 241.14, "text": "所以在这个阶段我们只给它只读的权限", "chunk": 0, "language": "zh", "avg_logprob": -0.08437998845027044, "no_speech_prob": 1.945393325597511e-11, "compression_ratio": 1.4485981308411215 }, { "start": 241.14, "end": 242.6, "text": "它只能看不能改", "chunk": 0, "language": "zh", "avg_logprob": -0.08437998845027044, "no_speech_prob": 1.945393325597511e-11, "compression_ratio": 1.4485981308411215 }, { "start": 242.6, "end": 244.08, "text": "第二步是规划", "chunk": 0, "language": "zh", "avg_logprob": -0.08437998845027044, "no_speech_prob": 1.945393325597511e-11, "compression_ratio": 1.4485981308411215 }, { "start": 244.08, "end": 247.34, "text": "它根据看懂的信息做一个具体的执行计划", "chunk": 0, "language": "zh", "avg_logprob": -0.08437998845027044, "no_speech_prob": 1.945393325597511e-11, "compression_ratio": 1.4485981308411215 }, { "start": 247.34, "end": 249.0, "text": "那么在计划做好之后", "chunk": 0, "language": "zh", "avg_logprob": -0.08437998845027044, "no_speech_prob": 1.945393325597511e-11, "compression_ratio": 1.4485981308411215 }, { "start": 249.0, "end": 250.34, "text": "我们可以加入人工确认", "chunk": 0, "language": "zh", "avg_logprob": -0.08437998845027044, "no_speech_prob": 1.945393325597511e-11, "compression_ratio": 1.4485981308411215 }, { "start": 250.34, "end": 251.48, "text": "也就是人机协同", "chunk": 0, "language": "zh", "avg_logprob": -0.08437998845027044, "no_speech_prob": 1.945393325597511e-11, "compression_ratio": 1.4485981308411215 }, { "start": 251.48, "end": 253.62, "text": "人类觉得计划可行再放行", "chunk": 0, "language": "zh", "avg_logprob": -0.08437998845027044, "no_speech_prob": 1.945393325597511e-11, "compression_ratio": 1.4485981308411215 }, { "start": 253.62, "end": 255.54, "text": "第三步才是执行", "chunk": 0, "language": "zh", "avg_logprob": -0.08437998845027044, "no_speech_prob": 1.945393325597511e-11, "compression_ratio": 1.4485981308411215 }, { "start": 255.54, "end": 256.92, "text": "他开始去写文件", "chunk": 0, "language": "zh", "avg_logprob": -0.11217935005443518, "no_speech_prob": 1.8479039479157855e-11, "compression_ratio": 1.3532608695652173 }, { "start": 256.92, "end": 257.68, "text": "运行命令", "chunk": 0, "language": "zh", "avg_logprob": -0.11217935005443518, "no_speech_prob": 1.8479039479157855e-11, "compression_ratio": 1.3532608695652173 }, { "start": 257.68, "end": 260.6, "text": "这个时候他才能拿到完整的操作权限", "chunk": 0, "language": "zh", "avg_logprob": -0.11217935005443518, "no_speech_prob": 1.8479039479157855e-11, "compression_ratio": 1.3532608695652173 }, { "start": 260.6, "end": 261.48, "text": "执行完之后", "chunk": 0, "language": "zh", "avg_logprob": -0.11217935005443518, "no_speech_prob": 1.8479039479157855e-11, "compression_ratio": 1.3532608695652173 }, { "start": 261.48, "end": 262.42, "text": "如果发现报错了", "chunk": 0, "language": "zh", "avg_logprob": -0.11217935005443518, "no_speech_prob": 1.8479039479157855e-11, "compression_ratio": 1.3532608695652173 }, { "start": 262.42, "end": 264.48, "text": "他会自动回到第一步重新探索", "chunk": 0, "language": "zh", "avg_logprob": -0.11217935005443518, "no_speech_prob": 1.8479039479157855e-11, "compression_ratio": 1.3532608695652173 }, { "start": 264.48, "end": 266.16, "text": "这样一套循环下来", "chunk": 0, "language": "zh", "avg_logprob": -0.11217935005443518, "no_speech_prob": 1.8479039479157855e-11, "compression_ratio": 1.3532608695652173 }, { "start": 266.16, "end": 268.14, "text": "任务就会被框得非常稳", "chunk": 0, "language": "zh", "avg_logprob": -0.11217935005443518, "no_speech_prob": 1.8479039479157855e-11, "compression_ratio": 1.3532608695652173 }, { "start": 268.14, "end": 269.44, "text": "那么除了这个循环", "chunk": 0, "language": "zh", "avg_logprob": -0.11217935005443518, "no_speech_prob": 1.8479039479157855e-11, "compression_ratio": 1.3532608695652173 }, { "start": 269.44, "end": 270.82, "text": "还有一个更关键的点", "chunk": 0, "language": "zh", "avg_logprob": -0.11217935005443518, "no_speech_prob": 1.8479039479157855e-11, "compression_ratio": 1.3532608695652173 }, { "start": 270.82, "end": 272.98, "text": "叫做上下文隔离子智能体", "chunk": 0, "language": "zh", "avg_logprob": -0.11217935005443518, "no_speech_prob": 1.8479039479157855e-11, "compression_ratio": 1.3532608695652173 }, { "start": 272.98, "end": 276.14, "text": "你看如果一个智能体要做的事情太多", "chunk": 0, "language": "zh", "avg_logprob": -0.11217935005443518, "no_speech_prob": 1.8479039479157855e-11, "compression_ratio": 1.3532608695652173 }, { "start": 276.14, "end": 278.42, "text": "它的上下文里就会塞满各种各样信息", "chunk": 0, "language": "zh", "avg_logprob": -0.11217935005443518, "no_speech_prob": 1.8479039479157855e-11, "compression_ratio": 1.3532608695652173 }, { "start": 278.42, "end": 280.58, "text": "这就好比我们自己开一家公司", "chunk": 0, "language": "zh", "avg_logprob": -0.11217935005443518, "no_speech_prob": 1.8479039479157855e-11, "compression_ratio": 1.3532608695652173 }, { "start": 280.58, "end": 283.14, "text": "你不能让一个人既当销售又当财务", "chunk": 0, "language": "zh", "avg_logprob": -0.11217935005443518, "no_speech_prob": 1.8479039479157855e-11, "compression_ratio": 1.3532608695652173 }, { "start": 283.14, "end": 284.54, "text": "还要当程序员对吧", "chunk": 0, "language": "zh", "avg_logprob": -0.11217935005443518, "no_speech_prob": 1.8479039479157855e-11, "compression_ratio": 1.3532608695652173 }, { "start": 284.54, "end": 286.24, "text": "信息一多他肯定会记错", "chunk": 0, "language": "zh", "avg_logprob": -0.07227330482923068, "no_speech_prob": 1.644748573448318e-11, "compression_ratio": 1.5096774193548388 }, { "start": 286.24, "end": 288.46, "text": "所以呢我们要设立岗位", "chunk": 0, "language": "zh", "avg_logprob": -0.07227330482923068, "no_speech_prob": 1.644748573448318e-11, "compression_ratio": 1.5096774193548388 }, { "start": 288.46, "end": 290.46, "text": "比如我们有一个主智能体", "chunk": 0, "language": "zh", "avg_logprob": -0.07227330482923068, "no_speech_prob": 1.644748573448318e-11, "compression_ratio": 1.5096774193548388 }, { "start": 290.46, "end": 293.6, "text": "他手底下呢有三个上下文隔离的子智能体", "chunk": 0, "language": "zh", "avg_logprob": -0.07227330482923068, "no_speech_prob": 1.644748573448318e-11, "compression_ratio": 1.5096774193548388 }, { "start": 293.6, "end": 296.18, "text": "一个专门找资料的搜索智能体", "chunk": 0, "language": "zh", "avg_logprob": -0.07227330482923068, "no_speech_prob": 1.644748573448318e-11, "compression_ratio": 1.5096774193548388 }, { "start": 296.18, "end": 297.98, "text": "一个做计划的规划智能体", "chunk": 0, "language": "zh", "avg_logprob": -0.07227330482923068, "no_speech_prob": 1.644748573448318e-11, "compression_ratio": 1.5096774193548388 }, { "start": 297.98, "end": 299.82, "text": "一个负责写代码的执行智能体", "chunk": 0, "language": "zh", "avg_logprob": -0.07227330482923068, "no_speech_prob": 1.644748573448318e-11, "compression_ratio": 1.5096774193548388 }, { "start": 299.82, "end": 300.96, "text": "最关键的是", "chunk": 0, "language": "zh", "avg_logprob": -0.07227330482923068, "no_speech_prob": 1.644748573448318e-11, "compression_ratio": 1.5096774193548388 }, { "start": 300.96, "end": 303.48, "text": "他们每个人只拥有自己专属的上下文", "chunk": 0, "language": "zh", "avg_logprob": -0.07227330482923068, "no_speech_prob": 1.644748573448318e-11, "compression_ratio": 1.5096774193548388 }, { "start": 303.48, "end": 305.44, "text": "写代码的不知道财务数据", "chunk": 0, "language": "zh", "avg_logprob": -0.07227330482923068, "no_speech_prob": 1.644748573448318e-11, "compression_ratio": 1.5096774193548388 }, { "start": 305.44, "end": 307.56, "text": "找资料的也不需要管代码里的bug", "chunk": 0, "language": "zh", "avg_logprob": -0.07227330482923068, "no_speech_prob": 1.644748573448318e-11, "compression_ratio": 1.5096774193548388 }, { "start": 307.56, "end": 312.24, "text": "他们之间只通过主智能体进行最标准最干净的数据传递", "chunk": 0, "language": "zh", "avg_logprob": -0.07227330482923068, "no_speech_prob": 1.644748573448318e-11, "compression_ratio": 1.5096774193548388 }, { "start": 312.24, "end": 314.7, "text": "通过这种各干各的相互隔离的设计", "chunk": 0, "language": "zh", "avg_logprob": -0.08183684238167696, "no_speech_prob": 1.812730521244532e-11, "compression_ratio": 1.3452685421994885 }, { "start": 314.7, "end": 317.52, "text": "每个子智能体面对的信息量都会极大减少", "chunk": 0, "language": "zh", "avg_logprob": -0.08183684238167696, "no_speech_prob": 1.812730521244532e-11, "compression_ratio": 1.3452685421994885 }, { "start": 317.52, "end": 319.4, "text": "任务被彻底拆干净了", "chunk": 0, "language": "zh", "avg_logprob": -0.08183684238167696, "no_speech_prob": 1.812730521244532e-11, "compression_ratio": 1.3452685421994885 }, { "start": 319.4, "end": 320.98, "text": "模型的准确率自然就上去了", "chunk": 0, "language": "zh", "avg_logprob": -0.08183684238167696, "no_speech_prob": 1.812730521244532e-11, "compression_ratio": 1.3452685421994885 }, { "start": 320.98, "end": 323.44, "text": "OK 聊完了业务怎么拆解", "chunk": 0, "language": "zh", "avg_logprob": -0.08183684238167696, "no_speech_prob": 1.812730521244532e-11, "compression_ratio": 1.3452685421994885 }, { "start": 323.44, "end": 325.26, "text": "接下来我们讲第二个核心能力", "chunk": 0, "language": "zh", "avg_logprob": -0.08183684238167696, "no_speech_prob": 1.812730521244532e-11, "compression_ratio": 1.3452685421994885 }, { "start": 325.26, "end": 326.98, "text": "工具调用的能力", "chunk": 0, "language": "zh", "avg_logprob": -0.08183684238167696, "no_speech_prob": 1.812730521244532e-11, "compression_ratio": 1.3452685421994885 }, { "start": 326.98, "end": 329.2, "text": "其实大模型项目在线上翻车", "chunk": 0, "language": "zh", "avg_logprob": -0.08183684238167696, "no_speech_prob": 1.812730521244532e-11, "compression_ratio": 1.3452685421994885 }, { "start": 329.2, "end": 331.82, "text": "绝大部分问题都不是出在模型不够聪明上", "chunk": 0, "language": "zh", "avg_logprob": -0.08183684238167696, "no_speech_prob": 1.812730521244532e-11, "compression_ratio": 1.3452685421994885 }, { "start": 331.82, "end": 333.74, "text": "而是我们的工具设计出了问题", "chunk": 0, "language": "zh", "avg_logprob": -0.08183684238167696, "no_speech_prob": 1.812730521244532e-11, "compression_ratio": 1.3452685421994885 }, { "start": 333.74, "end": 336.34, "text": "我们很多刚入行做应用开发的朋友", "chunk": 0, "language": "zh", "avg_logprob": -0.08183684238167696, "no_speech_prob": 1.812730521244532e-11, "compression_ratio": 1.3452685421994885 }, { "start": 336.34, "end": 339.24, "text": "在让大模型去修改文件或者运行命令的时候", "chunk": 0, "language": "zh", "avg_logprob": -0.08183684238167696, "no_speech_prob": 1.812730521244532e-11, "compression_ratio": 1.3452685421994885 }, { "start": 339.24, "end": 342.18, "text": "特别喜欢给他一个全能的Mesh工具", "chunk": 0, "language": "zh", "avg_logprob": -0.08183684238167696, "no_speech_prob": 1.812730521244532e-11, "compression_ratio": 1.3452685421994885 }, { "start": 342.18, "end": 344.12, "text": "也就是我们今天常说的命令行", "chunk": 0, "language": "zh", "avg_logprob": -0.1027894390679394, "no_speech_prob": 1.7854452291632406e-11, "compression_ratio": 1.4005763688760806 }, { "start": 344.12, "end": 346.12, "text": "大模型想要读文件", "chunk": 0, "language": "zh", "avg_logprob": -0.1027894390679394, "no_speech_prob": 1.7854452291632406e-11, "compression_ratio": 1.4005763688760806 }, { "start": 346.12, "end": 346.78, "text": "改代码", "chunk": 0, "language": "zh", "avg_logprob": -0.1027894390679394, "no_speech_prob": 1.7854452291632406e-11, "compression_ratio": 1.4005763688760806 }, { "start": 346.78, "end": 347.74, "text": "搜索关键词", "chunk": 0, "language": "zh", "avg_logprob": -0.1027894390679394, "no_speech_prob": 1.7854452291632406e-11, "compression_ratio": 1.4005763688760806 }, { "start": 347.74, "end": 349.48, "text": "全都通过这一个命令行", "chunk": 0, "language": "zh", "avg_logprob": -0.1027894390679394, "no_speech_prob": 1.7854452291632406e-11, "compression_ratio": 1.4005763688760806 }, { "start": 349.48, "end": 350.86, "text": "去敲各种指令来解决", "chunk": 0, "language": "zh", "avg_logprob": -0.1027894390679394, "no_speech_prob": 1.7854452291632406e-11, "compression_ratio": 1.4005763688760806 }, { "start": 350.86, "end": 353.76, "text": "他们觉得这样模型很自由很省事对吧", "chunk": 0, "language": "zh", "avg_logprob": -0.1027894390679394, "no_speech_prob": 1.7854452291632406e-11, "compression_ratio": 1.4005763688760806 }, { "start": 353.76, "end": 355.76, "text": "但实际上在工业级开发里", "chunk": 0, "language": "zh", "avg_logprob": -0.1027894390679394, "no_speech_prob": 1.7854452291632406e-11, "compression_ratio": 1.4005763688760806 }, { "start": 355.76, "end": 357.54, "text": "这是一个非常危险的反面模式", "chunk": 0, "language": "zh", "avg_logprob": -0.1027894390679394, "no_speech_prob": 1.7854452291632406e-11, "compression_ratio": 1.4005763688760806 }, { "start": 357.54, "end": 359.1, "text": "为什么这么说呢", "chunk": 0, "language": "zh", "avg_logprob": -0.1027894390679394, "no_speech_prob": 1.7854452291632406e-11, "compression_ratio": 1.4005763688760806 }, { "start": 359.1, "end": 360.16, "text": "你想想", "chunk": 0, "language": "zh", "avg_logprob": -0.1027894390679394, "no_speech_prob": 1.7854452291632406e-11, "compression_ratio": 1.4005763688760806 }, { "start": 360.16, "end": 362.06, "text": "如果大模型手里拿的是一个", "chunk": 0, "language": "zh", "avg_logprob": -0.1027894390679394, "no_speech_prob": 1.7854452291632406e-11, "compression_ratio": 1.4005763688760806 }, { "start": 362.06, "end": 363.8, "text": "可以为所欲为的万能命令行", "chunk": 0, "language": "zh", "avg_logprob": -0.1027894390679394, "no_speech_prob": 1.7854452291632406e-11, "compression_ratio": 1.4005763688760806 }, { "start": 363.8, "end": 366.2, "text": "它能干的事情就太多了", "chunk": 0, "language": "zh", "avg_logprob": -0.1027894390679394, "no_speech_prob": 1.7854452291632406e-11, "compression_ratio": 1.4005763688760806 }, { "start": 366.2, "end": 367.8, "text": "一旦它某一次产生了幻觉", "chunk": 0, "language": "zh", "avg_logprob": -0.1027894390679394, "no_speech_prob": 1.7854452291632406e-11, "compression_ratio": 1.4005763688760806 }, { "start": 367.8, "end": 369.1, "text": "或者拼错了一个字符", "chunk": 0, "language": "zh", "avg_logprob": -0.1027894390679394, "no_speech_prob": 1.7854452291632406e-11, "compression_ratio": 1.4005763688760806 }, { "start": 369.1, "end": 370.8, "text": "就非常容易拼接出一个", "chunk": 0, "language": "zh", "avg_logprob": -0.1027894390679394, "no_speech_prob": 1.7854452291632406e-11, "compression_ratio": 1.4005763688760806 }, { "start": 370.8, "end": 372.58, "text": "把系统文件彻底弄坏的错误指令", "chunk": 0, "language": "zh", "avg_logprob": -0.09786351230166374, "no_speech_prob": 1.7979407893053967e-11, "compression_ratio": 1.3584905660377358 }, { "start": 372.58, "end": 374.38, "text": "而且作为开发者", "chunk": 0, "language": "zh", "avg_logprob": -0.09786351230166374, "no_speech_prob": 1.7979407893053967e-11, "compression_ratio": 1.3584905660377358 }, { "start": 374.38, "end": 375.8, "text": "你怎么去审查它的权限", "chunk": 0, "language": "zh", "avg_logprob": -0.09786351230166374, "no_speech_prob": 1.7979407893053967e-11, "compression_ratio": 1.3584905660377358 }, { "start": 375.8, "end": 377.4, "text": "你根本没有办法", "chunk": 0, "language": "zh", "avg_logprob": -0.09786351230166374, "no_speech_prob": 1.7979407893053967e-11, "compression_ratio": 1.3584905660377358 }, { "start": 377.4, "end": 378.82, "text": "给它做精细化的安全授权", "chunk": 0, "language": "zh", "avg_logprob": -0.09786351230166374, "no_speech_prob": 1.7979407893053967e-11, "compression_ratio": 1.3584905660377358 }, { "start": 378.82, "end": 380.84, "text": "这就像是你给新来的实习生", "chunk": 0, "language": "zh", "avg_logprob": -0.09786351230166374, "no_speech_prob": 1.7979407893053967e-11, "compression_ratio": 1.3584905660377358 }, { "start": 380.84, "end": 381.98, "text": "配了一把万能钥匙", "chunk": 0, "language": "zh", "avg_logprob": -0.09786351230166374, "no_speech_prob": 1.7979407893053967e-11, "compression_ratio": 1.3584905660377358 }, { "start": 381.98, "end": 383.5, "text": "能拔开公司所有的门", "chunk": 0, "language": "zh", "avg_logprob": -0.09786351230166374, "no_speech_prob": 1.7979407893053967e-11, "compression_ratio": 1.3584905660377358 }, { "start": 383.5, "end": 385.48, "text": "这个安全隐患应该是非常大的", "chunk": 0, "language": "zh", "avg_logprob": -0.09786351230166374, "no_speech_prob": 1.7979407893053967e-11, "compression_ratio": 1.3584905660377358 }, { "start": 385.48, "end": 387.46, "text": "所以我们更提倡一种", "chunk": 0, "language": "zh", "avg_logprob": -0.09786351230166374, "no_speech_prob": 1.7979407893053967e-11, "compression_ratio": 1.3584905660377358 }, { "start": 387.46, "end": 388.8, "text": "单一职责工具模式", "chunk": 0, "language": "zh", "avg_logprob": -0.09786351230166374, "no_speech_prob": 1.7979407893053967e-11, "compression_ratio": 1.3584905660377358 }, { "start": 388.8, "end": 390.06, "text": "简单来说", "chunk": 0, "language": "zh", "avg_logprob": -0.09786351230166374, "no_speech_prob": 1.7979407893053967e-11, "compression_ratio": 1.3584905660377358 }, { "start": 390.06, "end": 391.82, "text": "就是把这些大而全的命令行", "chunk": 0, "language": "zh", "avg_logprob": -0.09786351230166374, "no_speech_prob": 1.7979407893053967e-11, "compression_ratio": 1.3584905660377358 }, { "start": 391.82, "end": 394.42, "text": "拆成一个个只盖一键小式的专业工具", "chunk": 0, "language": "zh", "avg_logprob": -0.09786351230166374, "no_speech_prob": 1.7979407893053967e-11, "compression_ratio": 1.3584905660377358 }, { "start": 394.42, "end": 396.36, "text": "读文件的工具就是read", "chunk": 0, "language": "zh", "avg_logprob": -0.09786351230166374, "no_speech_prob": 1.7979407893053967e-11, "compression_ratio": 1.3584905660377358 }, { "start": 396.36, "end": 398.26, "text": "改文件的呢就是edit", "chunk": 0, "language": "zh", "avg_logprob": -0.09786351230166374, "no_speech_prob": 1.7979407893053967e-11, "compression_ratio": 1.3584905660377358 }, { "start": 398.26, "end": 399.96, "text": "查找关键词的工具是grab", "chunk": 0, "language": "zh", "avg_logprob": -0.09786351230166374, "no_speech_prob": 1.7979407893053967e-11, "compression_ratio": 1.3584905660377358 }, { "start": 399.96, "end": 401.8, "text": "写文件的工具呢是write", "chunk": 0, "language": "zh", "avg_logprob": -0.09060645470252404, "no_speech_prob": 1.770276460144604e-11, "compression_ratio": 1.3818681318681318 }, { "start": 401.8, "end": 403.56, "text": "这样设计的好处显而易见", "chunk": 0, "language": "zh", "avg_logprob": -0.09060645470252404, "no_speech_prob": 1.770276460144604e-11, "compression_ratio": 1.3818681318681318 }, { "start": 403.56, "end": 406.28, "text": "每一个工具都有非常严格的参数规范", "chunk": 0, "language": "zh", "avg_logprob": -0.09060645470252404, "no_speech_prob": 1.770276460144604e-11, "compression_ratio": 1.3818681318681318 }, { "start": 406.28, "end": 407.04, "text": "也就是schema", "chunk": 0, "language": "zh", "avg_logprob": -0.09060645470252404, "no_speech_prob": 1.770276460144604e-11, "compression_ratio": 1.3818681318681318 }, { "start": 407.04, "end": 408.76, "text": "大模型在调用的时候", "chunk": 0, "language": "zh", "avg_logprob": -0.09060645470252404, "no_speech_prob": 1.770276460144604e-11, "compression_ratio": 1.3818681318681318 }, { "start": 408.76, "end": 409.82, "text": "边界非常明确", "chunk": 0, "language": "zh", "avg_logprob": -0.09060645470252404, "no_speech_prob": 1.770276460144604e-11, "compression_ratio": 1.3818681318681318 }, { "start": 409.82, "end": 412.94, "text": "而且我们还可以给不同的工具分配独立的权限", "chunk": 0, "language": "zh", "avg_logprob": -0.09060645470252404, "no_speech_prob": 1.770276460144604e-11, "compression_ratio": 1.3818681318681318 }, { "start": 412.94, "end": 414.82, "text": "接口的确定性变高了", "chunk": 0, "language": "zh", "avg_logprob": -0.09060645470252404, "no_speech_prob": 1.770276460144604e-11, "compression_ratio": 1.3818681318681318 }, { "start": 414.82, "end": 416.84, "text": "模型的理解偏差自然就小了", "chunk": 0, "language": "zh", "avg_logprob": -0.09060645470252404, "no_speech_prob": 1.770276460144604e-11, "compression_ratio": 1.3818681318681318 }, { "start": 416.84, "end": 419.1, "text": "那这时候新的问题就来了", "chunk": 0, "language": "zh", "avg_logprob": -0.09060645470252404, "no_speech_prob": 1.770276460144604e-11, "compression_ratio": 1.3818681318681318 }, { "start": 419.1, "end": 421.24, "text": "随着我们的项目越做越大", "chunk": 0, "language": "zh", "avg_logprob": -0.09060645470252404, "no_speech_prob": 1.770276460144604e-11, "compression_ratio": 1.3818681318681318 }, { "start": 421.24, "end": 423.2, "text": "大模型需要用的工具越来越多", "chunk": 0, "language": "zh", "avg_logprob": -0.09060645470252404, "no_speech_prob": 1.770276460144604e-11, "compression_ratio": 1.3818681318681318 }, { "start": 423.2, "end": 425.72, "text": "我们该怎么去管理和扩展这些工具呢", "chunk": 0, "language": "zh", "avg_logprob": -0.09060645470252404, "no_speech_prob": 1.770276460144604e-11, "compression_ratio": 1.3818681318681318 }, { "start": 425.72, "end": 429.2, "text": "另外如果大模型确实需要跑一些像删除文件", "chunk": 0, "language": "zh", "avg_logprob": -0.09060645470252404, "no_speech_prob": 1.770276460144604e-11, "compression_ratio": 1.3818681318681318 }, { "start": 429.2, "end": 431.1, "text": "还有推送代法这些高风险的命令", "chunk": 0, "language": "zh", "avg_logprob": -0.09247810828504223, "no_speech_prob": 1.906641337867665e-11, "compression_ratio": 1.4424778761061947 }, { "start": 431.1, "end": 433.4, "text": "我们又该怎么控制它的安全风险呢", "chunk": 0, "language": "zh", "avg_logprob": -0.09247810828504223, "no_speech_prob": 1.906641337867665e-11, "compression_ratio": 1.4424778761061947 }, { "start": 433.4, "end": 435.08, "text": "在实际工程落地中", "chunk": 0, "language": "zh", "avg_logprob": -0.09247810828504223, "no_speech_prob": 1.906641337867665e-11, "compression_ratio": 1.4424778761061947 }, { "start": 435.08, "end": 436.76, "text": "我们一般要遵循两个原则", "chunk": 0, "language": "zh", "avg_logprob": -0.09247810828504223, "no_speech_prob": 1.906641337867665e-11, "compression_ratio": 1.4424778761061947 }, { "start": 436.76, "end": 439.32, "text": "第一个原则叫渐进式工具扩展", "chunk": 0, "language": "zh", "avg_logprob": -0.09247810828504223, "no_speech_prob": 1.906641337867665e-11, "compression_ratio": 1.4424778761061947 }, { "start": 439.32, "end": 440.4, "text": "什么意思呢", "chunk": 0, "language": "zh", "avg_logprob": -0.09247810828504223, "no_speech_prob": 1.906641337867665e-11, "compression_ratio": 1.4424778761061947 }, { "start": 440.4, "end": 441.62, "text": "就是我们一开始", "chunk": 0, "language": "zh", "avg_logprob": -0.09247810828504223, "no_speech_prob": 1.906641337867665e-11, "compression_ratio": 1.4424778761061947 }, { "start": 441.62, "end": 444.38, "text": "千万别一上来就把几十个上百个工具", "chunk": 0, "language": "zh", "avg_logprob": -0.09247810828504223, "no_speech_prob": 1.906641337867665e-11, "compression_ratio": 1.4424778761061947 }, { "start": 444.38, "end": 445.46, "text": "全都喂给大木型", "chunk": 0, "language": "zh", "avg_logprob": -0.09247810828504223, "no_speech_prob": 1.906641337867665e-11, "compression_ratio": 1.4424778761061947 }, { "start": 445.46, "end": 446.92, "text": "工具给的太多", "chunk": 0, "language": "zh", "avg_logprob": -0.09247810828504223, "no_speech_prob": 1.906641337867665e-11, "compression_ratio": 1.4424778761061947 }, { "start": 446.92, "end": 448.24, "text": "大木型在做选择的时候", "chunk": 0, "language": "zh", "avg_logprob": -0.09247810828504223, "no_speech_prob": 1.906641337867665e-11, "compression_ratio": 1.4424778761061947 }, { "start": 448.24, "end": 449.3, "text": "非常容易挑花眼", "chunk": 0, "language": "zh", "avg_logprob": -0.09247810828504223, "no_speech_prob": 1.906641337867665e-11, "compression_ratio": 1.4424778761061947 }, { "start": 449.3, "end": 450.36, "text": "最后选错工具", "chunk": 0, "language": "zh", "avg_logprob": -0.09247810828504223, "no_speech_prob": 1.906641337867665e-11, "compression_ratio": 1.4424778761061947 }, { "start": 450.36, "end": 452.0, "text": "刚开始的时候", "chunk": 0, "language": "zh", "avg_logprob": -0.09247810828504223, "no_speech_prob": 1.906641337867665e-11, "compression_ratio": 1.4424778761061947 }, { "start": 452.0, "end": 453.56, "text": "我们本地只给他提供", "chunk": 0, "language": "zh", "avg_logprob": -0.09247810828504223, "no_speech_prob": 1.906641337867665e-11, "compression_ratio": 1.4424778761061947 }, { "start": 453.56, "end": 455.48, "text": "一二十个最常用最基础的工具", "chunk": 0, "language": "zh", "avg_logprob": -0.09247810828504223, "no_speech_prob": 1.906641337867665e-11, "compression_ratio": 1.4424778761061947 }, { "start": 455.48, "end": 457.9, "text": "像读写修改就完全够了", "chunk": 0, "language": "zh", "avg_logprob": -0.09247810828504223, "no_speech_prob": 1.906641337867665e-11, "compression_ratio": 1.4424778761061947 }, { "start": 457.9, "end": 460.88, "text": "如果后面我们发现需要接入一些外部系统", "chunk": 0, "language": "zh", "avg_logprob": -0.09494286716574489, "no_speech_prob": 1.3675468743534758e-11, "compression_ratio": 1.427710843373494 }, { "start": 460.88, "end": 462.4, "text": "比如要读公司的数据库", "chunk": 0, "language": "zh", "avg_logprob": -0.09494286716574489, "no_speech_prob": 1.3675468743534758e-11, "compression_ratio": 1.427710843373494 }, { "start": 462.4, "end": 464.18, "text": "要调用外部的第三方服务", "chunk": 0, "language": "zh", "avg_logprob": -0.09494286716574489, "no_speech_prob": 1.3675468743534758e-11, "compression_ratio": 1.427710843373494 }, { "start": 464.18, "end": 465.96, "text": "那我们再加上MCP", "chunk": 0, "language": "zh", "avg_logprob": -0.09494286716574489, "no_speech_prob": 1.3675468743534758e-11, "compression_ratio": 1.427710843373494 }, { "start": 465.96, "end": 467.62, "text": "也就是模型上下文协议", "chunk": 0, "language": "zh", "avg_logprob": -0.09494286716574489, "no_speech_prob": 1.3675468743534758e-11, "compression_ratio": 1.427710843373494 }, { "start": 467.62, "end": 470.18, "text": "标准化的把这些外部工具给加载进来", "chunk": 0, "language": "zh", "avg_logprob": -0.09494286716574489, "no_speech_prob": 1.3675468743534758e-11, "compression_ratio": 1.427710843373494 }, { "start": 470.18, "end": 473.44, "text": "如果再往后需要调用远程云端的主机", "chunk": 0, "language": "zh", "avg_logprob": -0.09494286716574489, "no_speech_prob": 1.3675468743534758e-11, "compression_ratio": 1.427710843373494 }, { "start": 473.44, "end": 474.3, "text": "或远程的数据库", "chunk": 0, "language": "zh", "avg_logprob": -0.09494286716574489, "no_speech_prob": 1.3675468743534758e-11, "compression_ratio": 1.427710843373494 }, { "start": 474.3, "end": 475.92, "text": "我们再介入远程工具", "chunk": 0, "language": "zh", "avg_logprob": -0.09494286716574489, "no_speech_prob": 1.3675468743534758e-11, "compression_ratio": 1.427710843373494 }, { "start": 475.92, "end": 477.98, "text": "一步一步按需加载", "chunk": 0, "language": "zh", "avg_logprob": -0.09494286716574489, "no_speech_prob": 1.3675468743534758e-11, "compression_ratio": 1.427710843373494 }, { "start": 477.98, "end": 480.44, "text": "这样模型的工具调用准确率才不会崩", "chunk": 0, "language": "zh", "avg_logprob": -0.09494286716574489, "no_speech_prob": 1.3675468743534758e-11, "compression_ratio": 1.427710843373494 }, { "start": 480.44, "end": 482.14, "text": "那么第二个原则", "chunk": 0, "language": "zh", "avg_logprob": -0.09494286716574489, "no_speech_prob": 1.3675468743534758e-11, "compression_ratio": 1.427710843373494 }, { "start": 482.14, "end": 484.44, "text": "是必须在底层建立一个安全杀箱", "chunk": 0, "language": "zh", "avg_logprob": -0.09494286716574489, "no_speech_prob": 1.3675468743534758e-11, "compression_ratio": 1.427710843373494 }, { "start": 484.44, "end": 486.14, "text": "做好命令的风险分级", "chunk": 0, "language": "zh", "avg_logprob": -0.09494286716574489, "no_speech_prob": 1.3675468743534758e-11, "compression_ratio": 1.427710843373494 }, { "start": 486.783, "end": 488.743, "text": "大模型它毕竟是模型", "chunk": 1, "language": "zh", "avg_logprob": -0.07644869201811985, "no_speech_prob": 1.798354173909722e-11, "compression_ratio": 1.424778761061947 }, { "start": 488.743, "end": 490.163, "text": "它有时候可能会失控", "chunk": 1, "language": "zh", "avg_logprob": -0.07644869201811985, "no_speech_prob": 1.798354173909722e-11, "compression_ratio": 1.424778761061947 }, { "start": 490.163, "end": 493.203, "text": "所以我们必须给它调用的所有命令定下规则", "chunk": 1, "language": "zh", "avg_logprob": -0.07644869201811985, "no_speech_prob": 1.798354173909722e-11, "compression_ratio": 1.424778761061947 }, { "start": 493.203, "end": 494.343, "text": "分出安全等级", "chunk": 1, "language": "zh", "avg_logprob": -0.07644869201811985, "no_speech_prob": 1.798354173909722e-11, "compression_ratio": 1.424778761061947 }, { "start": 494.343, "end": 497.443, "text": "比如像跑个测试或者看一下文件差异", "chunk": 1, "language": "zh", "avg_logprob": -0.07644869201811985, "no_speech_prob": 1.798354173909722e-11, "compression_ratio": 1.424778761061947 }, { "start": 497.443, "end": 499.343, "text": "这些完全是安全的命令", "chunk": 1, "language": "zh", "avg_logprob": -0.07644869201811985, "no_speech_prob": 1.798354173909722e-11, "compression_ratio": 1.424778761061947 }, { "start": 499.343, "end": 500.863, "text": "我们可以设置为自动执行", "chunk": 1, "language": "zh", "avg_logprob": -0.07644869201811985, "no_speech_prob": 1.798354173909722e-11, "compression_ratio": 1.424778761061947 }, { "start": 500.863, "end": 502.083, "text": "不需要打扰用户", "chunk": 1, "language": "zh", "avg_logprob": -0.10938324751677336, "no_speech_prob": 1.70345040156894e-11, "compression_ratio": 1.3892045454545454 }, { "start": 502.083, "end": 504.983, "text": "但如果是像强推代码这样的操作", "chunk": 1, "language": "zh", "avg_logprob": -0.10938324751677336, "no_speech_prob": 1.70345040156894e-11, "compression_ratio": 1.3892045454545454 }, { "start": 504.983, "end": 506.263, "text": "这就有一定的风险了", "chunk": 1, "language": "zh", "avg_logprob": -0.10938324751677336, "no_speech_prob": 1.70345040156894e-11, "compression_ratio": 1.3892045454545454 }, { "start": 506.263, "end": 507.923, "text": "系统在跑这个命令之前", "chunk": 1, "language": "zh", "avg_logprob": -0.10938324751677336, "no_speech_prob": 1.70345040156894e-11, "compression_ratio": 1.3892045454545454 }, { "start": 507.923, "end": 509.183, "text": "必须暂停下来", "chunk": 1, "language": "zh", "avg_logprob": -0.10938324751677336, "no_speech_prob": 1.70345040156894e-11, "compression_ratio": 1.3892045454545454 }, { "start": 509.183, "end": 510.643, "text": "弹出来要用户确认一下", "chunk": 1, "language": "zh", "avg_logprob": -0.10938324751677336, "no_speech_prob": 1.70345040156894e-11, "compression_ratio": 1.3892045454545454 }, { "start": 510.643, "end": 512.483, "text": "必须等我们人肉确认了", "chunk": 1, "language": "zh", "avg_logprob": -0.10938324751677336, "no_speech_prob": 1.70345040156894e-11, "compression_ratio": 1.3892045454545454 }, { "start": 512.483, "end": 513.103, "text": "他才能跑", "chunk": 1, "language": "zh", "avg_logprob": -0.10938324751677336, "no_speech_prob": 1.70345040156894e-11, "compression_ratio": 1.3892045454545454 }, { "start": 513.103, "end": 514.763, "text": "而如果是像这种", "chunk": 1, "language": "zh", "avg_logprob": -0.10938324751677336, "no_speech_prob": 1.70345040156894e-11, "compression_ratio": 1.3892045454545454 }, { "start": 514.763, "end": 516.383, "text": "会煽光系统的毁灭性指令", "chunk": 1, "language": "zh", "avg_logprob": -0.10938324751677336, "no_speech_prob": 1.70345040156894e-11, "compression_ratio": 1.3892045454545454 }, { "start": 516.383, "end": 518.183, "text": "不用问我们在沙箱底层", "chunk": 1, "language": "zh", "avg_logprob": -0.10938324751677336, "no_speech_prob": 1.70345040156894e-11, "compression_ratio": 1.3892045454545454 }, { "start": 518.183, "end": 519.403, "text": "直接硬变码彻底拉黑", "chunk": 1, "language": "zh", "avg_logprob": -0.10938324751677336, "no_speech_prob": 1.70345040156894e-11, "compression_ratio": 1.3892045454545454 }, { "start": 519.403, "end": 520.883, "text": "不给他任何运行的机会", "chunk": 1, "language": "zh", "avg_logprob": -0.10938324751677336, "no_speech_prob": 1.70345040156894e-11, "compression_ratio": 1.3892045454545454 }, { "start": 520.883, "end": 523.163, "text": "你看我们只有在底层", "chunk": 1, "language": "zh", "avg_logprob": -0.10938324751677336, "no_speech_prob": 1.70345040156894e-11, "compression_ratio": 1.3892045454545454 }, { "start": 523.163, "end": 525.403, "text": "把这一套分级拦截的沙箱机制打好了", "chunk": 1, "language": "zh", "avg_logprob": -0.10938324751677336, "no_speech_prob": 1.70345040156894e-11, "compression_ratio": 1.3892045454545454 }, { "start": 525.403, "end": 527.423, "text": "才能真正放心的把工具", "chunk": 1, "language": "zh", "avg_logprob": -0.10938324751677336, "no_speech_prob": 1.70345040156894e-11, "compression_ratio": 1.3892045454545454 }, { "start": 527.423, "end": 528.923, "text": "交到大模型手里去跑对吧", "chunk": 1, "language": "zh", "avg_logprob": -0.10938324751677336, "no_speech_prob": 1.70345040156894e-11, "compression_ratio": 1.3892045454545454 }, { "start": 528.923, "end": 532.043, "text": "那接下来我们讲第三个很容易被忽略", "chunk": 1, "language": "zh", "avg_logprob": -0.0936265269284526, "no_speech_prob": 1.2961492122653961e-11, "compression_ratio": 1.3628571428571428 }, { "start": 532.043, "end": 534.843, "text": "但是却决定了项目能不能真正交付的能力", "chunk": 1, "language": "zh", "avg_logprob": -0.0936265269284526, "no_speech_prob": 1.2961492122653961e-11, "compression_ratio": 1.3628571428571428 }, { "start": 534.843, "end": 536.263, "text": "评测和可观测性", "chunk": 1, "language": "zh", "avg_logprob": -0.0936265269284526, "no_speech_prob": 1.2961492122653961e-11, "compression_ratio": 1.3628571428571428 }, { "start": 536.263, "end": 538.523, "text": "平时我们自己调模型的时候", "chunk": 1, "language": "zh", "avg_logprob": -0.0936265269284526, "no_speech_prob": 1.2961492122653961e-11, "compression_ratio": 1.3628571428571428 }, { "start": 538.523, "end": 539.343, "text": "经常会想", "chunk": 1, "language": "zh", "avg_logprob": -0.0936265269284526, "no_speech_prob": 1.2961492122653961e-11, "compression_ratio": 1.3628571428571428 }, { "start": 539.343, "end": 541.963, "text": "我测了几次感觉效果还挺不错的", "chunk": 1, "language": "zh", "avg_logprob": -0.0936265269284526, "no_speech_prob": 1.2961492122653961e-11, "compression_ratio": 1.3628571428571428 }, { "start": 541.963, "end": 544.783, "text": "但是企业级最怕的就是感觉这两个字", "chunk": 1, "language": "zh", "avg_logprob": -0.0936265269284526, "no_speech_prob": 1.2961492122653961e-11, "compression_ratio": 1.3628571428571428 }, { "start": 544.783, "end": 547.383, "text": "你改了一个prompt或者换了一个模型", "chunk": 1, "language": "zh", "avg_logprob": -0.0936265269284526, "no_speech_prob": 1.2961492122653961e-11, "compression_ratio": 1.3628571428571428 }, { "start": 547.383, "end": 548.483, "text": "你感觉变好了", "chunk": 1, "language": "zh", "avg_logprob": -0.0936265269284526, "no_speech_prob": 1.2961492122653961e-11, "compression_ratio": 1.3628571428571428 }, { "start": 548.483, "end": 550.763, "text": "但可能在某个你没测到的业务角落", "chunk": 1, "language": "zh", "avg_logprob": -0.0936265269284526, "no_speech_prob": 1.2961492122653961e-11, "compression_ratio": 1.3628571428571428 }, { "start": 550.763, "end": 551.643, "text": "它其实退化了", "chunk": 1, "language": "zh", "avg_logprob": -0.0936265269284526, "no_speech_prob": 1.2961492122653961e-11, "compression_ratio": 1.3628571428571428 }, { "start": 551.643, "end": 553.803, "text": "这在工程上叫做静默退化", "chunk": 1, "language": "zh", "avg_logprob": -0.0936265269284526, "no_speech_prob": 1.2961492122653961e-11, "compression_ratio": 1.3628571428571428 }, { "start": 553.803, "end": 555.903, "text": "你根本不知道它什么时候", "chunk": 1, "language": "zh", "avg_logprob": -0.0936265269284526, "no_speech_prob": 1.2961492122653961e-11, "compression_ratio": 1.3628571428571428 }, { "start": 555.903, "end": 557.243, "text": "在什么地方变差了", "chunk": 1, "language": "zh", "avg_logprob": -0.0936265269284526, "no_speech_prob": 1.2961492122653961e-11, "compression_ratio": 1.3628571428571428 }, { "start": 557.243, "end": 559.803, "text": "所以我们必须告别主观体感", "chunk": 1, "language": "zh", "avg_logprob": -0.08840326309204101, "no_speech_prob": 1.3583538807648843e-11, "compression_ratio": 1.3519061583577712 }, { "start": 559.803, "end": 562.303, "text": "建立一套体系化的量化评测流程", "chunk": 1, "language": "zh", "avg_logprob": -0.08840326309204101, "no_speech_prob": 1.3583538807648843e-11, "compression_ratio": 1.3519061583577712 }, { "start": 562.303, "end": 563.963, "text": "这就像去医院体检", "chunk": 1, "language": "zh", "avg_logprob": -0.08840326309204101, "no_speech_prob": 1.3583538807648843e-11, "compression_ratio": 1.3519061583577712 }, { "start": 563.963, "end": 566.123, "text": "不能只靠医生问你感觉怎么样", "chunk": 1, "language": "zh", "avg_logprob": -0.08840326309204101, "no_speech_prob": 1.3583538807648843e-11, "compression_ratio": 1.3519061583577712 }, { "start": 566.123, "end": 567.543, "text": "而是要抽血化验", "chunk": 1, "language": "zh", "avg_logprob": -0.08840326309204101, "no_speech_prob": 1.3583538807648843e-11, "compression_ratio": 1.3519061583577712 }, { "start": 567.543, "end": 568.883, "text": "看各项指标的数值", "chunk": 1, "language": "zh", "avg_logprob": -0.08840326309204101, "no_speech_prob": 1.3583538807648843e-11, "compression_ratio": 1.3519061583577712 }, { "start": 568.883, "end": 570.363, "text": "具体在开发里", "chunk": 1, "language": "zh", "avg_logprob": -0.08840326309204101, "no_speech_prob": 1.3583538807648843e-11, "compression_ratio": 1.3519061583577712 }, { "start": 570.363, "end": 571.523, "text": "这个流程分为三步", "chunk": 1, "language": "zh", "avg_logprob": -0.08840326309204101, "no_speech_prob": 1.3583538807648843e-11, "compression_ratio": 1.3519061583577712 }, { "start": 571.523, "end": 574.343, "text": "第一步我们要构建一个黄金数据集", "chunk": 1, "language": "zh", "avg_logprob": -0.08840326309204101, "no_speech_prob": 1.3583538807648843e-11, "compression_ratio": 1.3519061583577712 }, { "start": 574.343, "end": 576.783, "text": "这个数据集要覆盖我们最核心", "chunk": 1, "language": "zh", "avg_logprob": -0.08840326309204101, "no_speech_prob": 1.3583538807648843e-11, "compression_ratio": 1.3519061583577712 }, { "start": 576.783, "end": 577.983, "text": "最真实的业务场景", "chunk": 1, "language": "zh", "avg_logprob": -0.08840326309204101, "no_speech_prob": 1.3583538807648843e-11, "compression_ratio": 1.3519061583577712 }, { "start": 577.983, "end": 581.443, "text": "比如包含50个或者100个经典的客户提问", "chunk": 1, "language": "zh", "avg_logprob": -0.08840326309204101, "no_speech_prob": 1.3583538807648843e-11, "compression_ratio": 1.3519061583577712 }, { "start": 581.443, "end": 582.403, "text": "和对应的标准答案", "chunk": 1, "language": "zh", "avg_logprob": -0.08840326309204101, "no_speech_prob": 1.3583538807648843e-11, "compression_ratio": 1.3519061583577712 }, { "start": 582.403, "end": 584.163, "text": "这就是我们的考试试卷", "chunk": 1, "language": "zh", "avg_logprob": -0.08840326309204101, "no_speech_prob": 1.3583538807648843e-11, "compression_ratio": 1.3519061583577712 }, { "start": 584.163, "end": 585.923, "text": "第二步有了试卷", "chunk": 1, "language": "zh", "avg_logprob": -0.08840326309204101, "no_speech_prob": 1.3583538807648843e-11, "compression_ratio": 1.3519061583577712 }, { "start": 585.923, "end": 588.543, "text": "我们得找个客观的乐卷老师来自动进行打分", "chunk": 1, "language": "zh", "avg_logprob": -0.10694864607348885, "no_speech_prob": 1.568868993051531e-11, "compression_ratio": 1.3121546961325967 }, { "start": 588.543, "end": 591.003, "text": "我们一般会引入一个强推力模型", "chunk": 1, "language": "zh", "avg_logprob": -0.10694864607348885, "no_speech_prob": 1.568868993051531e-11, "compression_ratio": 1.3121546961325967 }, { "start": 591.003, "end": 592.523, "text": "让大模型当裁判", "chunk": 1, "language": "zh", "avg_logprob": -0.10694864607348885, "no_speech_prob": 1.568868993051531e-11, "compression_ratio": 1.3121546961325967 }, { "start": 592.523, "end": 594.203, "text": "也就是业内常说的", "chunk": 1, "language": "zh", "avg_logprob": -0.10694864607348885, "no_speech_prob": 1.568868993051531e-11, "compression_ratio": 1.3121546961325967 }, { "start": 594.203, "end": 595.523, "text": "LM as a judge", "chunk": 1, "language": "zh", "avg_logprob": -0.10694864607348885, "no_speech_prob": 1.568868993051531e-11, "compression_ratio": 1.3121546961325967 }, { "start": 595.523, "end": 597.583, "text": "这样每次我们改的代码", "chunk": 1, "language": "zh", "avg_logprob": -0.10694864607348885, "no_speech_prob": 1.568868993051531e-11, "compression_ratio": 1.3121546961325967 }, { "start": 597.583, "end": 598.963, "text": "直接自动化跑一遍考试", "chunk": 1, "language": "zh", "avg_logprob": -0.10694864607348885, "no_speech_prob": 1.568868993051531e-11, "compression_ratio": 1.3121546961325967 }, { "start": 598.963, "end": 600.943, "text": "摆脱人工人肉测试的低效", "chunk": 1, "language": "zh", "avg_logprob": -0.10694864607348885, "no_speech_prob": 1.568868993051531e-11, "compression_ratio": 1.3121546961325967 }, { "start": 600.943, "end": 603.283, "text": "那第三步看成绩单", "chunk": 1, "language": "zh", "avg_logprob": -0.10694864607348885, "no_speech_prob": 1.568868993051531e-11, "compression_ratio": 1.3121546961325967 }, { "start": 603.283, "end": 605.623, "text": "这个成绩单不能只有一个总分", "chunk": 1, "language": "zh", "avg_logprob": -0.10694864607348885, "no_speech_prob": 1.568868993051531e-11, "compression_ratio": 1.3121546961325967 }, { "start": 605.623, "end": 607.923, "text": "而是要有一个多维度的指标体系", "chunk": 1, "language": "zh", "avg_logprob": -0.10694864607348885, "no_speech_prob": 1.568868993051531e-11, "compression_ratio": 1.3121546961325967 }, { "start": 607.923, "end": 610.423, "text": "比如我们要看幻觉率有没有上升", "chunk": 1, "language": "zh", "avg_logprob": -0.10694864607348885, "no_speech_prob": 1.568868993051531e-11, "compression_ratio": 1.3121546961325967 }, { "start": 610.423, "end": 612.763, "text": "模型给出的答案跟问题相不相关", "chunk": 1, "language": "zh", "avg_logprob": -0.10694864607348885, "no_speech_prob": 1.568868993051531e-11, "compression_ratio": 1.3121546961325967 }, { "start": 612.763, "end": 614.783, "text": "检索知识库的召回率高不高", "chunk": 1, "language": "zh", "avg_logprob": -0.10694864607348885, "no_speech_prob": 1.568868993051531e-11, "compression_ratio": 1.3121546961325967 }, { "start": 614.783, "end": 617.563, "text": "还有最实际的响应耗时有没有变慢", "chunk": 1, "language": "zh", "avg_logprob": -0.08468015182421403, "no_speech_prob": 1.4570625955778738e-11, "compression_ratio": 1.3846153846153846 }, { "start": 617.563, "end": 619.543, "text": "通过这样的一套闭环评测", "chunk": 1, "language": "zh", "avg_logprob": -0.08468015182421403, "no_speech_prob": 1.4570625955778738e-11, "compression_ratio": 1.3846153846153846 }, { "start": 619.543, "end": 621.023, "text": "我们每次迭代升级", "chunk": 1, "language": "zh", "avg_logprob": -0.08468015182421403, "no_speech_prob": 1.4570625955778738e-11, "compression_ratio": 1.3846153846153846 }, { "start": 621.023, "end": 622.243, "text": "心里才是有底的对吧", "chunk": 1, "language": "zh", "avg_logprob": -0.08468015182421403, "no_speech_prob": 1.4570625955778738e-11, "compression_ratio": 1.3846153846153846 }, { "start": 622.243, "end": 624.023, "text": "有了这一套评测流程", "chunk": 1, "language": "zh", "avg_logprob": -0.08468015182421403, "no_speech_prob": 1.4570625955778738e-11, "compression_ratio": 1.3846153846153846 }, { "start": 624.023, "end": 626.363, "text": "只能说明我们的应用在模拟考试的时候", "chunk": 1, "language": "zh", "avg_logprob": -0.08468015182421403, "no_speech_prob": 1.4570625955778738e-11, "compression_ratio": 1.3846153846153846 }, { "start": 626.363, "end": 627.143, "text": "表现还不错", "chunk": 1, "language": "zh", "avg_logprob": -0.08468015182421403, "no_speech_prob": 1.4570625955778738e-11, "compression_ratio": 1.3846153846153846 }, { "start": 627.143, "end": 629.123, "text": "但如果项目真的上线了", "chunk": 1, "language": "zh", "avg_logprob": -0.08468015182421403, "no_speech_prob": 1.4570625955778738e-11, "compression_ratio": 1.3846153846153846 }, { "start": 629.123, "end": 630.963, "text": "用户在使用过程中遇到了爆错", "chunk": 1, "language": "zh", "avg_logprob": -0.08468015182421403, "no_speech_prob": 1.4570625955778738e-11, "compression_ratio": 1.3846153846153846 }, { "start": 630.963, "end": 632.663, "text": "或者突然输出了一堆乱码", "chunk": 1, "language": "zh", "avg_logprob": -0.08468015182421403, "no_speech_prob": 1.4570625955778738e-11, "compression_ratio": 1.3846153846153846 }, { "start": 632.663, "end": 633.783, "text": "我们该怎么排查呢", "chunk": 1, "language": "zh", "avg_logprob": -0.08468015182421403, "no_speech_prob": 1.4570625955778738e-11, "compression_ratio": 1.3846153846153846 }, { "start": 633.783, "end": 635.483, "text": "很多项目上线之后", "chunk": 1, "language": "zh", "avg_logprob": -0.08468015182421403, "no_speech_prob": 1.4570625955778738e-11, "compression_ratio": 1.3846153846153846 }, { "start": 635.483, "end": 636.283, "text": "一旦爆错了", "chunk": 1, "language": "zh", "avg_logprob": -0.08468015182421403, "no_speech_prob": 1.4570625955778738e-11, "compression_ratio": 1.3846153846153846 }, { "start": 636.283, "end": 637.183, "text": "开发者就懵了", "chunk": 1, "language": "zh", "avg_logprob": -0.08468015182421403, "no_speech_prob": 1.4570625955778738e-11, "compression_ratio": 1.3846153846153846 }, { "start": 637.183, "end": 639.423, "text": "因为大模型应用就像个黑盒", "chunk": 1, "language": "zh", "avg_logprob": -0.08468015182421403, "no_speech_prob": 1.4570625955778738e-11, "compression_ratio": 1.3846153846153846 }, { "start": 639.423, "end": 641.543, "text": "你根本不知道中间哪一步出了问题", "chunk": 1, "language": "zh", "avg_logprob": -0.08468015182421403, "no_speech_prob": 1.4570625955778738e-11, "compression_ratio": 1.3846153846153846 }, { "start": 641.543, "end": 644.943, "text": "所以我们必须要做全链路可观测性设计", "chunk": 1, "language": "zh", "avg_logprob": -0.09520021764243522, "no_speech_prob": 1.412177839943407e-11, "compression_ratio": 1.3890489913544668 }, { "start": 644.943, "end": 647.123, "text": "每次推理每次工具调用", "chunk": 1, "language": "zh", "avg_logprob": -0.09520021764243522, "no_speech_prob": 1.412177839943407e-11, "compression_ratio": 1.3890489913544668 }, { "start": 647.123, "end": 649.743, "text": "它的整个生命周期都必须被完整的记录下来", "chunk": 1, "language": "zh", "avg_logprob": -0.09520021764243522, "no_speech_prob": 1.412177839943407e-11, "compression_ratio": 1.3890489913544668 }, { "start": 649.743, "end": 651.583, "text": "首先在输入阶段", "chunk": 1, "language": "zh", "avg_logprob": -0.09520021764243522, "no_speech_prob": 1.412177839943407e-11, "compression_ratio": 1.3890489913544668 }, { "start": 651.583, "end": 655.463, "text": "我们要捕获最原始的prompt和系统当时的状态", "chunk": 1, "language": "zh", "avg_logprob": -0.09520021764243522, "no_speech_prob": 1.412177839943407e-11, "compression_ratio": 1.3890489913544668 }, { "start": 655.463, "end": 656.863, "text": "接着在推理中间菜", "chunk": 1, "language": "zh", "avg_logprob": -0.09520021764243522, "no_speech_prob": 1.412177839943407e-11, "compression_ratio": 1.3890489913544668 }, { "start": 656.863, "end": 658.563, "text": "我们要监控模型的思维链", "chunk": 1, "language": "zh", "avg_logprob": -0.09520021764243522, "no_speech_prob": 1.412177839943407e-11, "compression_ratio": 1.3890489913544668 }, { "start": 658.563, "end": 660.283, "text": "看它是怎么一步步思考", "chunk": 1, "language": "zh", "avg_logprob": -0.09520021764243522, "no_speech_prob": 1.412177839943407e-11, "compression_ratio": 1.3890489913544668 }, { "start": 660.283, "end": 662.163, "text": "同时记录它消耗了多少token", "chunk": 1, "language": "zh", "avg_logprob": -0.09520021764243522, "no_speech_prob": 1.412177839943407e-11, "compression_ratio": 1.3890489913544668 }, { "start": 662.163, "end": 663.903, "text": "这直接关系到我们的账单成本", "chunk": 1, "language": "zh", "avg_logprob": -0.09520021764243522, "no_speech_prob": 1.412177839943407e-11, "compression_ratio": 1.3890489913544668 }, { "start": 663.903, "end": 665.903, "text": "然后在工具调用阶段", "chunk": 1, "language": "zh", "avg_logprob": -0.09520021764243522, "no_speech_prob": 1.412177839943407e-11, "compression_ratio": 1.3890489913544668 }, { "start": 665.903, "end": 667.843, "text": "大模型传了什么参数", "chunk": 1, "language": "zh", "avg_logprob": -0.09520021764243522, "no_speech_prob": 1.412177839943407e-11, "compression_ratio": 1.3890489913544668 }, { "start": 667.843, "end": 669.143, "text": "调用了什么接口", "chunk": 1, "language": "zh", "avg_logprob": -0.09520021764243522, "no_speech_prob": 1.412177839943407e-11, "compression_ratio": 1.3890489913544668 }, { "start": 669.143, "end": 670.743, "text": "接口响应花了多少时间", "chunk": 1, "language": "zh", "avg_logprob": -0.09520021764243522, "no_speech_prob": 1.412177839943407e-11, "compression_ratio": 1.3890489913544668 }, { "start": 670.743, "end": 671.683, "text": "也要记下来", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 671.683, "end": 673.283, "text": "最后在输出阶段", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 673.283, "end": 674.603, "text": "如果发生了异常", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 674.603, "end": 675.843, "text": "比如说超时了呀", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 675.843, "end": 676.923, "text": "或者是格式不对", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 676.923, "end": 678.983, "text": "或者触发了我们的服务降级策略", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 678.983, "end": 680.823, "text": "系统也要能够精准的捕捉", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 680.823, "end": 683.303, "text": "把这些数据完整的记录下来之后", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 683.303, "end": 684.343, "text": "最核心的好处", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 684.343, "end": 685.603, "text": "就是我们可以进行一次", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 685.603, "end": 687.383, "text": "故障复现和回放调试", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 687.383, "end": 689.623, "text": "比如线上用户反馈说", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 689.623, "end": 690.623, "text": "模型刚才瞎编了", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 690.623, "end": 691.903, "text": "我们不需要去猜", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 691.903, "end": 693.863, "text": "直接把那次调用的数据给导出来", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 693.863, "end": 695.283, "text": "在本地重新跑一遍", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 695.283, "end": 697.763, "text": "看看大模型在思维链的哪一步走偏了", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 697.763, "end": 699.143, "text": "或者是调工具的时候", "chunk": 1, "language": "zh", "avg_logprob": -0.09382459004720052, "no_speech_prob": 1.9379362697913294e-11, "compression_ratio": 1.3739837398373984 }, { "start": 699.143, "end": 700.503, "text": "传哪个参数传错了", "chunk": 1, "language": "zh", "avg_logprob": -0.10625349325577015, "no_speech_prob": 1.5935319036541884e-11, "compression_ratio": 1.3342776203966005 }, { "start": 700.503, "end": 702.903, "text": "然后我们就可以精准优化prompt", "chunk": 1, "language": "zh", "avg_logprob": -0.10625349325577015, "no_speech_prob": 1.5935319036541884e-11, "compression_ratio": 1.3342776203966005 }, { "start": 702.903, "end": 704.883, "text": "或者加上超时重视", "chunk": 1, "language": "zh", "avg_logprob": -0.10625349325577015, "no_speech_prob": 1.5935319036541884e-11, "compression_ratio": 1.3342776203966005 }, { "start": 704.883, "end": 706.563, "text": "格式纠错这样的容错机制", "chunk": 1, "language": "zh", "avg_logprob": -0.10625349325577015, "no_speech_prob": 1.5935319036541884e-11, "compression_ratio": 1.3342776203966005 }, { "start": 706.563, "end": 709.183, "text": "这才是真正的工程化调试思维", "chunk": 1, "language": "zh", "avg_logprob": -0.10625349325577015, "no_speech_prob": 1.5935319036541884e-11, "compression_ratio": 1.3342776203966005 }, { "start": 709.183, "end": 710.163, "text": "好", "chunk": 1, "language": "zh", "avg_logprob": -0.10625349325577015, "no_speech_prob": 1.5935319036541884e-11, "compression_ratio": 1.3342776203966005 }, { "start": 710.163, "end": 712.163, "text": "有了这个评测和可观测性", "chunk": 1, "language": "zh", "avg_logprob": -0.10625349325577015, "no_speech_prob": 1.5935319036541884e-11, "compression_ratio": 1.3342776203966005 }, { "start": 712.163, "end": 713.783, "text": "我们的应用在开发阶段", "chunk": 1, "language": "zh", "avg_logprob": -0.10625349325577015, "no_speech_prob": 1.5935319036541884e-11, "compression_ratio": 1.3342776203966005 }, { "start": 713.783, "end": 715.083, "text": "已经看起来很健康了", "chunk": 1, "language": "zh", "avg_logprob": -0.10625349325577015, "no_speech_prob": 1.5935319036541884e-11, "compression_ratio": 1.3342776203966005 }, { "start": 715.083, "end": 717.543, "text": "但是如果要真正推上线上", "chunk": 1, "language": "zh", "avg_logprob": -0.10625349325577015, "no_speech_prob": 1.5935319036541884e-11, "compression_ratio": 1.3342776203966005 }, { "start": 717.543, "end": 719.303, "text": "交给成千上万的用户去用", "chunk": 1, "language": "zh", "avg_logprob": -0.10625349325577015, "no_speech_prob": 1.5935319036541884e-11, "compression_ratio": 1.3342776203966005 }, { "start": 719.303, "end": 721.343, "text": "这就到了最拉开差距的地方", "chunk": 1, "language": "zh", "avg_logprob": -0.10625349325577015, "no_speech_prob": 1.5935319036541884e-11, "compression_ratio": 1.3342776203966005 }, { "start": 721.343, "end": 722.863, "text": "我们的第四个核心能力", "chunk": 1, "language": "zh", "avg_logprob": -0.10625349325577015, "no_speech_prob": 1.5935319036541884e-11, "compression_ratio": 1.3342776203966005 }, { "start": 722.863, "end": 724.063, "text": "生产环境能力", "chunk": 1, "language": "zh", "avg_logprob": -0.10625349325577015, "no_speech_prob": 1.5935319036541884e-11, "compression_ratio": 1.3342776203966005 }, { "start": 724.063, "end": 726.463, "text": "生产环境里有一个非常实际的痛点", "chunk": 1, "language": "zh", "avg_logprob": -0.10625349325577015, "no_speech_prob": 1.5935319036541884e-11, "compression_ratio": 1.3342776203966005 }, { "start": 726.463, "end": 727.923, "text": "就是模型特别健忘", "chunk": 1, "language": "zh", "avg_logprob": -0.10625349325577015, "no_speech_prob": 1.5935319036541884e-11, "compression_ratio": 1.3342776203966005 }, { "start": 727.923, "end": 729.483, "text": "而且绘画越长", "chunk": 1, "language": "zh", "avg_logprob": -0.08668840196397569, "no_speech_prob": 1.6190214099931488e-11, "compression_ratio": 1.2916666666666667 }, { "start": 729.483, "end": 730.563, "text": "上下文塞的越满", "chunk": 1, "language": "zh", "avg_logprob": -0.08668840196397569, "no_speech_prob": 1.6190214099931488e-11, "compression_ratio": 1.2916666666666667 }, { "start": 730.563, "end": 732.343, "text": "不仅模型响应变得越来越慢", "chunk": 1, "language": "zh", "avg_logprob": -0.08668840196397569, "no_speech_prob": 1.6190214099931488e-11, "compression_ratio": 1.2916666666666667 }, { "start": 732.343, "end": 734.843, "text": "公司的token资费也会像流水一样花出去", "chunk": 1, "language": "zh", "avg_logprob": -0.08668840196397569, "no_speech_prob": 1.6190214099931488e-11, "compression_ratio": 1.2916666666666667 }, { "start": 734.843, "end": 737.623, "text": "我们不能指望用户每次跟智能体聊天", "chunk": 1, "language": "zh", "avg_logprob": -0.08668840196397569, "no_speech_prob": 1.6190214099931488e-11, "compression_ratio": 1.2916666666666667 }, { "start": 737.623, "end": 739.423, "text": "都把项目规则和历史对话", "chunk": 1, "language": "zh", "avg_logprob": -0.08668840196397569, "no_speech_prob": 1.6190214099931488e-11, "compression_ratio": 1.2916666666666667 }, { "start": 739.423, "end": 740.503, "text": "复制粘贴一遍吧", "chunk": 1, "language": "zh", "avg_logprob": -0.08668840196397569, "no_speech_prob": 1.6190214099931488e-11, "compression_ratio": 1.2916666666666667 }, { "start": 740.503, "end": 742.843, "text": "所以在生产及状态管理里", "chunk": 1, "language": "zh", "avg_logprob": -0.08668840196397569, "no_speech_prob": 1.6190214099931488e-11, "compression_ratio": 1.2916666666666667 }, { "start": 742.843, "end": 744.363, "text": "我们首先要在代码库里", "chunk": 1, "language": "zh", "avg_logprob": -0.08668840196397569, "no_speech_prob": 1.6190214099931488e-11, "compression_ratio": 1.2916666666666667 }, { "start": 744.363, "end": 746.103, "text": "放一个持久化的配置指令文件", "chunk": 1, "language": "zh", "avg_logprob": -0.08668840196397569, "no_speech_prob": 1.6190214099931488e-11, "compression_ratio": 1.2916666666666667 }, { "start": 746.103, "end": 747.723, "text": "比如说cloud.md", "chunk": 1, "language": "zh", "avg_logprob": -0.08668840196397569, "no_speech_prob": 1.6190214099931488e-11, "compression_ratio": 1.2916666666666667 }, { "start": 747.723, "end": 750.203, "text": "这个文件是随我们的代码库", "chunk": 1, "language": "zh", "avg_logprob": -0.08668840196397569, "no_speech_prob": 1.6190214099931488e-11, "compression_ratio": 1.2916666666666667 }, { "start": 750.203, "end": 751.323, "text": "一起做版本控制的", "chunk": 1, "language": "zh", "avg_logprob": -0.08668840196397569, "no_speech_prob": 1.6190214099931488e-11, "compression_ratio": 1.2916666666666667 }, { "start": 751.323, "end": 752.863, "text": "当绘画启动时", "chunk": 1, "language": "zh", "avg_logprob": -0.08668840196397569, "no_speech_prob": 1.6190214099931488e-11, "compression_ratio": 1.2916666666666667 }, { "start": 752.863, "end": 754.763, "text": "系统会自动把项目级的规范", "chunk": 1, "language": "zh", "avg_logprob": -0.08668840196397569, "no_speech_prob": 1.6190214099931488e-11, "compression_ratio": 1.2916666666666667 }, { "start": 754.763, "end": 756.743, "text": "测试命令自动加载进去", "chunk": 1, "language": "zh", "avg_logprob": -0.08668840196397569, "no_speech_prob": 1.6190214099931488e-11, "compression_ratio": 1.2916666666666667 }, { "start": 756.743, "end": 757.583, "text": "模型一上线", "chunk": 1, "language": "zh", "avg_logprob": -0.12512675721441682, "no_speech_prob": 1.6036161981647368e-11, "compression_ratio": 1.3677811550151975 }, { "start": 757.583, "end": 759.523, "text": "立刻就知道自己身处什么项目", "chunk": 1, "language": "zh", "avg_logprob": -0.12512675721441682, "no_speech_prob": 1.6036161981647368e-11, "compression_ratio": 1.3677811550151975 }, { "start": 759.523, "end": 760.743, "text": "需要遵守什么规则", "chunk": 1, "language": "zh", "avg_logprob": -0.12512675721441682, "no_speech_prob": 1.6036161981647368e-11, "compression_ratio": 1.3677811550151975 }, { "start": 760.743, "end": 762.303, "text": "总进一步呢", "chunk": 1, "language": "zh", "avg_logprob": -0.12512675721441682, "no_speech_prob": 1.6036161981647368e-11, "compression_ratio": 1.3677811550151975 }, { "start": 762.303, "end": 764.183, "text": "我们要建立一个分层记忆机制", "chunk": 1, "language": "zh", "avg_logprob": -0.12512675721441682, "no_speech_prob": 1.6036161981647368e-11, "compression_ratio": 1.3677811550151975 }, { "start": 764.183, "end": 766.443, "text": "对对上计算机有缓存", "chunk": 1, "language": "zh", "avg_logprob": -0.12512675721441682, "no_speech_prob": 1.6036161981647368e-11, "compression_ratio": 1.3677811550151975 }, { "start": 766.443, "end": 767.103, "text": "有内存", "chunk": 1, "language": "zh", "avg_logprob": -0.12512675721441682, "no_speech_prob": 1.6036161981647368e-11, "compression_ratio": 1.3677811550151975 }, { "start": 767.103, "end": 767.663, "text": "有硬盘", "chunk": 1, "language": "zh", "avg_logprob": -0.12512675721441682, "no_speech_prob": 1.6036161981647368e-11, "compression_ratio": 1.3677811550151975 }, { "start": 767.663, "end": 770.743, "text": "第一层是永远加载的核心索引", "chunk": 1, "language": "zh", "avg_logprob": -0.12512675721441682, "no_speech_prob": 1.6036161981647368e-11, "compression_ratio": 1.3677811550151975 }, { "start": 770.743, "end": 773.423, "text": "这部分一般控制在200行以内", "chunk": 1, "language": "zh", "avg_logprob": -0.12512675721441682, "no_speech_prob": 1.6036161981647368e-11, "compression_ratio": 1.3677811550151975 }, { "start": 773.423, "end": 774.863, "text": "保持极低的响应延迟", "chunk": 1, "language": "zh", "avg_logprob": -0.12512675721441682, "no_speech_prob": 1.6036161981647368e-11, "compression_ratio": 1.3677811550151975 }, { "start": 774.863, "end": 778.983, "text": "然后第二层是按需加载的相关文档和当前状态", "chunk": 1, "language": "zh", "avg_logprob": -0.12512675721441682, "no_speech_prob": 1.6036161981647368e-11, "compression_ratio": 1.3677811550151975 }, { "start": 778.983, "end": 781.683, "text": "比如模型当前要修改某个代码文件", "chunk": 1, "language": "zh", "avg_logprob": -0.12512675721441682, "no_speech_prob": 1.6036161981647368e-11, "compression_ratio": 1.3677811550151975 }, { "start": 781.683, "end": 785.183, "text": "我们就只要把这个文件和当前的锻点状态加载进来", "chunk": 1, "language": "zh", "avg_logprob": -0.12512675721441682, "no_speech_prob": 1.6036161981647368e-11, "compression_ratio": 1.3677811550151975 }, { "start": 785.183, "end": 786.783, "text": "这里非常重要的一点", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 786.783, "end": 788.023, "text": "是支持断点恢复", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 788.023, "end": 789.683, "text": "万一服务中断了", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 789.683, "end": 790.823, "text": "智能体重启之后", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 790.823, "end": 791.903, "text": "能接着上次的进度干", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 791.903, "end": 792.763, "text": "不用从头跑", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 792.763, "end": 794.303, "text": "那第三层", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 794.303, "end": 796.043, "text": "则是只支持检索的", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 796.043, "end": 797.223, "text": "完整历史对话转录", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 797.223, "end": 798.823, "text": "不用到的历史记录", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 798.823, "end": 799.663, "text": "全部打包归档", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 799.663, "end": 801.343, "text": "只有需要的时候才去搜索", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 801.343, "end": 802.743, "text": "通过这种分层", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 802.743, "end": 804.563, "text": "我们既能处理长周期的任务", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 804.563, "end": 806.723, "text": "又能帮公司省下一大笔的投肯开销", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 806.723, "end": 809.043, "text": "不过有了分层记忆", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 809.043, "end": 810.923, "text": "如果用户一直跟智能体聊下去", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 810.923, "end": 812.103, "text": "聊了上百轮", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 812.103, "end": 813.583, "text": "当前的绘画窗口", "chunk": 1, "language": "zh", "avg_logprob": -0.08779157782500645, "no_speech_prob": 1.6067281186082916e-11, "compression_ratio": 1.3478260869565217 }, { "start": 813.583, "end": 815.283, "text": "还是会面临一个爆满的风险", "chunk": 1, "language": "zh", "avg_logprob": -0.10935474531299572, "no_speech_prob": 1.7856176606767526e-11, "compression_ratio": 1.42433234421365 }, { "start": 815.283, "end": 818.183, "text": "这时候我们要怎么防止上下纹爆掉呢", "chunk": 1, "language": "zh", "avg_logprob": -0.10935474531299572, "no_speech_prob": 1.7856176606767526e-11, "compression_ratio": 1.42433234421365 }, { "start": 818.183, "end": 821.403, "text": "其实智能体在这一方面和我们人类很像", "chunk": 1, "language": "zh", "avg_logprob": -0.10935474531299572, "no_speech_prob": 1.7856176606767526e-11, "compression_ratio": 1.42433234421365 }, { "start": 821.403, "end": 824.163, "text": "人总不能一直不睡觉拼命公渡对吧", "chunk": 1, "language": "zh", "avg_logprob": -0.10935474531299572, "no_speech_prob": 1.7856176606767526e-11, "compression_ratio": 1.42433234421365 }, { "start": 824.163, "end": 826.623, "text": "所以我们可以给智能体引入一个", "chunk": 1, "language": "zh", "avg_logprob": -0.10935474531299572, "no_speech_prob": 1.7856176606767526e-11, "compression_ratio": 1.42433234421365 }, { "start": 826.623, "end": 827.963, "text": "记忆睡眠巩固机制", "chunk": 1, "language": "zh", "avg_logprob": -0.10935474531299572, "no_speech_prob": 1.7856176606767526e-11, "compression_ratio": 1.42433234421365 }, { "start": 827.963, "end": 830.643, "text": "在系统空闲的时候默默运行", "chunk": 1, "language": "zh", "avg_logprob": -0.10935474531299572, "no_speech_prob": 1.7856176606767526e-11, "compression_ratio": 1.42433234421365 }, { "start": 830.643, "end": 832.703, "text": "我们可以让一个后台智能体", "chunk": 1, "language": "zh", "avg_logprob": -0.10935474531299572, "no_speech_prob": 1.7856176606767526e-11, "compression_ratio": 1.42433234421365 }, { "start": 832.703, "end": 834.263, "text": "在用户不说话的闲时", "chunk": 1, "language": "zh", "avg_logprob": -0.10935474531299572, "no_speech_prob": 1.7856176606767526e-11, "compression_ratio": 1.42433234421365 }, { "start": 834.263, "end": 836.063, "text": "去帮我们整理零段的上下纹", "chunk": 1, "language": "zh", "avg_logprob": -0.10935474531299572, "no_speech_prob": 1.7856176606767526e-11, "compression_ratio": 1.42433234421365 }, { "start": 836.063, "end": 839.643, "text": "它会自动把过期的冲突的事实给修剪掉", "chunk": 1, "language": "zh", "avg_logprob": -0.10935474531299572, "no_speech_prob": 1.7856176606767526e-11, "compression_ratio": 1.42433234421365 }, { "start": 839.643, "end": 842.423, "text": "去虫之后合并成一个干净的核心索引", "chunk": 1, "language": "zh", "avg_logprob": -0.10935474531299572, "no_speech_prob": 1.7856176606767526e-11, "compression_ratio": 1.42433234421365 }, { "start": 842.423, "end": 844.563, "text": "这就像我们睡觉做梦的时候", "chunk": 1, "language": "zh", "avg_logprob": -0.1035150330642174, "no_speech_prob": 1.6836374308604185e-11, "compression_ratio": 1.3734939759036144 }, { "start": 844.563, "end": 846.383, "text": "大脑会自动整理白天的记忆", "chunk": 1, "language": "zh", "avg_logprob": -0.1035150330642174, "no_speech_prob": 1.6836374308604185e-11, "compression_ratio": 1.3734939759036144 }, { "start": 846.383, "end": 847.583, "text": "丢弃无用信息一样", "chunk": 1, "language": "zh", "avg_logprob": -0.1035150330642174, "no_speech_prob": 1.6836374308604185e-11, "compression_ratio": 1.3734939759036144 }, { "start": 847.583, "end": 849.583, "text": "这是一种应用的自我进化", "chunk": 1, "language": "zh", "avg_logprob": -0.1035150330642174, "no_speech_prob": 1.6836374308604185e-11, "compression_ratio": 1.3734939759036144 }, { "start": 849.583, "end": 851.603, "text": "同时在绘画进行中", "chunk": 1, "language": "zh", "avg_logprob": -0.1035150330642174, "no_speech_prob": 1.6836374308604185e-11, "compression_ratio": 1.3734939759036144 }, { "start": 851.603, "end": 853.863, "text": "我们还要配合渐进式上下文压缩", "chunk": 1, "language": "zh", "avg_logprob": -0.1035150330642174, "no_speech_prob": 1.6836374308604185e-11, "compression_ratio": 1.3734939759036144 }, { "start": 853.863, "end": 855.883, "text": "随着对话轮次的增加", "chunk": 1, "language": "zh", "avg_logprob": -0.1035150330642174, "no_speech_prob": 1.6836374308604185e-11, "compression_ratio": 1.3734939759036144 }, { "start": 855.883, "end": 858.483, "text": "我们不是粗暴的直接截掉前面的对话", "chunk": 1, "language": "zh", "avg_logprob": -0.1035150330642174, "no_speech_prob": 1.6836374308604185e-11, "compression_ratio": 1.3734939759036144 }, { "start": 858.483, "end": 861.163, "text": "而是施加不同强度的压缩", "chunk": 1, "language": "zh", "avg_logprob": -0.1035150330642174, "no_speech_prob": 1.6836374308604185e-11, "compression_ratio": 1.3734939759036144 }, { "start": 861.163, "end": 863.183, "text": "近期对话保留最详细的明气", "chunk": 1, "language": "zh", "avg_logprob": -0.1035150330642174, "no_speech_prob": 1.6836374308604185e-11, "compression_ratio": 1.3734939759036144 }, { "start": 863.183, "end": 865.203, "text": "中期对话做轻度总结", "chunk": 1, "language": "zh", "avg_logprob": -0.1035150330642174, "no_speech_prob": 1.6836374308604185e-11, "compression_ratio": 1.3734939759036144 }, { "start": 865.203, "end": 867.663, "text": "而远期对话则进行一个极限折叠", "chunk": 1, "language": "zh", "avg_logprob": -0.1035150330642174, "no_speech_prob": 1.6836374308604185e-11, "compression_ratio": 1.3734939759036144 }, { "start": 867.663, "end": 869.263, "text": "这样层层递进", "chunk": 1, "language": "zh", "avg_logprob": -0.1035150330642174, "no_speech_prob": 1.6836374308604185e-11, "compression_ratio": 1.3734939759036144 }, { "start": 869.263, "end": 870.903, "text": "既保证了模型不会失忆", "chunk": 1, "language": "zh", "avg_logprob": -0.1035150330642174, "no_speech_prob": 1.6836374308604185e-11, "compression_ratio": 1.3734939759036144 }, { "start": 870.903, "end": 873.443, "text": "又彻底解决了上下文窗口爆棚的问题", "chunk": 1, "language": "zh", "avg_logprob": -0.08715771183823094, "no_speech_prob": 1.4571397907725547e-11, "compression_ratio": 1.2822299651567943 }, { "start": 873.443, "end": 874.663, "text": "OK", "chunk": 1, "language": "zh", "avg_logprob": -0.08715771183823094, "no_speech_prob": 1.4571397907725547e-11, "compression_ratio": 1.2822299651567943 }, { "start": 874.663, "end": 876.063, "text": "在生产环境里", "chunk": 1, "language": "zh", "avg_logprob": -0.08715771183823094, "no_speech_prob": 1.4571397907725547e-11, "compression_ratio": 1.2822299651567943 }, { "start": 876.063, "end": 876.963, "text": "除了记忆管理", "chunk": 1, "language": "zh", "avg_logprob": -0.08715771183823094, "no_speech_prob": 1.4571397907725547e-11, "compression_ratio": 1.2822299651567943 }, { "start": 876.963, "end": 878.303, "text": "我们还要解决两个问题", "chunk": 1, "language": "zh", "avg_logprob": -0.08715771183823094, "no_speech_prob": 1.4571397907725547e-11, "compression_ratio": 1.2822299651567943 }, { "start": 878.303, "end": 880.063, "text": "一个是系统的安全合规", "chunk": 1, "language": "zh", "avg_logprob": -0.08715771183823094, "no_speech_prob": 1.4571397907725547e-11, "compression_ratio": 1.2822299651567943 }, { "start": 880.063, "end": 882.243, "text": "另一个是高昂的API障碍成本", "chunk": 1, "language": "zh", "avg_logprob": -0.08715771183823094, "no_speech_prob": 1.4571397907725547e-11, "compression_ratio": 1.2822299651567943 }, { "start": 882.243, "end": 883.763, "text": "首先是系统安全", "chunk": 1, "language": "zh", "avg_logprob": -0.08715771183823094, "no_speech_prob": 1.4571397907725547e-11, "compression_ratio": 1.2822299651567943 }, { "start": 883.763, "end": 886.283, "text": "我们不能全指望prompt去叮嘱大模型", "chunk": 1, "language": "zh", "avg_logprob": -0.08715771183823094, "no_speech_prob": 1.4571397907725547e-11, "compression_ratio": 1.2822299651567943 }, { "start": 886.283, "end": 887.183, "text": "你不要违规", "chunk": 1, "language": "zh", "avg_logprob": -0.08715771183823094, "no_speech_prob": 1.4571397907725547e-11, "compression_ratio": 1.2822299651567943 }, { "start": 887.183, "end": 888.383, "text": "格式千万要写对", "chunk": 1, "language": "zh", "avg_logprob": -0.08715771183823094, "no_speech_prob": 1.4571397907725547e-11, "compression_ratio": 1.2822299651567943 }, { "start": 888.383, "end": 890.483, "text": "因为模型总有遗忘的时候", "chunk": 1, "language": "zh", "avg_logprob": -0.08715771183823094, "no_speech_prob": 1.4571397907725547e-11, "compression_ratio": 1.2822299651567943 }, { "start": 890.483, "end": 891.743, "text": "我们的原则是", "chunk": 1, "language": "zh", "avg_logprob": -0.08715771183823094, "no_speech_prob": 1.4571397907725547e-11, "compression_ratio": 1.2822299651567943 }, { "start": 891.743, "end": 893.807, "text": "必须把硬逻辑剥离到prompt", "chunk": 1, "language": "zh", "avg_logprob": -0.08715771183823094, "no_speech_prob": 1.4571397907725547e-11, "compression_ratio": 1.2822299651567943 }, { "start": 894.807, "end": 896.367, "text": "用一个确定性的生命周期钩子。", "chunk": 2, "language": "zh", "avg_logprob": -0.15874648977209022, "no_speech_prob": 2.0695169536399938e-11, "compression_ratio": 1.4232954545454546 }, { "start": 897.027, "end": 899.007, "text": "比如,在大模型调用外部工具之前,", "chunk": 2, "language": "zh", "avg_logprob": -0.15874648977209022, "no_speech_prob": 2.0695169536399938e-11, "compression_ratio": 1.4232954545454546 }, { "start": 899.507, "end": 901.787, "text": "系统在底层强行的插入一个前置钩子,", "chunk": 2, "language": "zh", "avg_logprob": -0.15874648977209022, "no_speech_prob": 2.0695169536399938e-11, "compression_ratio": 1.4232954545454546 }, { "start": 902.147, "end": 903.407, "text": "自动降业命令安不安全。", "chunk": 2, "language": "zh", "avg_logprob": -0.15874648977209022, "no_speech_prob": 2.0695169536399938e-11, "compression_ratio": 1.4232954545454546 }, { "start": 904.167, "end": 905.147, "text": "工具跑完了之后,", "chunk": 2, "language": "zh", "avg_logprob": -0.15874648977209022, "no_speech_prob": 2.0695169536399938e-11, "compression_ratio": 1.4232954545454546 }, { "start": 905.387, "end": 907.207, "text": "我们再强行插入一个后置钩子,", "chunk": 2, "language": "zh", "avg_logprob": -0.15874648977209022, "no_speech_prob": 2.0695169536399938e-11, "compression_ratio": 1.4232954545454546 }, { "start": 907.207, "end": 909.707, "text": "自动去跑一次代码格式化和测试", "chunk": 2, "language": "zh", "avg_logprob": -0.08879641673011238, "no_speech_prob": 1.6245746067844458e-11, "compression_ratio": 1.3658536585365855 }, { "start": 909.707, "end": 910.947, "text": "从再配置文件", "chunk": 2, "language": "zh", "avg_logprob": -0.08879641673011238, "no_speech_prob": 1.6245746067844458e-11, "compression_ratio": 1.3658536585365855 }, { "start": 910.947, "end": 914.187, "text": "这些钩子是在系统底层用代码硬编码的", "chunk": 2, "language": "zh", "avg_logprob": -0.08879641673011238, "no_speech_prob": 1.6245746067844458e-11, "compression_ratio": 1.3658536585365855 }, { "start": 914.187, "end": 916.067, "text": "不管大模型怎么产生幻觉", "chunk": 2, "language": "zh", "avg_logprob": -0.08879641673011238, "no_speech_prob": 1.6245746067844458e-11, "compression_ratio": 1.3658536585365855 }, { "start": 916.067, "end": 918.147, "text": "这些安全红线它绝对越不过去", "chunk": 2, "language": "zh", "avg_logprob": -0.08879641673011238, "no_speech_prob": 1.6245746067844458e-11, "compression_ratio": 1.3658536585365855 }, { "start": 918.147, "end": 920.407, "text": "然后就是商业化成本的控制", "chunk": 2, "language": "zh", "avg_logprob": -0.08879641673011238, "no_speech_prob": 1.6245746067844458e-11, "compression_ratio": 1.3658536585365855 }, { "start": 920.407, "end": 922.687, "text": "如果每天接上万个用户请求", "chunk": 2, "language": "zh", "avg_logprob": -0.08879641673011238, "no_speech_prob": 1.6245746067844458e-11, "compression_ratio": 1.3658536585365855 }, { "start": 922.687, "end": 925.227, "text": "所有请求都用最贵最强的大模型", "chunk": 2, "language": "zh", "avg_logprob": -0.08879641673011238, "no_speech_prob": 1.6245746067844458e-11, "compression_ratio": 1.3658536585365855 }, { "start": 925.227, "end": 926.947, "text": "公司可能很快就撑不住了", "chunk": 2, "language": "zh", "avg_logprob": -0.08879641673011238, "no_speech_prob": 1.6245746067844458e-11, "compression_ratio": 1.3658536585365855 }, { "start": 926.947, "end": 929.527, "text": "所以我们要设计一个混合模型路由机制", "chunk": 2, "language": "zh", "avg_logprob": -0.08879641673011238, "no_speech_prob": 1.6245746067844458e-11, "compression_ratio": 1.3658536585365855 }, { "start": 929.527, "end": 931.667, "text": "当用户的请求进来的时候", "chunk": 2, "language": "zh", "avg_logprob": -0.08879641673011238, "no_speech_prob": 1.6245746067844458e-11, "compression_ratio": 1.3658536585365855 }, { "start": 931.667, "end": 933.187, "text": "先过一个路由节点", "chunk": 2, "language": "zh", "avg_logprob": -0.08879641673011238, "no_speech_prob": 1.6245746067844458e-11, "compression_ratio": 1.3658536585365855 }, { "start": 933.187, "end": 934.487, "text": "做动态难度分发", "chunk": 2, "language": "zh", "avg_logprob": -0.08879641673011238, "no_speech_prob": 1.6245746067844458e-11, "compression_ratio": 1.3658536585365855 }, { "start": 934.487, "end": 937.127, "text": "如果是写系统架构这种高难度任务", "chunk": 2, "language": "zh", "avg_logprob": -0.08879641673011238, "no_speech_prob": 1.6245746067844458e-11, "compression_ratio": 1.3658536585365855 }, { "start": 937.127, "end": 939.827, "text": "路由就把请求分发给大型颗粒模型", "chunk": 2, "language": "zh", "avg_logprob": -0.09950016129691645, "no_speech_prob": 1.6575940273155787e-11, "compression_ratio": 1.314207650273224 }, { "start": 939.827, "end": 942.007, "text": "那如果是问路打招呼", "chunk": 2, "language": "zh", "avg_logprob": -0.09950016129691645, "no_speech_prob": 1.6575940273155787e-11, "compression_ratio": 1.314207650273224 }, { "start": 942.007, "end": 944.107, "text": "或者是简单做一个格式化这种任务", "chunk": 2, "language": "zh", "avg_logprob": -0.09950016129691645, "no_speech_prob": 1.6575940273155787e-11, "compression_ratio": 1.314207650273224 }, { "start": 944.107, "end": 946.127, "text": "路由就自动分流给急速小模型", "chunk": 2, "language": "zh", "avg_logprob": -0.09950016129691645, "no_speech_prob": 1.6575940273155787e-11, "compression_ratio": 1.314207650273224 }, { "start": 946.127, "end": 948.187, "text": "或者直接命中我们的提示词缓存", "chunk": 2, "language": "zh", "avg_logprob": -0.09950016129691645, "no_speech_prob": 1.6575940273155787e-11, "compression_ratio": 1.314207650273224 }, { "start": 948.187, "end": 950.947, "text": "这样既兼顾了系统的高并发响应", "chunk": 2, "language": "zh", "avg_logprob": -0.09950016129691645, "no_speech_prob": 1.6575940273155787e-11, "compression_ratio": 1.314207650273224 }, { "start": 950.947, "end": 953.047, "text": "又帮我们省下了大量的API资费", "chunk": 2, "language": "zh", "avg_logprob": -0.09950016129691645, "no_speech_prob": 1.6575940273155787e-11, "compression_ratio": 1.314207650273224 }, { "start": 953.047, "end": 954.867, "text": "这才是合格的商业级架构", "chunk": 2, "language": "zh", "avg_logprob": -0.09950016129691645, "no_speech_prob": 1.6575940273155787e-11, "compression_ratio": 1.314207650273224 }, { "start": 954.867, "end": 957.167, "text": "而在高频的业务场景里", "chunk": 2, "language": "zh", "avg_logprob": -0.09950016129691645, "no_speech_prob": 1.6575940273155787e-11, "compression_ratio": 1.314207650273224 }, { "start": 957.167, "end": 958.967, "text": "我们最常碰到的落地项目", "chunk": 2, "language": "zh", "avg_logprob": -0.09950016129691645, "no_speech_prob": 1.6575940273155787e-11, "compression_ratio": 1.314207650273224 }, { "start": 958.967, "end": 960.227, "text": "其实还是企业知识库", "chunk": 2, "language": "zh", "avg_logprob": -0.09950016129691645, "no_speech_prob": 1.6575940273155787e-11, "compression_ratio": 1.314207650273224 }, { "start": 960.227, "end": 961.547, "text": "也就是RG系统", "chunk": 2, "language": "zh", "avg_logprob": -0.09950016129691645, "no_speech_prob": 1.6575940273155787e-11, "compression_ratio": 1.314207650273224 }, { "start": 961.547, "end": 963.567, "text": "但是在生产环境里", "chunk": 2, "language": "zh", "avg_logprob": -0.09950016129691645, "no_speech_prob": 1.6575940273155787e-11, "compression_ratio": 1.314207650273224 }, { "start": 963.567, "end": 965.787, "text": "一个合格的RG绝对不是单向的", "chunk": 2, "language": "zh", "avg_logprob": -0.09950016129691645, "no_speech_prob": 1.6575940273155787e-11, "compression_ratio": 1.314207650273224 }, { "start": 965.787, "end": 966.907, "text": "检索完就丢给模型", "chunk": 2, "language": "zh", "avg_logprob": -0.08895785372022172, "no_speech_prob": 2.420642505884807e-11, "compression_ratio": 1.4367469879518073 }, { "start": 966.907, "end": 969.267, "text": "它应该是一个能自我进化的闭环", "chunk": 2, "language": "zh", "avg_logprob": -0.08895785372022172, "no_speech_prob": 2.420642505884807e-11, "compression_ratio": 1.4367469879518073 }, { "start": 969.267, "end": 970.647, "text": "在生产环境里", "chunk": 2, "language": "zh", "avg_logprob": -0.08895785372022172, "no_speech_prob": 2.420642505884807e-11, "compression_ratio": 1.4367469879518073 }, { "start": 970.647, "end": 973.087, "text": "我们要打通混合检索和重排架构", "chunk": 2, "language": "zh", "avg_logprob": -0.08895785372022172, "no_speech_prob": 2.420642505884807e-11, "compression_ratio": 1.4367469879518073 }, { "start": 973.087, "end": 975.707, "text": "形成一个数据沉淀模型反母的飞轮", "chunk": 2, "language": "zh", "avg_logprob": -0.08895785372022172, "no_speech_prob": 2.420642505884807e-11, "compression_ratio": 1.4367469879518073 }, { "start": 975.707, "end": 977.387, "text": "第一步混合检索", "chunk": 2, "language": "zh", "avg_logprob": -0.08895785372022172, "no_speech_prob": 2.420642505884807e-11, "compression_ratio": 1.4367469879518073 }, { "start": 977.387, "end": 979.007, "text": "当用户提出问题的时候", "chunk": 2, "language": "zh", "avg_logprob": -0.08895785372022172, "no_speech_prob": 2.420642505884807e-11, "compression_ratio": 1.4367469879518073 }, { "start": 979.007, "end": 981.287, "text": "我们把向量检索和关键词检索", "chunk": 2, "language": "zh", "avg_logprob": -0.08895785372022172, "no_speech_prob": 2.420642505884807e-11, "compression_ratio": 1.4367469879518073 }, { "start": 981.287, "end": 982.387, "text": "融合在一起", "chunk": 2, "language": "zh", "avg_logprob": -0.08895785372022172, "no_speech_prob": 2.420642505884807e-11, "compression_ratio": 1.4367469879518073 }, { "start": 982.387, "end": 984.107, "text": "保证无论用户怎么提问", "chunk": 2, "language": "zh", "avg_logprob": -0.08895785372022172, "no_speech_prob": 2.420642505884807e-11, "compression_ratio": 1.4367469879518073 }, { "start": 984.107, "end": 986.307, "text": "我们都能精准定位底层的原始文档", "chunk": 2, "language": "zh", "avg_logprob": -0.08895785372022172, "no_speech_prob": 2.420642505884807e-11, "compression_ratio": 1.4367469879518073 }, { "start": 986.307, "end": 988.707, "text": "然后第二步重排提纯", "chunk": 2, "language": "zh", "avg_logprob": -0.08895785372022172, "no_speech_prob": 2.420642505884807e-11, "compression_ratio": 1.4367469879518073 }, { "start": 988.707, "end": 991.347, "text": "搜出来的文档可能有几十个", "chunk": 2, "language": "zh", "avg_logprob": -0.08895785372022172, "no_speech_prob": 2.420642505884807e-11, "compression_ratio": 1.4367469879518073 }, { "start": 991.347, "end": 993.107, "text": "那模型看不完也看不过来", "chunk": 2, "language": "zh", "avg_logprob": -0.08895785372022172, "no_speech_prob": 2.420642505884807e-11, "compression_ratio": 1.4367469879518073 }, { "start": 993.367, "end": 995.187, "text": "我们用专门的重排模型", "chunk": 2, "language": "zh", "avg_logprob": -0.08895785372022172, "no_speech_prob": 2.420642505884807e-11, "compression_ratio": 1.4367469879518073 }, { "start": 995.187, "end": 997.047, "text": "挑选出最精准的几个节点", "chunk": 2, "language": "zh", "avg_logprob": -0.08474444071451823, "no_speech_prob": 1.669622946842697e-11, "compression_ratio": 1.3588235294117648 }, { "start": 997.047, "end": 998.307, "text": "送进大幕型的上下文窗口", "chunk": 2, "language": "zh", "avg_logprob": -0.08474444071451823, "no_speech_prob": 1.669622946842697e-11, "compression_ratio": 1.3588235294117648 }, { "start": 998.307, "end": 999.847, "text": "接下来第三步", "chunk": 2, "language": "zh", "avg_logprob": -0.08474444071451823, "no_speech_prob": 1.669622946842697e-11, "compression_ratio": 1.3588235294117648 }, { "start": 999.847, "end": 1001.147, "text": "就是结合业务逻辑", "chunk": 2, "language": "zh", "avg_logprob": -0.08474444071451823, "no_speech_prob": 1.669622946842697e-11, "compression_ratio": 1.3588235294117648 }, { "start": 1001.147, "end": 1002.987, "text": "生成最终的一个融合输出", "chunk": 2, "language": "zh", "avg_logprob": -0.08474444071451823, "no_speech_prob": 1.669622946842697e-11, "compression_ratio": 1.3588235294117648 }, { "start": 1002.987, "end": 1004.647, "text": "重点在第四步", "chunk": 2, "language": "zh", "avg_logprob": -0.08474444071451823, "no_speech_prob": 1.669622946842697e-11, "compression_ratio": 1.3588235294117648 }, { "start": 1004.647, "end": 1005.387, "text": "数据沉淀", "chunk": 2, "language": "zh", "avg_logprob": -0.08474444071451823, "no_speech_prob": 1.669622946842697e-11, "compression_ratio": 1.3588235294117648 }, { "start": 1005.387, "end": 1007.887, "text": "我们要补获用户的真实反馈", "chunk": 2, "language": "zh", "avg_logprob": -0.08474444071451823, "no_speech_prob": 1.669622946842697e-11, "compression_ratio": 1.3588235294117648 }, { "start": 1007.887, "end": 1009.687, "text": "记录高频的查询盲区", "chunk": 2, "language": "zh", "avg_logprob": -0.08474444071451823, "no_speech_prob": 1.669622946842697e-11, "compression_ratio": 1.3588235294117648 }, { "start": 1009.687, "end": 1011.227, "text": "那最后一步", "chunk": 2, "language": "zh", "avg_logprob": -0.08474444071451823, "no_speech_prob": 1.669622946842697e-11, "compression_ratio": 1.3588235294117648 }, { "start": 1011.227, "end": 1012.187, "text": "反补迭代", "chunk": 2, "language": "zh", "avg_logprob": -0.08474444071451823, "no_speech_prob": 1.669622946842697e-11, "compression_ratio": 1.3588235294117648 }, { "start": 1012.187, "end": 1013.587, "text": "根据这些反馈", "chunk": 2, "language": "zh", "avg_logprob": -0.08474444071451823, "no_speech_prob": 1.669622946842697e-11, "compression_ratio": 1.3588235294117648 }, { "start": 1013.587, "end": 1014.647, "text": "自动去更新", "chunk": 2, "language": "zh", "avg_logprob": -0.08474444071451823, "no_speech_prob": 1.669622946842697e-11, "compression_ratio": 1.3588235294117648 }, { "start": 1014.647, "end": 1016.287, "text": "修正和强化我们的企业知识库", "chunk": 2, "language": "zh", "avg_logprob": -0.08474444071451823, "no_speech_prob": 1.669622946842697e-11, "compression_ratio": 1.3588235294117648 }, { "start": 1016.287, "end": 1018.867, "text": "那么当这一套流程运转起来之后", "chunk": 2, "language": "zh", "avg_logprob": -0.08474444071451823, "no_speech_prob": 1.669622946842697e-11, "compression_ratio": 1.3588235294117648 }, { "start": 1018.867, "end": 1020.727, "text": "知识库里的盲区就会越来越少", "chunk": 2, "language": "zh", "avg_logprob": -0.08474444071451823, "no_speech_prob": 1.669622946842697e-11, "compression_ratio": 1.3588235294117648 }, { "start": 1020.727, "end": 1022.347, "text": "下一轮检索就会更准", "chunk": 2, "language": "zh", "avg_logprob": -0.08474444071451823, "no_speech_prob": 1.669622946842697e-11, "compression_ratio": 1.3588235294117648 }, { "start": 1022.347, "end": 1024.147, "text": "我们的应用也会越用越聪明", "chunk": 2, "language": "zh", "avg_logprob": -0.10680085609401513, "no_speech_prob": 1.612115996252328e-11, "compression_ratio": 1.3623595505617978 }, { "start": 1024.147, "end": 1025.847, "text": "好 聊到这里", "chunk": 2, "language": "zh", "avg_logprob": -0.10680085609401513, "no_speech_prob": 1.612115996252328e-11, "compression_ratio": 1.3623595505617978 }, { "start": 1025.847, "end": 1027.647, "text": "我们基本上把一个工业级的", "chunk": 2, "language": "zh", "avg_logprob": -0.10680085609401513, "no_speech_prob": 1.612115996252328e-11, "compression_ratio": 1.3623595505617978 }, { "start": 1027.647, "end": 1028.847, "text": "大模型应用开发工程师", "chunk": 2, "language": "zh", "avg_logprob": -0.10680085609401513, "no_speech_prob": 1.612115996252328e-11, "compression_ratio": 1.3623595505617978 }, { "start": 1028.847, "end": 1030.727, "text": "需要具备的能力全都梳理了一遍", "chunk": 2, "language": "zh", "avg_logprob": -0.10680085609401513, "no_speech_prob": 1.612115996252328e-11, "compression_ratio": 1.3623595505617978 }, { "start": 1030.727, "end": 1032.567, "text": "但最后我们来做一个复盘", "chunk": 2, "language": "zh", "avg_logprob": -0.10680085609401513, "no_speech_prob": 1.612115996252328e-11, "compression_ratio": 1.3623595505617978 }, { "start": 1032.567, "end": 1034.947, "text": "其实一个真正合格", "chunk": 2, "language": "zh", "avg_logprob": -0.10680085609401513, "no_speech_prob": 1.612115996252328e-11, "compression_ratio": 1.3623595505617978 }, { "start": 1034.947, "end": 1037.467, "text": "能够帮企业解决实际大模型落地的工程师", "chunk": 2, "language": "zh", "avg_logprob": -0.10680085609401513, "no_speech_prob": 1.612115996252328e-11, "compression_ratio": 1.3623595505617978 }, { "start": 1037.467, "end": 1040.147, "text": "它的成长路径其实是非常清晰的", "chunk": 2, "language": "zh", "avg_logprob": -0.10680085609401513, "no_speech_prob": 1.612115996252328e-11, "compression_ratio": 1.3623595505617978 }, { "start": 1040.147, "end": 1041.167, "text": "在业务架构上", "chunk": 2, "language": "zh", "avg_logprob": -0.10680085609401513, "no_speech_prob": 1.612115996252328e-11, "compression_ratio": 1.3623595505617978 }, { "start": 1041.167, "end": 1043.447, "text": "我们要从只会写单一prompt", "chunk": 2, "language": "zh", "avg_logprob": -0.10680085609401513, "no_speech_prob": 1.612115996252328e-11, "compression_ratio": 1.3623595505617978 }, { "start": 1043.447, "end": 1045.587, "text": "升级到能够驾驭多智能体编排", "chunk": 2, "language": "zh", "avg_logprob": -0.10680085609401513, "no_speech_prob": 1.612115996252328e-11, "compression_ratio": 1.3623595505617978 }, { "start": 1045.587, "end": 1047.427, "text": "在工具与协议上", "chunk": 2, "language": "zh", "avg_logprob": -0.10680085609401513, "no_speech_prob": 1.612115996252328e-11, "compression_ratio": 1.3623595505617978 }, { "start": 1047.427, "end": 1049.907, "text": "我们要从硬编码施API", "chunk": 2, "language": "zh", "avg_logprob": -0.10680085609401513, "no_speech_prob": 1.612115996252328e-11, "compression_ratio": 1.3623595505617978 }, { "start": 1049.907, "end": 1052.067, "text": "升级到运用标准的AMCP协议", "chunk": 2, "language": "zh", "avg_logprob": -0.10680085609401513, "no_speech_prob": 1.612115996252328e-11, "compression_ratio": 1.3623595505617978 }, { "start": 1052.067, "end": 1053.347, "text": "和单一职责工具", "chunk": 2, "language": "zh", "avg_logprob": -0.08183592938362284, "no_speech_prob": 1.2998851994794336e-11, "compression_ratio": 1.2678571428571428 }, { "start": 1053.347, "end": 1055.027, "text": "在生产机状态上", "chunk": 2, "language": "zh", "avg_logprob": -0.08183592938362284, "no_speech_prob": 1.2998851994794336e-11, "compression_ratio": 1.2678571428571428 }, { "start": 1055.027, "end": 1057.027, "text": "我们要从粗暴的内存阶段", "chunk": 2, "language": "zh", "avg_logprob": -0.08183592938362284, "no_speech_prob": 1.2998851994794336e-11, "compression_ratio": 1.2678571428571428 }, { "start": 1057.027, "end": 1059.987, "text": "升级到分层记忆和动态上下纹压缩", "chunk": 2, "language": "zh", "avg_logprob": -0.08183592938362284, "no_speech_prob": 1.2998851994794336e-11, "compression_ratio": 1.2678571428571428 }, { "start": 1059.987, "end": 1061.387, "text": "那在系统级整合上", "chunk": 2, "language": "zh", "avg_logprob": -0.08183592938362284, "no_speech_prob": 1.2998851994794336e-11, "compression_ratio": 1.2678571428571428 }, { "start": 1061.387, "end": 1063.067, "text": "我们要从基础的检索", "chunk": 2, "language": "zh", "avg_logprob": -0.08183592938362284, "no_speech_prob": 1.2998851994794336e-11, "compression_ratio": 1.2678571428571428 }, { "start": 1063.067, "end": 1066.027, "text": "升级到刚才讲的RG飞轮和模型路由", "chunk": 2, "language": "zh", "avg_logprob": -0.08183592938362284, "no_speech_prob": 1.2998851994794336e-11, "compression_ratio": 1.2678571428571428 }, { "start": 1066.027, "end": 1068.027, "text": "然后在质量可观测性上", "chunk": 2, "language": "zh", "avg_logprob": -0.08183592938362284, "no_speech_prob": 1.2998851994794336e-11, "compression_ratio": 1.2678571428571428 }, { "start": 1068.027, "end": 1070.307, "text": "升级到全列路追踪和量化评测", "chunk": 2, "language": "zh", "avg_logprob": -0.08183592938362284, "no_speech_prob": 1.2998851994794336e-11, "compression_ratio": 1.2678571428571428 } ], "segmented_transcription": { "status": "SEGMENTED_TRANSCRIPT_OK", "created_at": "2026-08-05T16:57:29", "manifest": "/Users/sagawa/AI_WORK/video_notes/url/20260805_165643__我愿称之为AI圈最具诚意的入门教程_ai_程序员_人工智能_agent/segmented_work/chunk_manifest.json", "audio": "/Users/sagawa/AI_WORK/video_notes/url/20260805_165643__我愿称之为AI圈最具诚意的入门教程_ai_程序员_人工智能_agent/audio_16k_mono.wav", "audio_sha256": "c9c67d324402e9ef194dafb8c26629a0ceb621adfe554fd30e449ad4302de000", "audio_duration": 1073.806813, "model": "mlx-community/whisper-large-v3-turbo", "language_hint": "zh", "condition_on_previous_text": false, "chunk_count": 3, "resumed_chunks": 0, "merged_segment_count": 575, "final_segment_end": 1070.307, "tail_repeat_run": 1, "tail_repeat_text": "根据这些反馈", "qc_ok": true, "qc_errors": [] } }