{
  "text": "你有没有过这种很无力的经验?\n叫AI帮你整理一份跨部门的会议报告。\n做到一半它忘了你要求的章节格式,用语音助手查形成。\n旁边同事插句话,它就把指令听成完全不相干的内容。\n只要任务超过三个步骤,需要跨两个以上工具,最后交付的结果一定歪掉大半。\n你大概也听过很多人说AI已经很聪明了,但真的拿来干活,总是卡在这些最不起眼的小地方。\n2026年9月的头两天\nMeta接连丢出了两项更新\n没有吹什么通用人工智能的牛皮\n却精准踩中了这些大家天天吐槽的痛点\n一个负责把真实世界里的声音听得更清楚\n一个负责把复杂的常任务做得更稳\n有人看完数据说Meta终于要在代理人战场超车了\n也有人冷静算完成本\n发现事情根本没有这么简单\n今天我们就不吹不黑\n把官方技术文件第三方独立评测\n还有那些藏在角落的小字限制\n全部摊开来讲\n看看这两款产品到底走到了哪一步\n又有哪些东西是营销不会告诉你的\n我们不从结论讲起\n因为真正有价值的判断往往都藏在每一个\n但是后面\n先帮大家把最容易搞混的产品线界限画清楚\n这也是很多自媒体最常讲错的地方\n9月1日登场的是Muse Voice Transcribe语音转写模型\n9月2日接棒上线的是Muse Spark 1.3代理人模型\n两天连续发布核心产品\n这在Meta Superintelligence Labs的历史上并不常见\n也能看出他们在2026年下半年的竞争焦虑\n但更重要的是\n很多人到现在还搞不清楚\nMuse Spark和Muse Glimmer\n根本是两条完全不同的路线\n不能混为一谈\nMuse Spark走的是云端前沿代理人路线\n权重完全必圆\n目前只开放API呼叫\n官方只说未来有开放权重的规划\n但既没有时间表\n也没说会用哪种授权条款\n所以你绝对不能把它理解成\n马上就可以下载来本地跑\n而Muse Glimmer是拥有300亿参数\n采用Apache 2.0授权的本地方案\n主打消费级显示卡就能执行\n讲白一点\nSpark负责把能力堆到最高\n强企业级和高端开发者市场\nGlimmer负责把能力扑到最广\n占领开源生态和终端装置\n这种双轨策略其实刚好对应了Meta在AI产业的定位转变\n前几年大家提到Meta的AI第一个想到的就是Lama系列开源模型\n靠免费打天下\n跟OpenAI的全B原路线正面抗衡\n但到了2026年\n整个行业的竞争已经从比参数 比跑分\n进到比代理人 比落地场景的阶段\nMeta必须拿出足够强的B原顶端产品\n才能跟GPT系列\nGemini系列抢商业客户\n同时又不能丢掉开源生态的优势\n为什么要花这么多时间讲这件事\n因为如果你误以为\nSpark 1.3是开源模型\n就会产生错误的部署预期\n如果你以为Glimmer只是Spark的缩小版\n就会低估两者在设计目标上的根本差异\n另外从迭代节奏来看\nMuse系列从7月的1.1版\n首次连上邮件行事例\n并能制作投影片到8月初1.2版聚焦大型程式库与复杂真错\n甚至能连续1000多次工具呼叫\n在24小时内反复编写编译GPU核心程式\n8月中再加入图像与视频理解能力\n到9月的1.3版转向打磨长期任务的可用性\n几乎保持每个月一次大版本更新的速度\n这种快速迭代在AI圈不算稀奇\n但难得的是Meta几乎每一版都会附上明确的内部对比数据\n同时引用第三方评测结果\n给了外界一个相对透明的观察窗口\n很多人只盯着分数涨了多少\n却忽略了每一代版本的定位差异\nSpark 1.3并不是全面碾压1.2的终极进化\n而是针对代理人任务的可靠性做了针对性强化\n甚至在部分涨上下文指标上还出现了小幅退步\n这种有得有失的升级\n恰恰说明模型能力不是简单的线性叠加\n而是一连串的取舍与权衡\n如果你只看营销文案里的亮点\n很快就会在实际使用中发现落差\n如果你翻完这次的全部技术文件\n会发现最核心的关键词\n不是更高的分数\n也不是更多的参数\n而是长任务代理人\n什么意思\n就是让AI不再只是一次性回答一个问题\n而是像一个真正的执行助理一样\n接下一个模糊的需求\n自己拆解步骤 呼叫工具\n遇到问题回头问你\n碰到危险操作\n先停下来确认最后再交付完整的成果\n这听起来很美好\n但过去大部分代理人模型的问题无非三种\n要么不敢动手只会空泛回答\n要么乱动手做到一半偏离目标\n更常见的是干到一半彻底忘了最原始的要求\nSpark 1.3就是冲着这些痛点做的优化\n举个很实际的场景\n如果你跟模型说\n帮我整理上礼拜的三场客户会议记录\n做成部门简报\n把每个客户的行动项列出来\n最后寄给主管\n换做之前的模型\n很可能做到一半就忘了要按客户分类\n或者漏了寄信的步骤\n甚至为了凑数编造几个不存在的行动项\n而Spark 1.3的处理逻辑会完全不同\n一开始如果觉得需求不够清楚\n它会主动追问你\n简报要的是正式版本还是讨论版\n行动项要不要加上预计完成时间\n寄信的时候要不要附上原始记录档\n执行过程中如果发现某份会议记录打不开或者权限不足\n它会停下来跟你要资讯\n而不是硬着头皮编一个结果\n如果碰到删除档案\n发送批量邮件\n修改正式环境设定这类有不可逆风险的操作\n它一定会先跳出确认\n问你确定要执行吗\n不会自己直接动手\n更重要的是\n它能分辨你现在是在补充当前任务的资讯\n还是要打断手头的事做别的\n又或者是要开启一个全新的任务\n这听起来好像是基础能力,但在多轮对话里,很多模型会把用户随口的插画当成新指令,直接把之前的上下文全部覆盖掉,导致任务做到一半就断了。\n针对长期任务最常见的约束遗忘问题,Spark 1.3也做了专门强化。\n比如你一开始就说过\n所有输出都要用正体中文\n不能用外部图片素材\n数据不能超过指定范围\n他会在整个执行过程中\n反复检查这些条件\n而不是跑个五六步之后就全忘了\n最关键的一点\n是他对自身能力边界的判断\n过去很多代理人模型\n最危险的习惯\n就是任务做不完的时候\n会假装做完了\n然后生成一份\n看起来很完整的报告\n里面全是编造的数据和结论\n普通用户根本看不出来\nBug1.3在训练时特别强化了这点遇到做不到的地方\n他更倾向于直接说\n我卡在这里了\n我缺少这个资讯\n而不是硬凑出一份看似完美的答案\nMeta内部工程师对比1.2版本后发现\n新版本平均减少了20%的工具呼叫次数\n同时减少了25%的token消耗\n这两个数字其实比跑分更有意义\n工具呼叫减少代表模型更清楚\n什么时候该用什么工具\n不会无脑尝试所有选项来回绕路\nToken消耗减少\n代表推理效率提升\n单位任务的成本直接下降\n很多人不知道代理人任务的成本\n通常是纯问答的3到5倍\n因为要不断多轮推理\n呼叫工具修正路线\n所以效率的提升\n直接决定了这项技术\n能不能大规模商业化\n不过也要说清楚\n这两个数字是Meta内部测试场景下的结果\n不是第三方独立评测\n实际业务中因为场景更复杂\n不一定能复现同样的比例\n官方也举了几个典型的常任务场景\n官方也举了几个典型的常任务场景,比如工程领域的CAD模型搭配流体模拟数据。最后按指定章节输出PDF报告,音乐制作领域根据时间码定位错误音符,从重复段落里挑出合适的替代片段。\n保持所有音轨长度不变\n再重新混音\n这些场景的共同点就是步骤多\n中间产物多\n需要跨工具协作\n差一步整个结果就不能用\n从这个角度看\nSpark 1.3根本不是一个聊天更厉害的模型\n而是一个干活更稳的模型\n它的评价标准早就不是会不会写诗\n能不能解数学题\n而是能不能在100步之后\n还不偏离最初的目标\n聊完了官方自己讲的升级\n我们来看看第三方独立评测的数据\n验证一下这些能力到底有没有落到实处\n这次的数据全部来自Artificial Analysis\n这是目前业界专门做代理人模型评测的独立机构\n不是Meta自己跑的分\n参考价值相对高一些\n在他们的综合智慧指数榜单上\nSpark 1.3的Shy版本拿到61分\n还在有限预览阶段的Max版本拿到62分\n比起1.2版本的57分\nExia整整涨了4分\n在AI评测里 4分的涨幅算不算大 要看在哪个区间\n如果从30分涨到34分 那只是入门级的进步\n但从57分涨到61分已经是第一梯队模型的明显提升\n因为越往上 每一分的难度都越高\n更有竞争力的其实是成本XJ版本完成一项综合指数任务\n平均花费0.55美元 在所有得分超过59分的模型里是最低的\n我们可以简单对比一下\n同级别里\nOpenAI的GPT-4O Advanced\n做类似任务\n大概要1.2美元\nAnthropic的Cloud Opus\n则要1.5美元左右\nMeta差不多只要一半的价格\n这对于需要大量呼叫API的企业用户来说\n吸引力非常大\n我们再拆开看具体任务的表现\nTile3-Bench Banking\n是专门的银行代理人基准测试\n模拟金融场景下的多步骤任务\n1.2版本\n当时只拿到35%的通过率\nSHAE版本提升到47%\nMAX版本更是达到52%\n这个涨幅其实相当惊人\n说明模型在处理合规\n要求高步骤繁琐的金融任务时\n可靠性比上一代强了很多\n但也要客观说\n47%的通过率\n依然代表超过一半的任务\n没办法完美完成\n距离真正取代人类专员\n还有非常远的距离\n大概相当于入行半年的助理水平\nTerminal-Bench 2.1是终端指令执行测试\n1.2版本通过率80%\nX-Shy提升到85%\nMax到86%\n这个分数看起来很高\n但终端任务相对结构化\n指令对错的标准很明确\n能拿到高分并不意外\n参考价值不如真实工作场景\n真正值得注意的反而是一项退步的指标\nAA-LCR长上下文基准\n1.2版本有83%的正确率\n1.3版本降到了79%\n为什么长上下文能力反而退步了\n难道模型越更新越差吗\n当然不是\n背后的原因是Spark 1.3调整了策略\n遇到不确定的资讯\n它更倾向于拒绝回答\n而不是硬猜一个答案\n也就是说它宁可少说几句\n也不说错话\n这种策略在AA-Omniscience准确率指标上\n也能看出来\n表面上准确率小幅回落\n但幻觉率明显降低了\n这其实是一个非常重要的取舍\n你要更高的召回率\n还是更高的准确率\n对于普通用户来说\n可能希望AI什么都能答\n哪怕错一点也没关系\n但对于企业用户\n尤其是金融 医疗 法律这些领域\n一个错误的答案\n代价可能比没有答案大得多\n之前就有企业用初代代理人模型\n处理报销审核\n结果模型编造了很多不存在的单据\n造成了不小的财务漏洞\n所以现在越是面向企业的模型\n越把不编造放在比达对更高的优先级\n从这个角度看\n这项退步反而是一种进步\n代表模型开始往实用化的方向走\n而不是为了跑分应凑答案\n另外很多人忽略的一点是\nMAX版本的分数虽然更高\n但代价是运算算力大幅增加\n根据Artificial Analysis的测试\n在GDPVAL-AAVFE任务里\nMAX版本比SY多消耗约62%的推理量\n就算是银行代理任务\n也多了28%的算力\n为什么会差这么多\n因为MAX版本会做更多的自我检查\n走更多的思维步骤\n相当于人做事情的时候反复检查好几遍\n自然更费时间更费算力\n但正确率也更高\n这其实带出了未来AI商业模式的一个可能方向\n以后的模型不会只有单一档位\n而是会根据任务的风险等级调整推理强度\n低风险的闲聊\n查资料用低档速度快又便宜\n高风险的财务\n法律任务用高档\n更慢 更贵 但更可靠\n用户可以按自己的需求选\n而不是为了偶尔的高需求\n全程付最贵的价格\n当然 目前这还只是观察\nMeta并没有正式推出这种分级定价\n但从技术设计上\n已经具备了这样的基础\n最后 还要再强调一次\nMax版本目前只是有限预览\n只有少数合作机构和企业能拿到权限\n普通开发者根本呼叫不到\n所以你看到62分那个数字的时候\n要知道它背后是更高的成本\n更严格的访问限制\n不能拿它去跟别人全面开放的API比\n然后说Meta已经赢了\n这种对比并不公平\n看完了负责思考和执行的大脑\n我们再转过头来讲这次连发的另一款产品\n负责感知输入的耳朵\nMuse Voice Transcribe\n很多人第一眼看到它\n会觉得不就是个语音转文字吗\n有什么特别的\n但实际上它的设计思路\n跟传统的自动语音识别完全不一样\n传统的ASR比如Whisper\n走的是离线转写路线\n拿到整段音频之后再慢慢处理\n可以利用上下文反复修正\n准确率很高\n但延迟也大\n根本没办法用来做实时对话\n而以前的实时语音助手\n比如手机上的Siri Google助理\n为了追求低延迟\n只能听一小段就马上输出\n结果准确率很差\n稍微吵一点\n有两个人叠话或者混个外文单字\n结果就歪掉\nMuse Voice Transcribe想解决的\n就是又要准又要快\n还要懂真实对话这个老难题\n它的做法是把音频切成80毫秒的小块\n每秒处理12月5日次\n每一小块音频先转换成一个软Token\n模型在判断是要继续等后面的声音\n还是立刻输出文字\n等到音频完全停止\n系统会加入一个特殊标记\n让模型把剩下没输出的文字一口气吐出来\n其中最关键的技术叫做自适应延迟\n模型不会给每个字都设定一样的等待时间\n简单确定的内容就立刻输出\n碰到需要后文才能判断的同音词\n专有名词\n就稍微多等一点点音频再决定\n这其实很像人类的同步口译\n译员会根据句子的难度\n调整自己的滞后时间\n不会一味求快\n也不会一味求准\n训练的时候\nMeta把词 错误率和延迟时间\n一起放进强化学习的奖励函数里\n如果模型只追求准确\n每句话都等很久\n得分不会高\n只追求速度\n输出一大堆错字同样过不了关\n最终目标就是在准确率和反应速度之间找到最佳平衡点\n那实际表现如何\n在Artificial Analysis的流逝语音榜单上\nMuse Voice Transcribe的最终词错误率约3.1%\n直接登上第一名\n这个数字是什么概念?在流逝实时的场景下,3.1%已经非常接近人类专业数据员的水平,人类大概是2%到3%左右,相当于每讲100个字只错3个。\n不过这里必须先打个预防针\n这个成绩是英语流逝对话的测试结果\n绝对不能直接套用到所有语言 所有场景\n中文因为同音字多 声调复杂\n同样的模型词错误率通常会比英语高2到3个百分点\n如果是嘈杂环境 多人强化 或者带方言口音\n表现还会再打折扣\n官方也明确说 模型训练覆盖了70多种语言\n其中只有25种经过充分验证\n其他语言可能只是有训练过\n但没有经过严格的效果评测\n除了基础的转写准确率\n还有几个功能特别值得一提\n首先是原生的说话人分离\n而且支持20人以上同时发言\n传统的会议转写通常是先做语音识别\n再另外用一套声文系统\n判断每段话是谁说的\n两个系统的时间轴\n只要差一点点就会出现文字\n人认错的情况\nMeta是在同一个生成过程里\n加入说话人切换标记和身份标签\n检测到换人就先标一轮新发言\n等听到足够声音再确定身份\n时间轴的准确度会高很多\n官方展示的长对话里包含多人插话\n重叠发言 自然停顿\n都能正确标记身份\n这对于会议记录 访谈整理 客服记录来说\n价值很高\n因为后续的摘要系统拿到的不只是一整段文字\n还知道谁提了问题 谁做了承诺\n谁反对哪个方案\n其次是端点侦测能力\n人类对话本来就不会每次讲完都按个结束钮\n系统要从语气停顿和上下文判断\n对方只是喘口气还是真的说完了\n如果端点判断不准\n要么用户都讲完半天了\n助手还没反应\n要么用户还没说完\n助手就急着插嘴体验都很差\nMuse Voice Transcribe\n专门针对这点做了优化\n能更准确判断语音的开始和结束\n这也是代理人能自然对话的前提\n第三个很实用的功能是\n句内和句间的多语言自由切换\n还有上下文偏制\n现实中很多双语使用者\n不会先切语言模式再讲话\n中文句子里夹个英文专有名词\n项目代号是家常便饭\n旧的系统很容易把英文词\n应转成发音相近的中文字\n或者切换语言之后直接丢了半句话\nMuse Voice Transcribe\n支持70多种语言的自由切换\n还可以让用户提前输入常用的人名\n公司名专业术语\n模型碰到相近发音的时候\n会优先选择这些符合上下文的词\n比如医生可以提前录入药品名称\n城市设计师可以加入函数名和项目代号\n企业用户可以上传内部数语表\n这样转写的准确率会大幅提升\n甚至可以直接进入工作流程\n不用再人工校对一遍\n这才是语音转写真正的价值\n不是把声音变成文字就好\n而是变成能直接用的文字\n目前Muse Voice Transcribe\n已经接入Meta Model API\n还有Meta AI4 Mac Muse Code Mac版本\n用户可以按住FN键直接语音输入\n把转写结果送到不同的应用程式里\n不过有个隐私细节一定要讲清楚\n官方网页演示的时候说麦克风音频只会用来生成转写\n不会被保存\n但这个承诺只限于网页演示场景\n不等于所有API呼叫\n所有产品部署都不会保存音频\n如果你是透过API上传录音\nMeta会不会保存\n保存多久\n会不会用来训练模型\n都要看你签署的服务条款和数据处理协议\n不能因为网页演示不保存\n就理所当然认为所有场景都不保存\n这在个资合规上是非常危险的认知\n讲到这里就必须好好聊一下\n隐私与安全的边界\n这也是很多人看AI产品的时候\n最容易忽略但绝对最重要的部分\n先讲Muse Spark 1.3的安全设计\n官方说这次提升了对提示注入\n和对抗输入的抵抗能力\n也更能识别哪些操作是不可逆的\n提示注入是什么\n就是攻击者在输入内容里\n嵌入看不见的恶意指令\n比如藏在PDF档案里\n网页里\n诱导AI代理人忽略原本的规则\n去转发机密档案\n修改设定甚至转账\n2025年就曾经出现过\n类似的攻击案例\n让不少企业对AI代理人望而却步\nSpark 1.3针对这类攻击做了专门训练\n但这不代表它完全免疫\n目前没有任何模型能100%防御所有提示注入攻击\n识别不可逆操作也是一样\n模型能标出删除档案\n发送批量邮件\n修改数据库这些高风险动作\n要求用户确认\n但光有模型层面的防护远远不够\n官方自己也强调实际部署的时候\n必须同时做好全线隔离\n白名单和操作日志\n全线隔离就是给代理人最小必要的全线\n比如负责整理表格的助理\n就不该有删除整个资料夹的权限\n白名单就是只允许模型呼叫预先核准的\n工具和API不能让它随便访问系统里的任何东西\n操作日志就是把代理人所有执行动作都记录下来\n方便事后审计和回溯\n这三层防护都是资安领域的基本常识\n却也是很多AI产品最容易省略的部分\nMeta愿意主动强调这点\n说明它是真的在考虑企业落地的实际需求\n而不是只做展示用的技术demo\n再回到语音部分\n当语音转写加上代理人执行\n再搭配Meta旗下的ray减班智慧眼镜\n想象空间确实很大\n手机上的语音助手要等你点开APP才能用\n但眼镜可以一直戴在脸上\n随时处在声音环境里\n如果他能听懂谁在讲话\n什么时候该接话\n什么时候不该插嘴\n就能在会议后自动整理每个人的行动项\n访谈的时候自动标记重点\n甚至日常聊天里提到的代办事项都能自动记下来\n但这也带来了更尖锐的隐私问题\n眼镜上的麦克风什么时候在听\n听到的内容传到哪里去了\n用户有没有明确的控制权\n旁边的人知不知道自己正在被录音\n这些问题目前都还没有完全公开的答案\n不过我们可以从Meta的双轨策略里推出一个很合理的假设\n未来轻量级的语音转写\n简单的指令理解\n很可能会直接跑在眼镜本地的端侧芯片上\n不需要传到云端\n敏感数据不会离开设备\n大幅降低隐私风险\n而复杂的常任务代理\n多工具协作\n再交给云端的Spark模型处理\n刚好Meta有Glimmer这条本地开源路线\n专门做轻量化端侧部署\n刚好可以跟眼镜这类终端装置搭配\n云端负责复杂思考\n本地负责实时感知\n既保证体验又兼顾隐私\n这很可能就是Meta下一代AI入口的完整架构\n当然这只是基于现有产品线的推测\n官方还没有证实这个方向\n讲完了技术亮点和想象空间\n我们也得冷静下来\n聊聊这次发布的口径边界和实际风险\n首先Max版本目前只是有线预览\n普通开发者能用的只有Shi版本\n网上所有关于Max的评测数据\n都来自少数拿到权限的机构\n一般用户现在根本用不到\n所以62分那个数字只能当成技术上限的参考\n不能作为采购或者使用决策的依据\n其次所有的基准测试数据\n都是在标准化的干净环境里跑出来的\n真实世界的场景只会更复杂更混乱\n比如银行代理测试里的数据都是结构化的\n格式统一全线清晰\n但真实的企业里,数据可能分散在十几个老旧系统里,格式乱七八糟、权限层层卡关。\n你在测试里拿到47%的通过率,到了真实环境可能连30%都不到。\n这不是Meta的问题 是所有AI模型评测的普遍现象 跑分永远不等于真实生产力\n第三 关于开源的期待也要放低\nSpark 1.3目前是完全避原的 官方只说未来有开放权重的规划 但没有时间表 也没有说授权条款\n参考Meta之前Lama系列的节奏 通常是新一代出来之后才会把上一代的权重逐步开放\n而且早期都是非商业授权 要等到更后面才会用更宽松的授权\n所以不要期待短期内就能在自己的电脑上\n跑Spark 1.3核心的代理人能力\nMeta肯定会先握在手里当商业护城河\n最后回到产业竞争的层面\nMeta这次的更新确实扎实\n在代理人可靠性和流逝语音转写上都往前走了一步\n成本控制也做得很有竞争力\n但要说已经赢了还太早\nOpenEye的GPT系列代理人能力同样在快速迭代\n谷歌的Gemini在多模态和端侧整合上有自己的优势\nAnthropic在长上下文和安全合规上依然领先\n还有Eleven Labs这类专业语音公司也在往代理人方向走\nMeta现在的领先更像是某几个细分场景的小幅度领先\n还远没有到拉开差距的程度\n而且真正的竞争从来都不是比模型本身\n而是比谁能先把技术变成用户愿意天天用的产品\n谁能解决好隐私安全\n权限这些麻烦的落地问题\n模型再聪明\n如果用户不敢把重要的工作交给他\n那也只是个高级玩具\n讲到这里\n我们把这次Meta两款产品的技术细节\n评测数据\n产品线定位\n隐私风险\n还有产业格局全部拆完了\n你会发现每一个亮眼的数据背后\n都有一个限定条件\n每一项技术突破的旁边\n都站着一个带解决的问题\n这不是说Meta做得不好\n恰恰相反\n只有当你看清了所有的边界和限制\n才能真正评估它的价值\nMuse Spark 1.3确实把常任务代理人的可靠性往前推了一步\n成本优势也很明显\nMuse Voice Transcribe在英语流逝转写上\n做到了业界顶尖\n准确率和延迟的平衡做得可圈可点\n但它们都还只是阶段性的进展\n离大家想象中什么都能做的AI助理\n还有非常远的距离\n2026年的AI产业\n其实正在经历一个很重要的转向\n前几年大家都在比谁的模型更大\n谁的跑分更高\n谁的牛皮吹得更响\n但从今年开始竞争慢慢落到了实处\n比的是谁能把长任务做稳\n谁能把真实声音听懂\n谁能让企业敢放心用\n谁能把隐私控制在合理范围\n那些看起来不怎么耀眼的细节优化\n反而决定了产品能不能真正走进日常生活和工作流\n就像造车\n早期大家都比马力 比急速 等到市场成熟了 比的是隔音 是油耗 是故障率 是乘坐舒适度\n这些才是决定你愿不愿意每天开的关\n见如果你喜欢这种不吹不黑 把数据和限制都讲清楚的深度科技解析\n欢迎订阅这个频道 并且打开通知小铃铛\n我们会持续追踪AI代理人 语音交互 端策AI这些领域的最新进展\n帮你过滤掉夸大的营销话术\n留下真正值得关注的技术变化\n也欢迎在留言区告诉我\n你最期待AI帮你完成什么样的任务\n是整理会议记录\n还是处理繁琐的文书工作\n我们下期影片再见",
  "language": "zh",
  "duration": 1629.855063,
  "segments": [
    {
      "start": 0.0,
      "end": 2.3,
      "text": "你有没有过这种很无力的经验?",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09952453702513935,
      "no_speech_prob": 9.36235377418937e-12,
      "compression_ratio": 1.2661064425770308
    },
    {
      "start": 2.3,
      "end": 5.0,
      "text": "叫AI帮你整理一份跨部门的会议报告。",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09952453702513935,
      "no_speech_prob": 9.36235377418937e-12,
      "compression_ratio": 1.2661064425770308
    },
    {
      "start": 5.0,
      "end": 9.3,
      "text": "做到一半它忘了你要求的章节格式,用语音助手查形成。",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09952453702513935,
      "no_speech_prob": 9.36235377418937e-12,
      "compression_ratio": 1.2661064425770308
    },
    {
      "start": 9.3,
      "end": 13.7,
      "text": "旁边同事插句话,它就把指令听成完全不相干的内容。",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09952453702513935,
      "no_speech_prob": 9.36235377418937e-12,
      "compression_ratio": 1.2661064425770308
    },
    {
      "start": 13.7,
      "end": 20.5,
      "text": "只要任务超过三个步骤,需要跨两个以上工具,最后交付的结果一定歪掉大半。",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09952453702513935,
      "no_speech_prob": 9.36235377418937e-12,
      "compression_ratio": 1.2661064425770308
    },
    {
      "start": 20.5,
      "end": 27.5,
      "text": "你大概也听过很多人说AI已经很聪明了,但真的拿来干活,总是卡在这些最不起眼的小地方。",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09952453702513935,
      "no_speech_prob": 9.36235377418937e-12,
      "compression_ratio": 1.2661064425770308
    },
    {
      "start": 27.5,
      "end": 29.98,
      "text": "2026年9月的头两天",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.054914754979750686,
      "no_speech_prob": 1.8702352166943825e-11,
      "compression_ratio": 1.2761627906976745
    },
    {
      "start": 29.98,
      "end": 32.4,
      "text": "Meta接连丢出了两项更新",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.054914754979750686,
      "no_speech_prob": 1.8702352166943825e-11,
      "compression_ratio": 1.2761627906976745
    },
    {
      "start": 32.4,
      "end": 34.84,
      "text": "没有吹什么通用人工智能的牛皮",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.054914754979750686,
      "no_speech_prob": 1.8702352166943825e-11,
      "compression_ratio": 1.2761627906976745
    },
    {
      "start": 34.84,
      "end": 38.48,
      "text": "却精准踩中了这些大家天天吐槽的痛点",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.054914754979750686,
      "no_speech_prob": 1.8702352166943825e-11,
      "compression_ratio": 1.2761627906976745
    },
    {
      "start": 38.48,
      "end": 41.76,
      "text": "一个负责把真实世界里的声音听得更清楚",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.054914754979750686,
      "no_speech_prob": 1.8702352166943825e-11,
      "compression_ratio": 1.2761627906976745
    },
    {
      "start": 41.76,
      "end": 44.78,
      "text": "一个负责把复杂的常任务做得更稳",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.054914754979750686,
      "no_speech_prob": 1.8702352166943825e-11,
      "compression_ratio": 1.2761627906976745
    },
    {
      "start": 44.78,
      "end": 48.72,
      "text": "有人看完数据说Meta终于要在代理人战场超车了",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.054914754979750686,
      "no_speech_prob": 1.8702352166943825e-11,
      "compression_ratio": 1.2761627906976745
    },
    {
      "start": 48.72,
      "end": 50.42,
      "text": "也有人冷静算完成本",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.054914754979750686,
      "no_speech_prob": 1.8702352166943825e-11,
      "compression_ratio": 1.2761627906976745
    },
    {
      "start": 50.42,
      "end": 52.54,
      "text": "发现事情根本没有这么简单",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.054914754979750686,
      "no_speech_prob": 1.8702352166943825e-11,
      "compression_ratio": 1.2761627906976745
    },
    {
      "start": 52.54,
      "end": 54.18,
      "text": "今天我们就不吹不黑",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.054914754979750686,
      "no_speech_prob": 1.8702352166943825e-11,
      "compression_ratio": 1.2761627906976745
    },
    {
      "start": 54.18,
      "end": 57.22,
      "text": "把官方技术文件第三方独立评测",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.054914754979750686,
      "no_speech_prob": 1.8702352166943825e-11,
      "compression_ratio": 1.2761627906976745
    },
    {
      "start": 57.22,
      "end": 60.14,
      "text": "还有那些藏在角落的小字限制",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07029183287369578,
      "no_speech_prob": 1.4124537477122612e-11,
      "compression_ratio": 1.1810089020771513
    },
    {
      "start": 60.14,
      "end": 61.84,
      "text": "全部摊开来讲",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07029183287369578,
      "no_speech_prob": 1.4124537477122612e-11,
      "compression_ratio": 1.1810089020771513
    },
    {
      "start": 61.84,
      "end": 64.62,
      "text": "看看这两款产品到底走到了哪一步",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07029183287369578,
      "no_speech_prob": 1.4124537477122612e-11,
      "compression_ratio": 1.1810089020771513
    },
    {
      "start": 64.62,
      "end": 67.3,
      "text": "又有哪些东西是营销不会告诉你的",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07029183287369578,
      "no_speech_prob": 1.4124537477122612e-11,
      "compression_ratio": 1.1810089020771513
    },
    {
      "start": 67.3,
      "end": 68.86,
      "text": "我们不从结论讲起",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07029183287369578,
      "no_speech_prob": 1.4124537477122612e-11,
      "compression_ratio": 1.1810089020771513
    },
    {
      "start": 68.86,
      "end": 72.28,
      "text": "因为真正有价值的判断往往都藏在每一个",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07029183287369578,
      "no_speech_prob": 1.4124537477122612e-11,
      "compression_ratio": 1.1810089020771513
    },
    {
      "start": 72.28,
      "end": 73.38,
      "text": "但是后面",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07029183287369578,
      "no_speech_prob": 1.4124537477122612e-11,
      "compression_ratio": 1.1810089020771513
    },
    {
      "start": 73.38,
      "end": 77.3,
      "text": "先帮大家把最容易搞混的产品线界限画清楚",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07029183287369578,
      "no_speech_prob": 1.4124537477122612e-11,
      "compression_ratio": 1.1810089020771513
    },
    {
      "start": 77.3,
      "end": 80.1,
      "text": "这也是很多自媒体最常讲错的地方",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07029183287369578,
      "no_speech_prob": 1.4124537477122612e-11,
      "compression_ratio": 1.1810089020771513
    },
    {
      "start": 80.1,
      "end": 84.16,
      "text": "9月1日登场的是Muse Voice Transcribe语音转写模型",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07029183287369578,
      "no_speech_prob": 1.4124537477122612e-11,
      "compression_ratio": 1.1810089020771513
    },
    {
      "start": 84.16,
      "end": 88.16,
      "text": "9月2日接棒上线的是Muse Spark 1.3代理人模型",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08180356621742249,
      "no_speech_prob": 1.2959733980411059e-11,
      "compression_ratio": 1.1975308641975309
    },
    {
      "start": 88.16,
      "end": 90.56,
      "text": "两天连续发布核心产品",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08180356621742249,
      "no_speech_prob": 1.2959733980411059e-11,
      "compression_ratio": 1.1975308641975309
    },
    {
      "start": 90.56,
      "end": 94.34,
      "text": "这在Meta Superintelligence Labs的历史上并不常见",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08180356621742249,
      "no_speech_prob": 1.2959733980411059e-11,
      "compression_ratio": 1.1975308641975309
    },
    {
      "start": 94.34,
      "end": 98.52,
      "text": "也能看出他们在2026年下半年的竞争焦虑",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08180356621742249,
      "no_speech_prob": 1.2959733980411059e-11,
      "compression_ratio": 1.1975308641975309
    },
    {
      "start": 98.52,
      "end": 99.5,
      "text": "但更重要的是",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08180356621742249,
      "no_speech_prob": 1.2959733980411059e-11,
      "compression_ratio": 1.1975308641975309
    },
    {
      "start": 99.5,
      "end": 101.78,
      "text": "很多人到现在还搞不清楚",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08180356621742249,
      "no_speech_prob": 1.2959733980411059e-11,
      "compression_ratio": 1.1975308641975309
    },
    {
      "start": 101.78,
      "end": 103.96,
      "text": "Muse Spark和Muse Glimmer",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08180356621742249,
      "no_speech_prob": 1.2959733980411059e-11,
      "compression_ratio": 1.1975308641975309
    },
    {
      "start": 103.96,
      "end": 106.86,
      "text": "根本是两条完全不同的路线",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08180356621742249,
      "no_speech_prob": 1.2959733980411059e-11,
      "compression_ratio": 1.1975308641975309
    },
    {
      "start": 106.86,
      "end": 108.18,
      "text": "不能混为一谈",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08180356621742249,
      "no_speech_prob": 1.2959733980411059e-11,
      "compression_ratio": 1.1975308641975309
    },
    {
      "start": 108.18,
      "end": 111.28,
      "text": "Muse Spark走的是云端前沿代理人路线",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08180356621742249,
      "no_speech_prob": 1.2959733980411059e-11,
      "compression_ratio": 1.1975308641975309
    },
    {
      "start": 111.28,
      "end": 112.72,
      "text": "权重完全必圆",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08180356621742249,
      "no_speech_prob": 1.2959733980411059e-11,
      "compression_ratio": 1.1975308641975309
    },
    {
      "start": 112.72,
      "end": 114.82,
      "text": "目前只开放API呼叫",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057756668610098476,
      "no_speech_prob": 1.2463792151140574e-11,
      "compression_ratio": 1.2082111436950147
    },
    {
      "start": 114.82,
      "end": 117.56,
      "text": "官方只说未来有开放权重的规划",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057756668610098476,
      "no_speech_prob": 1.2463792151140574e-11,
      "compression_ratio": 1.2082111436950147
    },
    {
      "start": 117.56,
      "end": 119.0,
      "text": "但既没有时间表",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057756668610098476,
      "no_speech_prob": 1.2463792151140574e-11,
      "compression_ratio": 1.2082111436950147
    },
    {
      "start": 119.0,
      "end": 121.28,
      "text": "也没说会用哪种授权条款",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057756668610098476,
      "no_speech_prob": 1.2463792151140574e-11,
      "compression_ratio": 1.2082111436950147
    },
    {
      "start": 121.28,
      "end": 123.42,
      "text": "所以你绝对不能把它理解成",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057756668610098476,
      "no_speech_prob": 1.2463792151140574e-11,
      "compression_ratio": 1.2082111436950147
    },
    {
      "start": 123.42,
      "end": 125.7,
      "text": "马上就可以下载来本地跑",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057756668610098476,
      "no_speech_prob": 1.2463792151140574e-11,
      "compression_ratio": 1.2082111436950147
    },
    {
      "start": 125.7,
      "end": 128.54,
      "text": "而Muse Glimmer是拥有300亿参数",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057756668610098476,
      "no_speech_prob": 1.2463792151140574e-11,
      "compression_ratio": 1.2082111436950147
    },
    {
      "start": 128.54,
      "end": 131.86,
      "text": "采用Apache 2.0授权的本地方案",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057756668610098476,
      "no_speech_prob": 1.2463792151140574e-11,
      "compression_ratio": 1.2082111436950147
    },
    {
      "start": 131.86,
      "end": 134.14,
      "text": "主打消费级显示卡就能执行",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057756668610098476,
      "no_speech_prob": 1.2463792151140574e-11,
      "compression_ratio": 1.2082111436950147
    },
    {
      "start": 134.14,
      "end": 134.78,
      "text": "讲白一点",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057756668610098476,
      "no_speech_prob": 1.2463792151140574e-11,
      "compression_ratio": 1.2082111436950147
    },
    {
      "start": 134.78,
      "end": 137.1,
      "text": "Spark负责把能力堆到最高",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057756668610098476,
      "no_speech_prob": 1.2463792151140574e-11,
      "compression_ratio": 1.2082111436950147
    },
    {
      "start": 137.1,
      "end": 139.68,
      "text": "强企业级和高端开发者市场",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057756668610098476,
      "no_speech_prob": 1.2463792151140574e-11,
      "compression_ratio": 1.2082111436950147
    },
    {
      "start": 139.68,
      "end": 141.84,
      "text": "Glimmer负责把能力扑到最广",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057756668610098476,
      "no_speech_prob": 1.2463792151140574e-11,
      "compression_ratio": 1.2082111436950147
    },
    {
      "start": 141.84,
      "end": 144.1,
      "text": "占领开源生态和终端装置",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07546072901941794,
      "no_speech_prob": 1.4404363118947217e-11,
      "compression_ratio": 1.1702127659574468
    },
    {
      "start": 144.1,
      "end": 148.68,
      "text": "这种双轨策略其实刚好对应了Meta在AI产业的定位转变",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07546072901941794,
      "no_speech_prob": 1.4404363118947217e-11,
      "compression_ratio": 1.1702127659574468
    },
    {
      "start": 148.68,
      "end": 153.4,
      "text": "前几年大家提到Meta的AI第一个想到的就是Lama系列开源模型",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07546072901941794,
      "no_speech_prob": 1.4404363118947217e-11,
      "compression_ratio": 1.1702127659574468
    },
    {
      "start": 153.4,
      "end": 154.96,
      "text": "靠免费打天下",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07546072901941794,
      "no_speech_prob": 1.4404363118947217e-11,
      "compression_ratio": 1.1702127659574468
    },
    {
      "start": 154.96,
      "end": 157.98,
      "text": "跟OpenAI的全B原路线正面抗衡",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07546072901941794,
      "no_speech_prob": 1.4404363118947217e-11,
      "compression_ratio": 1.1702127659574468
    },
    {
      "start": 157.98,
      "end": 160.06,
      "text": "但到了2026年",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07546072901941794,
      "no_speech_prob": 1.4404363118947217e-11,
      "compression_ratio": 1.1702127659574468
    },
    {
      "start": 160.06,
      "end": 163.76,
      "text": "整个行业的竞争已经从比参数 比跑分",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07546072901941794,
      "no_speech_prob": 1.4404363118947217e-11,
      "compression_ratio": 1.1702127659574468
    },
    {
      "start": 163.76,
      "end": 166.9,
      "text": "进到比代理人 比落地场景的阶段",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07546072901941794,
      "no_speech_prob": 1.4404363118947217e-11,
      "compression_ratio": 1.1702127659574468
    },
    {
      "start": 166.9,
      "end": 170.14,
      "text": "Meta必须拿出足够强的B原顶端产品",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07546072901941794,
      "no_speech_prob": 1.4404363118947217e-11,
      "compression_ratio": 1.1702127659574468
    },
    {
      "start": 170.14,
      "end": 171.6,
      "text": "才能跟GPT系列",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06806739171346028,
      "no_speech_prob": 1.3743662458098882e-11,
      "compression_ratio": 1.1840490797546013
    },
    {
      "start": 171.6,
      "end": 173.7,
      "text": "Gemini系列抢商业客户",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06806739171346028,
      "no_speech_prob": 1.3743662458098882e-11,
      "compression_ratio": 1.1840490797546013
    },
    {
      "start": 173.7,
      "end": 176.76,
      "text": "同时又不能丢掉开源生态的优势",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06806739171346028,
      "no_speech_prob": 1.3743662458098882e-11,
      "compression_ratio": 1.1840490797546013
    },
    {
      "start": 176.76,
      "end": 179.24,
      "text": "为什么要花这么多时间讲这件事",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06806739171346028,
      "no_speech_prob": 1.3743662458098882e-11,
      "compression_ratio": 1.1840490797546013
    },
    {
      "start": 179.24,
      "end": 180.54,
      "text": "因为如果你误以为",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06806739171346028,
      "no_speech_prob": 1.3743662458098882e-11,
      "compression_ratio": 1.1840490797546013
    },
    {
      "start": 180.54,
      "end": 182.58,
      "text": "Spark 1.3是开源模型",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06806739171346028,
      "no_speech_prob": 1.3743662458098882e-11,
      "compression_ratio": 1.1840490797546013
    },
    {
      "start": 182.58,
      "end": 184.66,
      "text": "就会产生错误的部署预期",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06806739171346028,
      "no_speech_prob": 1.3743662458098882e-11,
      "compression_ratio": 1.1840490797546013
    },
    {
      "start": 184.66,
      "end": 188.22,
      "text": "如果你以为Glimmer只是Spark的缩小版",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06806739171346028,
      "no_speech_prob": 1.3743662458098882e-11,
      "compression_ratio": 1.1840490797546013
    },
    {
      "start": 188.22,
      "end": 191.7,
      "text": "就会低估两者在设计目标上的根本差异",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06806739171346028,
      "no_speech_prob": 1.3743662458098882e-11,
      "compression_ratio": 1.1840490797546013
    },
    {
      "start": 191.7,
      "end": 193.16,
      "text": "另外从迭代节奏来看",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06806739171346028,
      "no_speech_prob": 1.3743662458098882e-11,
      "compression_ratio": 1.1840490797546013
    },
    {
      "start": 193.16,
      "end": 195.5,
      "text": "Muse系列从7月的1.1版",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06806739171346028,
      "no_speech_prob": 1.3743662458098882e-11,
      "compression_ratio": 1.1840490797546013
    },
    {
      "start": 195.5,
      "end": 197.36,
      "text": "首次连上邮件行事例",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06806739171346028,
      "no_speech_prob": 1.3743662458098882e-11,
      "compression_ratio": 1.1840490797546013
    },
    {
      "start": 197.36,
      "end": 203.16,
      "text": "并能制作投影片到8月初1.2版聚焦大型程式库与复杂真错",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04671802160874853,
      "no_speech_prob": 1.208129776775202e-11,
      "compression_ratio": 1.167808219178082
    },
    {
      "start": 203.16,
      "end": 205.7,
      "text": "甚至能连续1000多次工具呼叫",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04671802160874853,
      "no_speech_prob": 1.208129776775202e-11,
      "compression_ratio": 1.167808219178082
    },
    {
      "start": 205.7,
      "end": 210.26,
      "text": "在24小时内反复编写编译GPU核心程式",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04671802160874853,
      "no_speech_prob": 1.208129776775202e-11,
      "compression_ratio": 1.167808219178082
    },
    {
      "start": 210.26,
      "end": 213.6,
      "text": "8月中再加入图像与视频理解能力",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04671802160874853,
      "no_speech_prob": 1.208129776775202e-11,
      "compression_ratio": 1.167808219178082
    },
    {
      "start": 213.6,
      "end": 217.14,
      "text": "到9月的1.3版转向打磨长期任务的可用性",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04671802160874853,
      "no_speech_prob": 1.208129776775202e-11,
      "compression_ratio": 1.167808219178082
    },
    {
      "start": 217.14,
      "end": 220.5,
      "text": "几乎保持每个月一次大版本更新的速度",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04671802160874853,
      "no_speech_prob": 1.208129776775202e-11,
      "compression_ratio": 1.167808219178082
    },
    {
      "start": 220.5,
      "end": 223.0,
      "text": "这种快速迭代在AI圈不算稀奇",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04671802160874853,
      "no_speech_prob": 1.208129776775202e-11,
      "compression_ratio": 1.167808219178082
    },
    {
      "start": 223.0,
      "end": 227.68,
      "text": "但难得的是Meta几乎每一版都会附上明确的内部对比数据",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.052933514436935995,
      "no_speech_prob": 1.3114118298240829e-11,
      "compression_ratio": 1.2148997134670487
    },
    {
      "start": 227.68,
      "end": 229.96,
      "text": "同时引用第三方评测结果",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.052933514436935995,
      "no_speech_prob": 1.3114118298240829e-11,
      "compression_ratio": 1.2148997134670487
    },
    {
      "start": 229.96,
      "end": 232.88,
      "text": "给了外界一个相对透明的观察窗口",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.052933514436935995,
      "no_speech_prob": 1.3114118298240829e-11,
      "compression_ratio": 1.2148997134670487
    },
    {
      "start": 232.88,
      "end": 235.22,
      "text": "很多人只盯着分数涨了多少",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.052933514436935995,
      "no_speech_prob": 1.3114118298240829e-11,
      "compression_ratio": 1.2148997134670487
    },
    {
      "start": 235.22,
      "end": 238.02,
      "text": "却忽略了每一代版本的定位差异",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.052933514436935995,
      "no_speech_prob": 1.3114118298240829e-11,
      "compression_ratio": 1.2148997134670487
    },
    {
      "start": 238.02,
      "end": 242.2,
      "text": "Spark 1.3并不是全面碾压1.2的终极进化",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.052933514436935995,
      "no_speech_prob": 1.3114118298240829e-11,
      "compression_ratio": 1.2148997134670487
    },
    {
      "start": 242.2,
      "end": 245.6,
      "text": "而是针对代理人任务的可靠性做了针对性强化",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.052933514436935995,
      "no_speech_prob": 1.3114118298240829e-11,
      "compression_ratio": 1.2148997134670487
    },
    {
      "start": 245.6,
      "end": 249.42,
      "text": "甚至在部分涨上下文指标上还出现了小幅退步",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.052933514436935995,
      "no_speech_prob": 1.3114118298240829e-11,
      "compression_ratio": 1.2148997134670487
    },
    {
      "start": 249.42,
      "end": 251.24,
      "text": "这种有得有失的升级",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.052933514436935995,
      "no_speech_prob": 1.3114118298240829e-11,
      "compression_ratio": 1.2148997134670487
    },
    {
      "start": 251.24,
      "end": 254.86,
      "text": "恰恰说明模型能力不是简单的线性叠加",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06625737557864056,
      "no_speech_prob": 1.4014711266496782e-11,
      "compression_ratio": 1.3529411764705883
    },
    {
      "start": 254.86,
      "end": 257.66,
      "text": "而是一连串的取舍与权衡",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06625737557864056,
      "no_speech_prob": 1.4014711266496782e-11,
      "compression_ratio": 1.3529411764705883
    },
    {
      "start": 257.66,
      "end": 259.96,
      "text": "如果你只看营销文案里的亮点",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06625737557864056,
      "no_speech_prob": 1.4014711266496782e-11,
      "compression_ratio": 1.3529411764705883
    },
    {
      "start": 259.96,
      "end": 263.14,
      "text": "很快就会在实际使用中发现落差",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06625737557864056,
      "no_speech_prob": 1.4014711266496782e-11,
      "compression_ratio": 1.3529411764705883
    },
    {
      "start": 263.14,
      "end": 265.5,
      "text": "如果你翻完这次的全部技术文件",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06625737557864056,
      "no_speech_prob": 1.4014711266496782e-11,
      "compression_ratio": 1.3529411764705883
    },
    {
      "start": 265.5,
      "end": 267.4,
      "text": "会发现最核心的关键词",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06625737557864056,
      "no_speech_prob": 1.4014711266496782e-11,
      "compression_ratio": 1.3529411764705883
    },
    {
      "start": 267.4,
      "end": 268.76,
      "text": "不是更高的分数",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06625737557864056,
      "no_speech_prob": 1.4014711266496782e-11,
      "compression_ratio": 1.3529411764705883
    },
    {
      "start": 268.76,
      "end": 270.28,
      "text": "也不是更多的参数",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06625737557864056,
      "no_speech_prob": 1.4014711266496782e-11,
      "compression_ratio": 1.3529411764705883
    },
    {
      "start": 270.28,
      "end": 272.3,
      "text": "而是长任务代理人",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06625737557864056,
      "no_speech_prob": 1.4014711266496782e-11,
      "compression_ratio": 1.3529411764705883
    },
    {
      "start": 272.3,
      "end": 273.44,
      "text": "什么意思",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06625737557864056,
      "no_speech_prob": 1.4014711266496782e-11,
      "compression_ratio": 1.3529411764705883
    },
    {
      "start": 273.44,
      "end": 276.5,
      "text": "就是让AI不再只是一次性回答一个问题",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06625737557864056,
      "no_speech_prob": 1.4014711266496782e-11,
      "compression_ratio": 1.3529411764705883
    },
    {
      "start": 276.5,
      "end": 279.16,
      "text": "而是像一个真正的执行助理一样",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06625737557864056,
      "no_speech_prob": 1.4014711266496782e-11,
      "compression_ratio": 1.3529411764705883
    },
    {
      "start": 279.16,
      "end": 281.22,
      "text": "接下一个模糊的需求",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06625737557864056,
      "no_speech_prob": 1.4014711266496782e-11,
      "compression_ratio": 1.3529411764705883
    },
    {
      "start": 281.22,
      "end": 283.48,
      "text": "自己拆解步骤 呼叫工具",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06597136976707041,
      "no_speech_prob": 1.5569556061079126e-11,
      "compression_ratio": 1.2526881720430108
    },
    {
      "start": 283.48,
      "end": 284.88,
      "text": "遇到问题回头问你",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06597136976707041,
      "no_speech_prob": 1.5569556061079126e-11,
      "compression_ratio": 1.2526881720430108
    },
    {
      "start": 284.88,
      "end": 286.1,
      "text": "碰到危险操作",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06597136976707041,
      "no_speech_prob": 1.5569556061079126e-11,
      "compression_ratio": 1.2526881720430108
    },
    {
      "start": 286.1,
      "end": 289.38,
      "text": "先停下来确认最后再交付完整的成果",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06597136976707041,
      "no_speech_prob": 1.5569556061079126e-11,
      "compression_ratio": 1.2526881720430108
    },
    {
      "start": 289.38,
      "end": 290.48,
      "text": "这听起来很美好",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06597136976707041,
      "no_speech_prob": 1.5569556061079126e-11,
      "compression_ratio": 1.2526881720430108
    },
    {
      "start": 290.48,
      "end": 293.68,
      "text": "但过去大部分代理人模型的问题无非三种",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06597136976707041,
      "no_speech_prob": 1.5569556061079126e-11,
      "compression_ratio": 1.2526881720430108
    },
    {
      "start": 293.68,
      "end": 295.88,
      "text": "要么不敢动手只会空泛回答",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06597136976707041,
      "no_speech_prob": 1.5569556061079126e-11,
      "compression_ratio": 1.2526881720430108
    },
    {
      "start": 295.88,
      "end": 298.68,
      "text": "要么乱动手做到一半偏离目标",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06597136976707041,
      "no_speech_prob": 1.5569556061079126e-11,
      "compression_ratio": 1.2526881720430108
    },
    {
      "start": 298.68,
      "end": 302.0,
      "text": "更常见的是干到一半彻底忘了最原始的要求",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06597136976707041,
      "no_speech_prob": 1.5569556061079126e-11,
      "compression_ratio": 1.2526881720430108
    },
    {
      "start": 302.0,
      "end": 305.02,
      "text": "Spark 1.3就是冲着这些痛点做的优化",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06597136976707041,
      "no_speech_prob": 1.5569556061079126e-11,
      "compression_ratio": 1.2526881720430108
    },
    {
      "start": 305.02,
      "end": 306.46,
      "text": "举个很实际的场景",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06597136976707041,
      "no_speech_prob": 1.5569556061079126e-11,
      "compression_ratio": 1.2526881720430108
    },
    {
      "start": 306.46,
      "end": 307.62,
      "text": "如果你跟模型说",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06597136976707041,
      "no_speech_prob": 1.5569556061079126e-11,
      "compression_ratio": 1.2526881720430108
    },
    {
      "start": 307.62,
      "end": 310.52,
      "text": "帮我整理上礼拜的三场客户会议记录",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06597136976707041,
      "no_speech_prob": 1.5569556061079126e-11,
      "compression_ratio": 1.2526881720430108
    },
    {
      "start": 310.52,
      "end": 311.8,
      "text": "做成部门简报",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.060401537031403385,
      "no_speech_prob": 1.296110874876577e-11,
      "compression_ratio": 1.300578034682081
    },
    {
      "start": 311.8,
      "end": 314.12,
      "text": "把每个客户的行动项列出来",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.060401537031403385,
      "no_speech_prob": 1.296110874876577e-11,
      "compression_ratio": 1.300578034682081
    },
    {
      "start": 314.12,
      "end": 315.44,
      "text": "最后寄给主管",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.060401537031403385,
      "no_speech_prob": 1.296110874876577e-11,
      "compression_ratio": 1.300578034682081
    },
    {
      "start": 315.44,
      "end": 316.68,
      "text": "换做之前的模型",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.060401537031403385,
      "no_speech_prob": 1.296110874876577e-11,
      "compression_ratio": 1.300578034682081
    },
    {
      "start": 316.68,
      "end": 319.78,
      "text": "很可能做到一半就忘了要按客户分类",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.060401537031403385,
      "no_speech_prob": 1.296110874876577e-11,
      "compression_ratio": 1.300578034682081
    },
    {
      "start": 319.78,
      "end": 321.8,
      "text": "或者漏了寄信的步骤",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.060401537031403385,
      "no_speech_prob": 1.296110874876577e-11,
      "compression_ratio": 1.300578034682081
    },
    {
      "start": 321.8,
      "end": 324.98,
      "text": "甚至为了凑数编造几个不存在的行动项",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.060401537031403385,
      "no_speech_prob": 1.296110874876577e-11,
      "compression_ratio": 1.300578034682081
    },
    {
      "start": 324.98,
      "end": 327.92,
      "text": "而Spark 1.3的处理逻辑会完全不同",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.060401537031403385,
      "no_speech_prob": 1.296110874876577e-11,
      "compression_ratio": 1.300578034682081
    },
    {
      "start": 327.92,
      "end": 330.28,
      "text": "一开始如果觉得需求不够清楚",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.060401537031403385,
      "no_speech_prob": 1.296110874876577e-11,
      "compression_ratio": 1.300578034682081
    },
    {
      "start": 330.28,
      "end": 331.92,
      "text": "它会主动追问你",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.060401537031403385,
      "no_speech_prob": 1.296110874876577e-11,
      "compression_ratio": 1.300578034682081
    },
    {
      "start": 331.92,
      "end": 334.8,
      "text": "简报要的是正式版本还是讨论版",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.060401537031403385,
      "no_speech_prob": 1.296110874876577e-11,
      "compression_ratio": 1.300578034682081
    },
    {
      "start": 334.8,
      "end": 337.5,
      "text": "行动项要不要加上预计完成时间",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.060401537031403385,
      "no_speech_prob": 1.296110874876577e-11,
      "compression_ratio": 1.300578034682081
    },
    {
      "start": 337.5,
      "end": 340.44,
      "text": "寄信的时候要不要附上原始记录档",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.060401537031403385,
      "no_speech_prob": 1.296110874876577e-11,
      "compression_ratio": 1.300578034682081
    },
    {
      "start": 340.44,
      "end": 344.66,
      "text": "执行过程中如果发现某份会议记录打不开或者权限不足",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04757784359419166,
      "no_speech_prob": 1.5391561289379574e-11,
      "compression_ratio": 1.3421052631578947
    },
    {
      "start": 344.66,
      "end": 346.58,
      "text": "它会停下来跟你要资讯",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04757784359419166,
      "no_speech_prob": 1.5391561289379574e-11,
      "compression_ratio": 1.3421052631578947
    },
    {
      "start": 346.58,
      "end": 349.16,
      "text": "而不是硬着头皮编一个结果",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04757784359419166,
      "no_speech_prob": 1.5391561289379574e-11,
      "compression_ratio": 1.3421052631578947
    },
    {
      "start": 349.16,
      "end": 350.46,
      "text": "如果碰到删除档案",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04757784359419166,
      "no_speech_prob": 1.5391561289379574e-11,
      "compression_ratio": 1.3421052631578947
    },
    {
      "start": 350.46,
      "end": 351.64,
      "text": "发送批量邮件",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04757784359419166,
      "no_speech_prob": 1.5391561289379574e-11,
      "compression_ratio": 1.3421052631578947
    },
    {
      "start": 351.64,
      "end": 355.08,
      "text": "修改正式环境设定这类有不可逆风险的操作",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04757784359419166,
      "no_speech_prob": 1.5391561289379574e-11,
      "compression_ratio": 1.3421052631578947
    },
    {
      "start": 355.08,
      "end": 357.02,
      "text": "它一定会先跳出确认",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04757784359419166,
      "no_speech_prob": 1.5391561289379574e-11,
      "compression_ratio": 1.3421052631578947
    },
    {
      "start": 357.02,
      "end": 358.48,
      "text": "问你确定要执行吗",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04757784359419166,
      "no_speech_prob": 1.5391561289379574e-11,
      "compression_ratio": 1.3421052631578947
    },
    {
      "start": 358.48,
      "end": 359.98,
      "text": "不会自己直接动手",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04757784359419166,
      "no_speech_prob": 1.5391561289379574e-11,
      "compression_ratio": 1.3421052631578947
    },
    {
      "start": 359.98,
      "end": 360.88,
      "text": "更重要的是",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04757784359419166,
      "no_speech_prob": 1.5391561289379574e-11,
      "compression_ratio": 1.3421052631578947
    },
    {
      "start": 360.88,
      "end": 364.18,
      "text": "它能分辨你现在是在补充当前任务的资讯",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04757784359419166,
      "no_speech_prob": 1.5391561289379574e-11,
      "compression_ratio": 1.3421052631578947
    },
    {
      "start": 364.18,
      "end": 366.6,
      "text": "还是要打断手头的事做别的",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04757784359419166,
      "no_speech_prob": 1.5391561289379574e-11,
      "compression_ratio": 1.3421052631578947
    },
    {
      "start": 366.6,
      "end": 369.38,
      "text": "又或者是要开启一个全新的任务",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.04757784359419166,
      "no_speech_prob": 1.5391561289379574e-11,
      "compression_ratio": 1.3421052631578947
    },
    {
      "start": 369.38,
      "end": 381.48,
      "text": "这听起来好像是基础能力,但在多轮对话里,很多模型会把用户随口的插画当成新指令,直接把之前的上下文全部覆盖掉,导致任务做到一半就断了。",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.05171555803533186,
      "no_speech_prob": 1.6193586402368787e-11,
      "compression_ratio": 1.088
    },
    {
      "start": 381.48,
      "end": 388.0,
      "text": "针对长期任务最常见的约束遗忘问题,Spark 1.3也做了专门强化。",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.05171555803533186,
      "no_speech_prob": 1.6193586402368787e-11,
      "compression_ratio": 1.088
    },
    {
      "start": 388.0,
      "end": 389.6,
      "text": "比如你一开始就说过",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057497377395629885,
      "no_speech_prob": 1.5939641967444018e-11,
      "compression_ratio": 1.319327731092437
    },
    {
      "start": 389.6,
      "end": 391.88,
      "text": "所有输出都要用正体中文",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057497377395629885,
      "no_speech_prob": 1.5939641967444018e-11,
      "compression_ratio": 1.319327731092437
    },
    {
      "start": 391.88,
      "end": 393.52,
      "text": "不能用外部图片素材",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057497377395629885,
      "no_speech_prob": 1.5939641967444018e-11,
      "compression_ratio": 1.319327731092437
    },
    {
      "start": 393.52,
      "end": 395.4,
      "text": "数据不能超过指定范围",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057497377395629885,
      "no_speech_prob": 1.5939641967444018e-11,
      "compression_ratio": 1.319327731092437
    },
    {
      "start": 395.4,
      "end": 397.16,
      "text": "他会在整个执行过程中",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057497377395629885,
      "no_speech_prob": 1.5939641967444018e-11,
      "compression_ratio": 1.319327731092437
    },
    {
      "start": 397.16,
      "end": 398.76,
      "text": "反复检查这些条件",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057497377395629885,
      "no_speech_prob": 1.5939641967444018e-11,
      "compression_ratio": 1.319327731092437
    },
    {
      "start": 398.76,
      "end": 401.38,
      "text": "而不是跑个五六步之后就全忘了",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057497377395629885,
      "no_speech_prob": 1.5939641967444018e-11,
      "compression_ratio": 1.319327731092437
    },
    {
      "start": 401.38,
      "end": 402.44,
      "text": "最关键的一点",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057497377395629885,
      "no_speech_prob": 1.5939641967444018e-11,
      "compression_ratio": 1.319327731092437
    },
    {
      "start": 402.44,
      "end": 404.92,
      "text": "是他对自身能力边界的判断",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057497377395629885,
      "no_speech_prob": 1.5939641967444018e-11,
      "compression_ratio": 1.319327731092437
    },
    {
      "start": 404.92,
      "end": 406.4,
      "text": "过去很多代理人模型",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057497377395629885,
      "no_speech_prob": 1.5939641967444018e-11,
      "compression_ratio": 1.319327731092437
    },
    {
      "start": 406.4,
      "end": 407.42,
      "text": "最危险的习惯",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057497377395629885,
      "no_speech_prob": 1.5939641967444018e-11,
      "compression_ratio": 1.319327731092437
    },
    {
      "start": 407.42,
      "end": 408.88,
      "text": "就是任务做不完的时候",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057497377395629885,
      "no_speech_prob": 1.5939641967444018e-11,
      "compression_ratio": 1.319327731092437
    },
    {
      "start": 408.88,
      "end": 410.12,
      "text": "会假装做完了",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057497377395629885,
      "no_speech_prob": 1.5939641967444018e-11,
      "compression_ratio": 1.319327731092437
    },
    {
      "start": 410.12,
      "end": 411.26,
      "text": "然后生成一份",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057497377395629885,
      "no_speech_prob": 1.5939641967444018e-11,
      "compression_ratio": 1.319327731092437
    },
    {
      "start": 411.26,
      "end": 412.86,
      "text": "看起来很完整的报告",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057497377395629885,
      "no_speech_prob": 1.5939641967444018e-11,
      "compression_ratio": 1.319327731092437
    },
    {
      "start": 412.86,
      "end": 415.4,
      "text": "里面全是编造的数据和结论",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057497377395629885,
      "no_speech_prob": 1.5939641967444018e-11,
      "compression_ratio": 1.319327731092437
    },
    {
      "start": 415.4,
      "end": 417.22,
      "text": "普通用户根本看不出来",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.057497377395629885,
      "no_speech_prob": 1.5939641967444018e-11,
      "compression_ratio": 1.319327731092437
    },
    {
      "start": 417.22,
      "end": 421.26,
      "text": "Bug1.3在训练时特别强化了这点遇到做不到的地方",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06431577384816428,
      "no_speech_prob": 1.2759732506695265e-11,
      "compression_ratio": 1.2577639751552796
    },
    {
      "start": 421.26,
      "end": 422.86,
      "text": "他更倾向于直接说",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06431577384816428,
      "no_speech_prob": 1.2759732506695265e-11,
      "compression_ratio": 1.2577639751552796
    },
    {
      "start": 422.86,
      "end": 424.16,
      "text": "我卡在这里了",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06431577384816428,
      "no_speech_prob": 1.2759732506695265e-11,
      "compression_ratio": 1.2577639751552796
    },
    {
      "start": 424.16,
      "end": 425.5,
      "text": "我缺少这个资讯",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06431577384816428,
      "no_speech_prob": 1.2759732506695265e-11,
      "compression_ratio": 1.2577639751552796
    },
    {
      "start": 425.5,
      "end": 429.24,
      "text": "而不是硬凑出一份看似完美的答案",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06431577384816428,
      "no_speech_prob": 1.2759732506695265e-11,
      "compression_ratio": 1.2577639751552796
    },
    {
      "start": 429.24,
      "end": 432.06,
      "text": "Meta内部工程师对比1.2版本后发现",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06431577384816428,
      "no_speech_prob": 1.2759732506695265e-11,
      "compression_ratio": 1.2577639751552796
    },
    {
      "start": 432.06,
      "end": 435.82,
      "text": "新版本平均减少了20%的工具呼叫次数",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06431577384816428,
      "no_speech_prob": 1.2759732506695265e-11,
      "compression_ratio": 1.2577639751552796
    },
    {
      "start": 435.82,
      "end": 438.88,
      "text": "同时减少了25%的token消耗",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06431577384816428,
      "no_speech_prob": 1.2759732506695265e-11,
      "compression_ratio": 1.2577639751552796
    },
    {
      "start": 438.88,
      "end": 441.34,
      "text": "这两个数字其实比跑分更有意义",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06431577384816428,
      "no_speech_prob": 1.2759732506695265e-11,
      "compression_ratio": 1.2577639751552796
    },
    {
      "start": 441.34,
      "end": 444.02,
      "text": "工具呼叫减少代表模型更清楚",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06431577384816428,
      "no_speech_prob": 1.2759732506695265e-11,
      "compression_ratio": 1.2577639751552796
    },
    {
      "start": 444.02,
      "end": 445.74,
      "text": "什么时候该用什么工具",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06431577384816428,
      "no_speech_prob": 1.2759732506695265e-11,
      "compression_ratio": 1.2577639751552796
    },
    {
      "start": 445.74,
      "end": 448.58,
      "text": "不会无脑尝试所有选项来回绕路",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06311481475830077,
      "no_speech_prob": 1.33725287787545e-11,
      "compression_ratio": 1.2792022792022792
    },
    {
      "start": 448.58,
      "end": 450.0,
      "text": "Token消耗减少",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06311481475830077,
      "no_speech_prob": 1.33725287787545e-11,
      "compression_ratio": 1.2792022792022792
    },
    {
      "start": 450.0,
      "end": 451.68,
      "text": "代表推理效率提升",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06311481475830077,
      "no_speech_prob": 1.33725287787545e-11,
      "compression_ratio": 1.2792022792022792
    },
    {
      "start": 451.68,
      "end": 453.84,
      "text": "单位任务的成本直接下降",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06311481475830077,
      "no_speech_prob": 1.33725287787545e-11,
      "compression_ratio": 1.2792022792022792
    },
    {
      "start": 453.84,
      "end": 455.88,
      "text": "很多人不知道代理人任务的成本",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06311481475830077,
      "no_speech_prob": 1.33725287787545e-11,
      "compression_ratio": 1.2792022792022792
    },
    {
      "start": 455.88,
      "end": 458.3,
      "text": "通常是纯问答的3到5倍",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06311481475830077,
      "no_speech_prob": 1.33725287787545e-11,
      "compression_ratio": 1.2792022792022792
    },
    {
      "start": 458.3,
      "end": 460.12,
      "text": "因为要不断多轮推理",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06311481475830077,
      "no_speech_prob": 1.33725287787545e-11,
      "compression_ratio": 1.2792022792022792
    },
    {
      "start": 460.12,
      "end": 461.8,
      "text": "呼叫工具修正路线",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06311481475830077,
      "no_speech_prob": 1.33725287787545e-11,
      "compression_ratio": 1.2792022792022792
    },
    {
      "start": 461.8,
      "end": 463.08,
      "text": "所以效率的提升",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06311481475830077,
      "no_speech_prob": 1.33725287787545e-11,
      "compression_ratio": 1.2792022792022792
    },
    {
      "start": 463.08,
      "end": 464.52,
      "text": "直接决定了这项技术",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06311481475830077,
      "no_speech_prob": 1.33725287787545e-11,
      "compression_ratio": 1.2792022792022792
    },
    {
      "start": 464.52,
      "end": 466.22,
      "text": "能不能大规模商业化",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06311481475830077,
      "no_speech_prob": 1.33725287787545e-11,
      "compression_ratio": 1.2792022792022792
    },
    {
      "start": 466.22,
      "end": 467.38,
      "text": "不过也要说清楚",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06311481475830077,
      "no_speech_prob": 1.33725287787545e-11,
      "compression_ratio": 1.2792022792022792
    },
    {
      "start": 467.38,
      "end": 470.9,
      "text": "这两个数字是Meta内部测试场景下的结果",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06311481475830077,
      "no_speech_prob": 1.33725287787545e-11,
      "compression_ratio": 1.2792022792022792
    },
    {
      "start": 470.9,
      "end": 472.88,
      "text": "不是第三方独立评测",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06311481475830077,
      "no_speech_prob": 1.33725287787545e-11,
      "compression_ratio": 1.2792022792022792
    },
    {
      "start": 472.88,
      "end": 475.26,
      "text": "实际业务中因为场景更复杂",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.06311481475830077,
      "no_speech_prob": 1.33725287787545e-11,
      "compression_ratio": 1.2792022792022792
    },
    {
      "start": 475.26,
      "end": 477.34,
      "text": "不一定能复现同样的比例",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07353733607700892,
      "no_speech_prob": 9.329810361780044e-12,
      "compression_ratio": 0.896551724137931
    },
    {
      "start": 477.34,
      "end": 480.0,
      "text": "官方也举了几个典型的常任务场景",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07353733607700892,
      "no_speech_prob": 9.329810361780044e-12,
      "compression_ratio": 0.896551724137931
    },
    {
      "start": 480.0,
      "end": 492.98,
      "text": "官方也举了几个典型的常任务场景,比如工程领域的CAD模型搭配流体模拟数据。最后按指定章节输出PDF报告,音乐制作领域根据时间码定位错误音符,从重复段落里挑出合适的替代片段。",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06585335189645941,
      "no_speech_prob": 9.030163769518929e-12,
      "compression_ratio": 1.2477611940298508
    },
    {
      "start": 492.98,
      "end": 494.92,
      "text": "保持所有音轨长度不变",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06964309825453648,
      "no_speech_prob": 1.4976842682701275e-11,
      "compression_ratio": 1.2724458204334366
    },
    {
      "start": 494.92,
      "end": 496.04,
      "text": "再重新混音",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06964309825453648,
      "no_speech_prob": 1.4976842682701275e-11,
      "compression_ratio": 1.2724458204334366
    },
    {
      "start": 496.04,
      "end": 498.64,
      "text": "这些场景的共同点就是步骤多",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06964309825453648,
      "no_speech_prob": 1.4976842682701275e-11,
      "compression_ratio": 1.2724458204334366
    },
    {
      "start": 498.64,
      "end": 499.84,
      "text": "中间产物多",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06964309825453648,
      "no_speech_prob": 1.4976842682701275e-11,
      "compression_ratio": 1.2724458204334366
    },
    {
      "start": 499.84,
      "end": 501.4,
      "text": "需要跨工具协作",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06964309825453648,
      "no_speech_prob": 1.4976842682701275e-11,
      "compression_ratio": 1.2724458204334366
    },
    {
      "start": 501.4,
      "end": 503.86,
      "text": "差一步整个结果就不能用",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06964309825453648,
      "no_speech_prob": 1.4976842682701275e-11,
      "compression_ratio": 1.2724458204334366
    },
    {
      "start": 503.86,
      "end": 504.94,
      "text": "从这个角度看",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06964309825453648,
      "no_speech_prob": 1.4976842682701275e-11,
      "compression_ratio": 1.2724458204334366
    },
    {
      "start": 504.94,
      "end": 508.54,
      "text": "Spark 1.3根本不是一个聊天更厉害的模型",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06964309825453648,
      "no_speech_prob": 1.4976842682701275e-11,
      "compression_ratio": 1.2724458204334366
    },
    {
      "start": 508.54,
      "end": 510.72,
      "text": "而是一个干活更稳的模型",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06964309825453648,
      "no_speech_prob": 1.4976842682701275e-11,
      "compression_ratio": 1.2724458204334366
    },
    {
      "start": 510.72,
      "end": 513.58,
      "text": "它的评价标准早就不是会不会写诗",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06964309825453648,
      "no_speech_prob": 1.4976842682701275e-11,
      "compression_ratio": 1.2724458204334366
    },
    {
      "start": 513.58,
      "end": 514.94,
      "text": "能不能解数学题",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06964309825453648,
      "no_speech_prob": 1.4976842682701275e-11,
      "compression_ratio": 1.2724458204334366
    },
    {
      "start": 514.94,
      "end": 516.96,
      "text": "而是能不能在100步之后",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06964309825453648,
      "no_speech_prob": 1.4976842682701275e-11,
      "compression_ratio": 1.2724458204334366
    },
    {
      "start": 516.96,
      "end": 518.78,
      "text": "还不偏离最初的目标",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06964309825453648,
      "no_speech_prob": 1.4976842682701275e-11,
      "compression_ratio": 1.2724458204334366
    },
    {
      "start": 518.78,
      "end": 520.58,
      "text": "聊完了官方自己讲的升级",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06964309825453648,
      "no_speech_prob": 1.4976842682701275e-11,
      "compression_ratio": 1.2724458204334366
    },
    {
      "start": 520.58,
      "end": 523.22,
      "text": "我们来看看第三方独立评测的数据",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07907958030700683,
      "no_speech_prob": 1.3687819974683713e-11,
      "compression_ratio": 1.2046109510086456
    },
    {
      "start": 523.22,
      "end": 526.3,
      "text": "验证一下这些能力到底有没有落到实处",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07907958030700683,
      "no_speech_prob": 1.3687819974683713e-11,
      "compression_ratio": 1.2046109510086456
    },
    {
      "start": 526.3,
      "end": 528.88,
      "text": "这次的数据全部来自Artificial Analysis",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07907958030700683,
      "no_speech_prob": 1.3687819974683713e-11,
      "compression_ratio": 1.2046109510086456
    },
    {
      "start": 528.88,
      "end": 532.8,
      "text": "这是目前业界专门做代理人模型评测的独立机构",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07907958030700683,
      "no_speech_prob": 1.3687819974683713e-11,
      "compression_ratio": 1.2046109510086456
    },
    {
      "start": 532.8,
      "end": 534.4,
      "text": "不是Meta自己跑的分",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07907958030700683,
      "no_speech_prob": 1.3687819974683713e-11,
      "compression_ratio": 1.2046109510086456
    },
    {
      "start": 534.4,
      "end": 536.12,
      "text": "参考价值相对高一些",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07907958030700683,
      "no_speech_prob": 1.3687819974683713e-11,
      "compression_ratio": 1.2046109510086456
    },
    {
      "start": 536.12,
      "end": 538.76,
      "text": "在他们的综合智慧指数榜单上",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07907958030700683,
      "no_speech_prob": 1.3687819974683713e-11,
      "compression_ratio": 1.2046109510086456
    },
    {
      "start": 538.76,
      "end": 542.14,
      "text": "Spark 1.3的Shy版本拿到61分",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07907958030700683,
      "no_speech_prob": 1.3687819974683713e-11,
      "compression_ratio": 1.2046109510086456
    },
    {
      "start": 542.14,
      "end": 546.22,
      "text": "还在有限预览阶段的Max版本拿到62分",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07907958030700683,
      "no_speech_prob": 1.3687819974683713e-11,
      "compression_ratio": 1.2046109510086456
    },
    {
      "start": 546.22,
      "end": 548.4,
      "text": "比起1.2版本的57分",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07907958030700683,
      "no_speech_prob": 1.3687819974683713e-11,
      "compression_ratio": 1.2046109510086456
    },
    {
      "start": 548.4,
      "end": 550.36,
      "text": "Exia整整涨了4分",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07907958030700683,
      "no_speech_prob": 1.3687819974683713e-11,
      "compression_ratio": 1.2046109510086456
    },
    {
      "start": 550.36,
      "end": 555.3,
      "text": "在AI评测里 4分的涨幅算不算大 要看在哪个区间",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05868639444050036,
      "no_speech_prob": 1.1754796788721045e-11,
      "compression_ratio": 1.2039473684210527
    },
    {
      "start": 555.3,
      "end": 559.3,
      "text": "如果从30分涨到34分 那只是入门级的进步",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05868639444050036,
      "no_speech_prob": 1.1754796788721045e-11,
      "compression_ratio": 1.2039473684210527
    },
    {
      "start": 559.3,
      "end": 564.32,
      "text": "但从57分涨到61分已经是第一梯队模型的明显提升",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05868639444050036,
      "no_speech_prob": 1.1754796788721045e-11,
      "compression_ratio": 1.2039473684210527
    },
    {
      "start": 564.32,
      "end": 567.08,
      "text": "因为越往上 每一分的难度都越高",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05868639444050036,
      "no_speech_prob": 1.1754796788721045e-11,
      "compression_ratio": 1.2039473684210527
    },
    {
      "start": 567.08,
      "end": 572.2,
      "text": "更有竞争力的其实是成本XJ版本完成一项综合指数任务",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05868639444050036,
      "no_speech_prob": 1.1754796788721045e-11,
      "compression_ratio": 1.2039473684210527
    },
    {
      "start": 572.2,
      "end": 577.98,
      "text": "平均花费0.55美元 在所有得分超过59分的模型里是最低的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05868639444050036,
      "no_speech_prob": 1.1754796788721045e-11,
      "compression_ratio": 1.2039473684210527
    },
    {
      "start": 577.98,
      "end": 579.76,
      "text": "我们可以简单对比一下",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010391649992569,
      "no_speech_prob": 1.3422810606067426e-11,
      "compression_ratio": 1.1271676300578035
    },
    {
      "start": 579.76,
      "end": 580.64,
      "text": "同级别里",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010391649992569,
      "no_speech_prob": 1.3422810606067426e-11,
      "compression_ratio": 1.1271676300578035
    },
    {
      "start": 580.64,
      "end": 583.04,
      "text": "OpenAI的GPT-4O Advanced",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010391649992569,
      "no_speech_prob": 1.3422810606067426e-11,
      "compression_ratio": 1.1271676300578035
    },
    {
      "start": 583.04,
      "end": 584.1,
      "text": "做类似任务",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010391649992569,
      "no_speech_prob": 1.3422810606067426e-11,
      "compression_ratio": 1.1271676300578035
    },
    {
      "start": 584.1,
      "end": 585.56,
      "text": "大概要1.2美元",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010391649992569,
      "no_speech_prob": 1.3422810606067426e-11,
      "compression_ratio": 1.1271676300578035
    },
    {
      "start": 585.56,
      "end": 587.7,
      "text": "Anthropic的Cloud Opus",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010391649992569,
      "no_speech_prob": 1.3422810606067426e-11,
      "compression_ratio": 1.1271676300578035
    },
    {
      "start": 587.7,
      "end": 589.5,
      "text": "则要1.5美元左右",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010391649992569,
      "no_speech_prob": 1.3422810606067426e-11,
      "compression_ratio": 1.1271676300578035
    },
    {
      "start": 589.5,
      "end": 591.62,
      "text": "Meta差不多只要一半的价格",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010391649992569,
      "no_speech_prob": 1.3422810606067426e-11,
      "compression_ratio": 1.1271676300578035
    },
    {
      "start": 591.62,
      "end": 594.98,
      "text": "这对于需要大量呼叫API的企业用户来说",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010391649992569,
      "no_speech_prob": 1.3422810606067426e-11,
      "compression_ratio": 1.1271676300578035
    },
    {
      "start": 594.98,
      "end": 596.3,
      "text": "吸引力非常大",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010391649992569,
      "no_speech_prob": 1.3422810606067426e-11,
      "compression_ratio": 1.1271676300578035
    },
    {
      "start": 596.3,
      "end": 598.56,
      "text": "我们再拆开看具体任务的表现",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010391649992569,
      "no_speech_prob": 1.3422810606067426e-11,
      "compression_ratio": 1.1271676300578035
    },
    {
      "start": 598.56,
      "end": 600.28,
      "text": "Tile3-Bench Banking",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010391649992569,
      "no_speech_prob": 1.3422810606067426e-11,
      "compression_ratio": 1.1271676300578035
    },
    {
      "start": 600.28,
      "end": 602.68,
      "text": "是专门的银行代理人基准测试",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010391649992569,
      "no_speech_prob": 1.3422810606067426e-11,
      "compression_ratio": 1.1271676300578035
    },
    {
      "start": 602.68,
      "end": 605.24,
      "text": "模拟金融场景下的多步骤任务",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010391649992569,
      "no_speech_prob": 1.3422810606067426e-11,
      "compression_ratio": 1.1271676300578035
    },
    {
      "start": 605.24,
      "end": 606.34,
      "text": "1.2版本",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010391649992569,
      "no_speech_prob": 1.3422810606067426e-11,
      "compression_ratio": 1.1271676300578035
    },
    {
      "start": 606.34,
      "end": 608.74,
      "text": "当时只拿到35%的通过率",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07984705656272931,
      "no_speech_prob": 1.6835234595280468e-11,
      "compression_ratio": 1.2278481012658229
    },
    {
      "start": 608.74,
      "end": 610.96,
      "text": "SHAE版本提升到47%",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07984705656272931,
      "no_speech_prob": 1.6835234595280468e-11,
      "compression_ratio": 1.2278481012658229
    },
    {
      "start": 610.96,
      "end": 613.64,
      "text": "MAX版本更是达到52%",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07984705656272931,
      "no_speech_prob": 1.6835234595280468e-11,
      "compression_ratio": 1.2278481012658229
    },
    {
      "start": 613.64,
      "end": 615.36,
      "text": "这个涨幅其实相当惊人",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07984705656272931,
      "no_speech_prob": 1.6835234595280468e-11,
      "compression_ratio": 1.2278481012658229
    },
    {
      "start": 615.36,
      "end": 617.26,
      "text": "说明模型在处理合规",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07984705656272931,
      "no_speech_prob": 1.6835234595280468e-11,
      "compression_ratio": 1.2278481012658229
    },
    {
      "start": 617.26,
      "end": 620.24,
      "text": "要求高步骤繁琐的金融任务时",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07984705656272931,
      "no_speech_prob": 1.6835234595280468e-11,
      "compression_ratio": 1.2278481012658229
    },
    {
      "start": 620.24,
      "end": 622.84,
      "text": "可靠性比上一代强了很多",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07984705656272931,
      "no_speech_prob": 1.6835234595280468e-11,
      "compression_ratio": 1.2278481012658229
    },
    {
      "start": 622.84,
      "end": 624.0,
      "text": "但也要客观说",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07984705656272931,
      "no_speech_prob": 1.6835234595280468e-11,
      "compression_ratio": 1.2278481012658229
    },
    {
      "start": 624.0,
      "end": 625.58,
      "text": "47%的通过率",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07984705656272931,
      "no_speech_prob": 1.6835234595280468e-11,
      "compression_ratio": 1.2278481012658229
    },
    {
      "start": 625.58,
      "end": 627.72,
      "text": "依然代表超过一半的任务",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07984705656272931,
      "no_speech_prob": 1.6835234595280468e-11,
      "compression_ratio": 1.2278481012658229
    },
    {
      "start": 627.72,
      "end": 629.04,
      "text": "没办法完美完成",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07984705656272931,
      "no_speech_prob": 1.6835234595280468e-11,
      "compression_ratio": 1.2278481012658229
    },
    {
      "start": 629.04,
      "end": 631.04,
      "text": "距离真正取代人类专员",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07984705656272931,
      "no_speech_prob": 1.6835234595280468e-11,
      "compression_ratio": 1.2278481012658229
    },
    {
      "start": 631.04,
      "end": 632.78,
      "text": "还有非常远的距离",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07984705656272931,
      "no_speech_prob": 1.6835234595280468e-11,
      "compression_ratio": 1.2278481012658229
    },
    {
      "start": 632.78,
      "end": 635.7,
      "text": "大概相当于入行半年的助理水平",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07984705656272931,
      "no_speech_prob": 1.6835234595280468e-11,
      "compression_ratio": 1.2278481012658229
    },
    {
      "start": 635.7,
      "end": 638.98,
      "text": "Terminal-Bench 2.1是终端指令执行测试",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0908617738817559,
      "no_speech_prob": 1.240930101731319e-11,
      "compression_ratio": 1.130030959752322
    },
    {
      "start": 638.98,
      "end": 641.3,
      "text": "1.2版本通过率80%",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0908617738817559,
      "no_speech_prob": 1.240930101731319e-11,
      "compression_ratio": 1.130030959752322
    },
    {
      "start": 641.3,
      "end": 643.6,
      "text": "X-Shy提升到85%",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0908617738817559,
      "no_speech_prob": 1.240930101731319e-11,
      "compression_ratio": 1.130030959752322
    },
    {
      "start": 643.6,
      "end": 645.3,
      "text": "Max到86%",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0908617738817559,
      "no_speech_prob": 1.240930101731319e-11,
      "compression_ratio": 1.130030959752322
    },
    {
      "start": 645.3,
      "end": 646.62,
      "text": "这个分数看起来很高",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0908617738817559,
      "no_speech_prob": 1.240930101731319e-11,
      "compression_ratio": 1.130030959752322
    },
    {
      "start": 646.62,
      "end": 648.64,
      "text": "但终端任务相对结构化",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0908617738817559,
      "no_speech_prob": 1.240930101731319e-11,
      "compression_ratio": 1.130030959752322
    },
    {
      "start": 648.64,
      "end": 650.5,
      "text": "指令对错的标准很明确",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0908617738817559,
      "no_speech_prob": 1.240930101731319e-11,
      "compression_ratio": 1.130030959752322
    },
    {
      "start": 650.5,
      "end": 652.24,
      "text": "能拿到高分并不意外",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0908617738817559,
      "no_speech_prob": 1.240930101731319e-11,
      "compression_ratio": 1.130030959752322
    },
    {
      "start": 652.24,
      "end": 654.6,
      "text": "参考价值不如真实工作场景",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0908617738817559,
      "no_speech_prob": 1.240930101731319e-11,
      "compression_ratio": 1.130030959752322
    },
    {
      "start": 654.6,
      "end": 657.32,
      "text": "真正值得注意的反而是一项退步的指标",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0908617738817559,
      "no_speech_prob": 1.240930101731319e-11,
      "compression_ratio": 1.130030959752322
    },
    {
      "start": 657.32,
      "end": 660.1,
      "text": "AA-LCR长上下文基准",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0908617738817559,
      "no_speech_prob": 1.240930101731319e-11,
      "compression_ratio": 1.130030959752322
    },
    {
      "start": 660.1,
      "end": 662.66,
      "text": "1.2版本有83%的正确率",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0908617738817559,
      "no_speech_prob": 1.240930101731319e-11,
      "compression_ratio": 1.130030959752322
    },
    {
      "start": 662.66,
      "end": 665.24,
      "text": "1.3版本降到了79%",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0908617738817559,
      "no_speech_prob": 1.240930101731319e-11,
      "compression_ratio": 1.130030959752322
    },
    {
      "start": 665.24,
      "end": 667.9,
      "text": "为什么长上下文能力反而退步了",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07063393791516621,
      "no_speech_prob": 2.203290327962293e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 667.9,
      "end": 670.06,
      "text": "难道模型越更新越差吗",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07063393791516621,
      "no_speech_prob": 2.203290327962293e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 670.06,
      "end": 670.92,
      "text": "当然不是",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07063393791516621,
      "no_speech_prob": 2.203290327962293e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 670.92,
      "end": 673.84,
      "text": "背后的原因是Spark 1.3调整了策略",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07063393791516621,
      "no_speech_prob": 2.203290327962293e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 673.84,
      "end": 675.14,
      "text": "遇到不确定的资讯",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07063393791516621,
      "no_speech_prob": 2.203290327962293e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 675.14,
      "end": 676.8,
      "text": "它更倾向于拒绝回答",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07063393791516621,
      "no_speech_prob": 2.203290327962293e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 676.8,
      "end": 678.66,
      "text": "而不是硬猜一个答案",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07063393791516621,
      "no_speech_prob": 2.203290327962293e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 678.66,
      "end": 680.78,
      "text": "也就是说它宁可少说几句",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07063393791516621,
      "no_speech_prob": 2.203290327962293e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 680.78,
      "end": 682.0,
      "text": "也不说错话",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07063393791516621,
      "no_speech_prob": 2.203290327962293e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 682.0,
      "end": 685.66,
      "text": "这种策略在AA-Omniscience准确率指标上",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07063393791516621,
      "no_speech_prob": 2.203290327962293e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 685.66,
      "end": 686.5,
      "text": "也能看出来",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07063393791516621,
      "no_speech_prob": 2.203290327962293e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 686.5,
      "end": 688.62,
      "text": "表面上准确率小幅回落",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07063393791516621,
      "no_speech_prob": 2.203290327962293e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 688.62,
      "end": 690.88,
      "text": "但幻觉率明显降低了",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07063393791516621,
      "no_speech_prob": 2.203290327962293e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 690.88,
      "end": 693.3,
      "text": "这其实是一个非常重要的取舍",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07063393791516621,
      "no_speech_prob": 2.203290327962293e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 693.3,
      "end": 694.96,
      "text": "你要更高的召回率",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07063393791516621,
      "no_speech_prob": 2.203290327962293e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 694.96,
      "end": 696.76,
      "text": "还是更高的准确率",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05889078745475182,
      "no_speech_prob": 1.569171181881046e-11,
      "compression_ratio": 1.3538011695906433
    },
    {
      "start": 696.76,
      "end": 698.06,
      "text": "对于普通用户来说",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05889078745475182,
      "no_speech_prob": 1.569171181881046e-11,
      "compression_ratio": 1.3538011695906433
    },
    {
      "start": 698.06,
      "end": 699.72,
      "text": "可能希望AI什么都能答",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05889078745475182,
      "no_speech_prob": 1.569171181881046e-11,
      "compression_ratio": 1.3538011695906433
    },
    {
      "start": 699.72,
      "end": 701.26,
      "text": "哪怕错一点也没关系",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05889078745475182,
      "no_speech_prob": 1.569171181881046e-11,
      "compression_ratio": 1.3538011695906433
    },
    {
      "start": 701.26,
      "end": 702.54,
      "text": "但对于企业用户",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05889078745475182,
      "no_speech_prob": 1.569171181881046e-11,
      "compression_ratio": 1.3538011695906433
    },
    {
      "start": 702.54,
      "end": 705.3,
      "text": "尤其是金融 医疗 法律这些领域",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05889078745475182,
      "no_speech_prob": 1.569171181881046e-11,
      "compression_ratio": 1.3538011695906433
    },
    {
      "start": 705.3,
      "end": 706.56,
      "text": "一个错误的答案",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05889078745475182,
      "no_speech_prob": 1.569171181881046e-11,
      "compression_ratio": 1.3538011695906433
    },
    {
      "start": 706.56,
      "end": 708.62,
      "text": "代价可能比没有答案大得多",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05889078745475182,
      "no_speech_prob": 1.569171181881046e-11,
      "compression_ratio": 1.3538011695906433
    },
    {
      "start": 708.62,
      "end": 711.16,
      "text": "之前就有企业用初代代理人模型",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05889078745475182,
      "no_speech_prob": 1.569171181881046e-11,
      "compression_ratio": 1.3538011695906433
    },
    {
      "start": 711.16,
      "end": 712.44,
      "text": "处理报销审核",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05889078745475182,
      "no_speech_prob": 1.569171181881046e-11,
      "compression_ratio": 1.3538011695906433
    },
    {
      "start": 712.44,
      "end": 715.5,
      "text": "结果模型编造了很多不存在的单据",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05889078745475182,
      "no_speech_prob": 1.569171181881046e-11,
      "compression_ratio": 1.3538011695906433
    },
    {
      "start": 715.5,
      "end": 717.6,
      "text": "造成了不小的财务漏洞",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05889078745475182,
      "no_speech_prob": 1.569171181881046e-11,
      "compression_ratio": 1.3538011695906433
    },
    {
      "start": 717.6,
      "end": 720.0,
      "text": "所以现在越是面向企业的模型",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05889078745475182,
      "no_speech_prob": 1.569171181881046e-11,
      "compression_ratio": 1.3538011695906433
    },
    {
      "start": 720.0,
      "end": 723.66,
      "text": "越把不编造放在比达对更高的优先级",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05889078745475182,
      "no_speech_prob": 1.569171181881046e-11,
      "compression_ratio": 1.3538011695906433
    },
    {
      "start": 723.66,
      "end": 724.76,
      "text": "从这个角度看",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.05889078745475182,
      "no_speech_prob": 1.569171181881046e-11,
      "compression_ratio": 1.3538011695906433
    },
    {
      "start": 724.76,
      "end": 726.98,
      "text": "这项退步反而是一种进步",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09467398552667528,
      "no_speech_prob": 1.4516057626756673e-11,
      "compression_ratio": 1.1454545454545455
    },
    {
      "start": 726.98,
      "end": 729.5,
      "text": "代表模型开始往实用化的方向走",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09467398552667528,
      "no_speech_prob": 1.4516057626756673e-11,
      "compression_ratio": 1.1454545454545455
    },
    {
      "start": 729.5,
      "end": 731.7,
      "text": "而不是为了跑分应凑答案",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09467398552667528,
      "no_speech_prob": 1.4516057626756673e-11,
      "compression_ratio": 1.1454545454545455
    },
    {
      "start": 731.7,
      "end": 733.34,
      "text": "另外很多人忽略的一点是",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09467398552667528,
      "no_speech_prob": 1.4516057626756673e-11,
      "compression_ratio": 1.1454545454545455
    },
    {
      "start": 733.34,
      "end": 735.48,
      "text": "MAX版本的分数虽然更高",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09467398552667528,
      "no_speech_prob": 1.4516057626756673e-11,
      "compression_ratio": 1.1454545454545455
    },
    {
      "start": 735.48,
      "end": 738.2,
      "text": "但代价是运算算力大幅增加",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09467398552667528,
      "no_speech_prob": 1.4516057626756673e-11,
      "compression_ratio": 1.1454545454545455
    },
    {
      "start": 738.2,
      "end": 740.48,
      "text": "根据Artificial Analysis的测试",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09467398552667528,
      "no_speech_prob": 1.4516057626756673e-11,
      "compression_ratio": 1.1454545454545455
    },
    {
      "start": 740.48,
      "end": 743.56,
      "text": "在GDPVAL-AAVFE任务里",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09467398552667528,
      "no_speech_prob": 1.4516057626756673e-11,
      "compression_ratio": 1.1454545454545455
    },
    {
      "start": 743.56,
      "end": 747.54,
      "text": "MAX版本比SY多消耗约62%的推理量",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09467398552667528,
      "no_speech_prob": 1.4516057626756673e-11,
      "compression_ratio": 1.1454545454545455
    },
    {
      "start": 747.54,
      "end": 749.54,
      "text": "就算是银行代理任务",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09467398552667528,
      "no_speech_prob": 1.4516057626756673e-11,
      "compression_ratio": 1.1454545454545455
    },
    {
      "start": 749.54,
      "end": 751.78,
      "text": "也多了28%的算力",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09467398552667528,
      "no_speech_prob": 1.4516057626756673e-11,
      "compression_ratio": 1.1454545454545455
    },
    {
      "start": 751.78,
      "end": 753.18,
      "text": "为什么会差这么多",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09467398552667528,
      "no_speech_prob": 1.4516057626756673e-11,
      "compression_ratio": 1.1454545454545455
    },
    {
      "start": 753.18,
      "end": 756.06,
      "text": "因为MAX版本会做更多的自我检查",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08106018272606103,
      "no_speech_prob": 1.4665308897821028e-11,
      "compression_ratio": 1.3069620253164558
    },
    {
      "start": 756.06,
      "end": 757.62,
      "text": "走更多的思维步骤",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08106018272606103,
      "no_speech_prob": 1.4665308897821028e-11,
      "compression_ratio": 1.3069620253164558
    },
    {
      "start": 757.62,
      "end": 761.2,
      "text": "相当于人做事情的时候反复检查好几遍",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08106018272606103,
      "no_speech_prob": 1.4665308897821028e-11,
      "compression_ratio": 1.3069620253164558
    },
    {
      "start": 761.2,
      "end": 763.26,
      "text": "自然更费时间更费算力",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08106018272606103,
      "no_speech_prob": 1.4665308897821028e-11,
      "compression_ratio": 1.3069620253164558
    },
    {
      "start": 763.26,
      "end": 764.86,
      "text": "但正确率也更高",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08106018272606103,
      "no_speech_prob": 1.4665308897821028e-11,
      "compression_ratio": 1.3069620253164558
    },
    {
      "start": 764.86,
      "end": 768.4,
      "text": "这其实带出了未来AI商业模式的一个可能方向",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08106018272606103,
      "no_speech_prob": 1.4665308897821028e-11,
      "compression_ratio": 1.3069620253164558
    },
    {
      "start": 768.4,
      "end": 770.5,
      "text": "以后的模型不会只有单一档位",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08106018272606103,
      "no_speech_prob": 1.4665308897821028e-11,
      "compression_ratio": 1.3069620253164558
    },
    {
      "start": 770.5,
      "end": 774.0,
      "text": "而是会根据任务的风险等级调整推理强度",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08106018272606103,
      "no_speech_prob": 1.4665308897821028e-11,
      "compression_ratio": 1.3069620253164558
    },
    {
      "start": 774.0,
      "end": 775.14,
      "text": "低风险的闲聊",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08106018272606103,
      "no_speech_prob": 1.4665308897821028e-11,
      "compression_ratio": 1.3069620253164558
    },
    {
      "start": 775.14,
      "end": 778.18,
      "text": "查资料用低档速度快又便宜",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08106018272606103,
      "no_speech_prob": 1.4665308897821028e-11,
      "compression_ratio": 1.3069620253164558
    },
    {
      "start": 778.18,
      "end": 779.48,
      "text": "高风险的财务",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08106018272606103,
      "no_speech_prob": 1.4665308897821028e-11,
      "compression_ratio": 1.3069620253164558
    },
    {
      "start": 779.48,
      "end": 781.18,
      "text": "法律任务用高档",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08106018272606103,
      "no_speech_prob": 1.4665308897821028e-11,
      "compression_ratio": 1.3069620253164558
    },
    {
      "start": 781.18,
      "end": 783.22,
      "text": "更慢 更贵 但更可靠",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07183174812356839,
      "no_speech_prob": 2.3329617751532972e-11,
      "compression_ratio": 1.25
    },
    {
      "start": 783.22,
      "end": 785.16,
      "text": "用户可以按自己的需求选",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07183174812356839,
      "no_speech_prob": 2.3329617751532972e-11,
      "compression_ratio": 1.25
    },
    {
      "start": 785.16,
      "end": 787.02,
      "text": "而不是为了偶尔的高需求",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07183174812356839,
      "no_speech_prob": 2.3329617751532972e-11,
      "compression_ratio": 1.25
    },
    {
      "start": 787.02,
      "end": 788.94,
      "text": "全程付最贵的价格",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07183174812356839,
      "no_speech_prob": 2.3329617751532972e-11,
      "compression_ratio": 1.25
    },
    {
      "start": 788.94,
      "end": 790.86,
      "text": "当然 目前这还只是观察",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07183174812356839,
      "no_speech_prob": 2.3329617751532972e-11,
      "compression_ratio": 1.25
    },
    {
      "start": 790.86,
      "end": 793.82,
      "text": "Meta并没有正式推出这种分级定价",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07183174812356839,
      "no_speech_prob": 2.3329617751532972e-11,
      "compression_ratio": 1.25
    },
    {
      "start": 793.82,
      "end": 795.22,
      "text": "但从技术设计上",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07183174812356839,
      "no_speech_prob": 2.3329617751532972e-11,
      "compression_ratio": 1.25
    },
    {
      "start": 795.22,
      "end": 797.02,
      "text": "已经具备了这样的基础",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07183174812356839,
      "no_speech_prob": 2.3329617751532972e-11,
      "compression_ratio": 1.25
    },
    {
      "start": 797.02,
      "end": 798.66,
      "text": "最后 还要再强调一次",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07183174812356839,
      "no_speech_prob": 2.3329617751532972e-11,
      "compression_ratio": 1.25
    },
    {
      "start": 798.66,
      "end": 800.8,
      "text": "Max版本目前只是有限预览",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07183174812356839,
      "no_speech_prob": 2.3329617751532972e-11,
      "compression_ratio": 1.25
    },
    {
      "start": 800.8,
      "end": 803.54,
      "text": "只有少数合作机构和企业能拿到权限",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07183174812356839,
      "no_speech_prob": 2.3329617751532972e-11,
      "compression_ratio": 1.25
    },
    {
      "start": 803.54,
      "end": 805.44,
      "text": "普通开发者根本呼叫不到",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07183174812356839,
      "no_speech_prob": 2.3329617751532972e-11,
      "compression_ratio": 1.25
    },
    {
      "start": 805.44,
      "end": 807.92,
      "text": "所以你看到62分那个数字的时候",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07183174812356839,
      "no_speech_prob": 2.3329617751532972e-11,
      "compression_ratio": 1.25
    },
    {
      "start": 807.92,
      "end": 810.42,
      "text": "要知道它背后是更高的成本",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07183174812356839,
      "no_speech_prob": 2.3329617751532972e-11,
      "compression_ratio": 1.25
    },
    {
      "start": 810.42,
      "end": 812.3,
      "text": "更严格的访问限制",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06809938947359721,
      "no_speech_prob": 1.8418691918875574e-11,
      "compression_ratio": 1.203804347826087
    },
    {
      "start": 812.3,
      "end": 815.4,
      "text": "不能拿它去跟别人全面开放的API比",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06809938947359721,
      "no_speech_prob": 1.8418691918875574e-11,
      "compression_ratio": 1.203804347826087
    },
    {
      "start": 815.4,
      "end": 817.16,
      "text": "然后说Meta已经赢了",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06809938947359721,
      "no_speech_prob": 1.8418691918875574e-11,
      "compression_ratio": 1.203804347826087
    },
    {
      "start": 817.16,
      "end": 818.58,
      "text": "这种对比并不公平",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06809938947359721,
      "no_speech_prob": 1.8418691918875574e-11,
      "compression_ratio": 1.203804347826087
    },
    {
      "start": 818.58,
      "end": 820.92,
      "text": "看完了负责思考和执行的大脑",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06809938947359721,
      "no_speech_prob": 1.8418691918875574e-11,
      "compression_ratio": 1.203804347826087
    },
    {
      "start": 820.92,
      "end": 824.22,
      "text": "我们再转过头来讲这次连发的另一款产品",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06809938947359721,
      "no_speech_prob": 1.8418691918875574e-11,
      "compression_ratio": 1.203804347826087
    },
    {
      "start": 824.22,
      "end": 826.0,
      "text": "负责感知输入的耳朵",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06809938947359721,
      "no_speech_prob": 1.8418691918875574e-11,
      "compression_ratio": 1.203804347826087
    },
    {
      "start": 826.0,
      "end": 827.48,
      "text": "Muse Voice Transcribe",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06809938947359721,
      "no_speech_prob": 1.8418691918875574e-11,
      "compression_ratio": 1.203804347826087
    },
    {
      "start": 827.48,
      "end": 828.98,
      "text": "很多人第一眼看到它",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06809938947359721,
      "no_speech_prob": 1.8418691918875574e-11,
      "compression_ratio": 1.203804347826087
    },
    {
      "start": 828.98,
      "end": 831.28,
      "text": "会觉得不就是个语音转文字吗",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06809938947359721,
      "no_speech_prob": 1.8418691918875574e-11,
      "compression_ratio": 1.203804347826087
    },
    {
      "start": 831.28,
      "end": 832.36,
      "text": "有什么特别的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06809938947359721,
      "no_speech_prob": 1.8418691918875574e-11,
      "compression_ratio": 1.203804347826087
    },
    {
      "start": 832.36,
      "end": 833.98,
      "text": "但实际上它的设计思路",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06809938947359721,
      "no_speech_prob": 1.8418691918875574e-11,
      "compression_ratio": 1.203804347826087
    },
    {
      "start": 833.98,
      "end": 836.72,
      "text": "跟传统的自动语音识别完全不一样",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06809938947359721,
      "no_speech_prob": 1.8418691918875574e-11,
      "compression_ratio": 1.203804347826087
    },
    {
      "start": 836.72,
      "end": 838.64,
      "text": "传统的ASR比如Whisper",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06809938947359721,
      "no_speech_prob": 1.8418691918875574e-11,
      "compression_ratio": 1.203804347826087
    },
    {
      "start": 838.64,
      "end": 840.44,
      "text": "走的是离线转写路线",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06546151141325633,
      "no_speech_prob": 2.478255101134863e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 840.44,
      "end": 842.84,
      "text": "拿到整段音频之后再慢慢处理",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06546151141325633,
      "no_speech_prob": 2.478255101134863e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 842.84,
      "end": 844.84,
      "text": "可以利用上下文反复修正",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06546151141325633,
      "no_speech_prob": 2.478255101134863e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 844.84,
      "end": 845.88,
      "text": "准确率很高",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06546151141325633,
      "no_speech_prob": 2.478255101134863e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 845.88,
      "end": 846.96,
      "text": "但延迟也大",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06546151141325633,
      "no_speech_prob": 2.478255101134863e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 846.96,
      "end": 849.56,
      "text": "根本没办法用来做实时对话",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06546151141325633,
      "no_speech_prob": 2.478255101134863e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 849.56,
      "end": 851.38,
      "text": "而以前的实时语音助手",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06546151141325633,
      "no_speech_prob": 2.478255101134863e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 851.38,
      "end": 853.94,
      "text": "比如手机上的Siri Google助理",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06546151141325633,
      "no_speech_prob": 2.478255101134863e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 853.94,
      "end": 855.32,
      "text": "为了追求低延迟",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06546151141325633,
      "no_speech_prob": 2.478255101134863e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 855.32,
      "end": 857.54,
      "text": "只能听一小段就马上输出",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06546151141325633,
      "no_speech_prob": 2.478255101134863e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 857.54,
      "end": 859.16,
      "text": "结果准确率很差",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06546151141325633,
      "no_speech_prob": 2.478255101134863e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 859.16,
      "end": 860.14,
      "text": "稍微吵一点",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06546151141325633,
      "no_speech_prob": 2.478255101134863e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 860.14,
      "end": 862.96,
      "text": "有两个人叠话或者混个外文单字",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06546151141325633,
      "no_speech_prob": 2.478255101134863e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 862.96,
      "end": 864.02,
      "text": "结果就歪掉",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06546151141325633,
      "no_speech_prob": 2.478255101134863e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 864.02,
      "end": 865.96,
      "text": "Muse Voice Transcribe想解决的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06546151141325633,
      "no_speech_prob": 2.478255101134863e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 865.96,
      "end": 867.6,
      "text": "就是又要准又要快",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06546151141325633,
      "no_speech_prob": 2.478255101134863e-11,
      "compression_ratio": 1.2225352112676056
    },
    {
      "start": 867.6,
      "end": 869.84,
      "text": "还要懂真实对话这个老难题",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06637357216513617,
      "no_speech_prob": 1.4402820949777073e-11,
      "compression_ratio": 1.2951807228915662
    },
    {
      "start": 869.84,
      "end": 872.96,
      "text": "它的做法是把音频切成80毫秒的小块",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06637357216513617,
      "no_speech_prob": 1.4402820949777073e-11,
      "compression_ratio": 1.2951807228915662
    },
    {
      "start": 872.96,
      "end": 874.94,
      "text": "每秒处理12月5日次",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06637357216513617,
      "no_speech_prob": 1.4402820949777073e-11,
      "compression_ratio": 1.2951807228915662
    },
    {
      "start": 874.94,
      "end": 877.98,
      "text": "每一小块音频先转换成一个软Token",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06637357216513617,
      "no_speech_prob": 1.4402820949777073e-11,
      "compression_ratio": 1.2951807228915662
    },
    {
      "start": 877.98,
      "end": 881.08,
      "text": "模型在判断是要继续等后面的声音",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06637357216513617,
      "no_speech_prob": 1.4402820949777073e-11,
      "compression_ratio": 1.2951807228915662
    },
    {
      "start": 881.08,
      "end": 882.76,
      "text": "还是立刻输出文字",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06637357216513617,
      "no_speech_prob": 1.4402820949777073e-11,
      "compression_ratio": 1.2951807228915662
    },
    {
      "start": 882.76,
      "end": 884.5,
      "text": "等到音频完全停止",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06637357216513617,
      "no_speech_prob": 1.4402820949777073e-11,
      "compression_ratio": 1.2951807228915662
    },
    {
      "start": 884.5,
      "end": 886.62,
      "text": "系统会加入一个特殊标记",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06637357216513617,
      "no_speech_prob": 1.4402820949777073e-11,
      "compression_ratio": 1.2951807228915662
    },
    {
      "start": 886.62,
      "end": 890.44,
      "text": "让模型把剩下没输出的文字一口气吐出来",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06637357216513617,
      "no_speech_prob": 1.4402820949777073e-11,
      "compression_ratio": 1.2951807228915662
    },
    {
      "start": 890.44,
      "end": 893.5,
      "text": "其中最关键的技术叫做自适应延迟",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06637357216513617,
      "no_speech_prob": 1.4402820949777073e-11,
      "compression_ratio": 1.2951807228915662
    },
    {
      "start": 893.5,
      "end": 897.08,
      "text": "模型不会给每个字都设定一样的等待时间",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06637357216513617,
      "no_speech_prob": 1.4402820949777073e-11,
      "compression_ratio": 1.2951807228915662
    },
    {
      "start": 897.08,
      "end": 899.48,
      "text": "简单确定的内容就立刻输出",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06953531804710927,
      "no_speech_prob": 1.4687593155593426e-11,
      "compression_ratio": 1.3100303951367782
    },
    {
      "start": 899.48,
      "end": 902.22,
      "text": "碰到需要后文才能判断的同音词",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06953531804710927,
      "no_speech_prob": 1.4687593155593426e-11,
      "compression_ratio": 1.3100303951367782
    },
    {
      "start": 902.22,
      "end": 903.14,
      "text": "专有名词",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06953531804710927,
      "no_speech_prob": 1.4687593155593426e-11,
      "compression_ratio": 1.3100303951367782
    },
    {
      "start": 903.14,
      "end": 905.94,
      "text": "就稍微多等一点点音频再决定",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06953531804710927,
      "no_speech_prob": 1.4687593155593426e-11,
      "compression_ratio": 1.3100303951367782
    },
    {
      "start": 905.94,
      "end": 908.16,
      "text": "这其实很像人类的同步口译",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06953531804710927,
      "no_speech_prob": 1.4687593155593426e-11,
      "compression_ratio": 1.3100303951367782
    },
    {
      "start": 908.16,
      "end": 910.06,
      "text": "译员会根据句子的难度",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06953531804710927,
      "no_speech_prob": 1.4687593155593426e-11,
      "compression_ratio": 1.3100303951367782
    },
    {
      "start": 910.06,
      "end": 911.68,
      "text": "调整自己的滞后时间",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06953531804710927,
      "no_speech_prob": 1.4687593155593426e-11,
      "compression_ratio": 1.3100303951367782
    },
    {
      "start": 911.68,
      "end": 913.04,
      "text": "不会一味求快",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06953531804710927,
      "no_speech_prob": 1.4687593155593426e-11,
      "compression_ratio": 1.3100303951367782
    },
    {
      "start": 913.04,
      "end": 914.4,
      "text": "也不会一味求准",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06953531804710927,
      "no_speech_prob": 1.4687593155593426e-11,
      "compression_ratio": 1.3100303951367782
    },
    {
      "start": 914.4,
      "end": 915.18,
      "text": "训练的时候",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06953531804710927,
      "no_speech_prob": 1.4687593155593426e-11,
      "compression_ratio": 1.3100303951367782
    },
    {
      "start": 915.18,
      "end": 917.54,
      "text": "Meta把词 错误率和延迟时间",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06953531804710927,
      "no_speech_prob": 1.4687593155593426e-11,
      "compression_ratio": 1.3100303951367782
    },
    {
      "start": 917.54,
      "end": 920.18,
      "text": "一起放进强化学习的奖励函数里",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06953531804710927,
      "no_speech_prob": 1.4687593155593426e-11,
      "compression_ratio": 1.3100303951367782
    },
    {
      "start": 920.18,
      "end": 921.9,
      "text": "如果模型只追求准确",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06953531804710927,
      "no_speech_prob": 1.4687593155593426e-11,
      "compression_ratio": 1.3100303951367782
    },
    {
      "start": 921.9,
      "end": 923.28,
      "text": "每句话都等很久",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06953531804710927,
      "no_speech_prob": 1.4687593155593426e-11,
      "compression_ratio": 1.3100303951367782
    },
    {
      "start": 923.28,
      "end": 924.24,
      "text": "得分不会高",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06953531804710927,
      "no_speech_prob": 1.4687593155593426e-11,
      "compression_ratio": 1.3100303951367782
    },
    {
      "start": 924.24,
      "end": 925.4,
      "text": "只追求速度",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.06953531804710927,
      "no_speech_prob": 1.4687593155593426e-11,
      "compression_ratio": 1.3100303951367782
    },
    {
      "start": 925.4,
      "end": 927.9,
      "text": "输出一大堆错字同样过不了关",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07038392011935894,
      "no_speech_prob": 1.5198472688715547e-11,
      "compression_ratio": 1.008130081300813
    },
    {
      "start": 927.9,
      "end": 932.46,
      "text": "最终目标就是在准确率和反应速度之间找到最佳平衡点",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07038392011935894,
      "no_speech_prob": 1.5198472688715547e-11,
      "compression_ratio": 1.008130081300813
    },
    {
      "start": 932.46,
      "end": 933.68,
      "text": "那实际表现如何",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07038392011935894,
      "no_speech_prob": 1.5198472688715547e-11,
      "compression_ratio": 1.008130081300813
    },
    {
      "start": 933.68,
      "end": 937.06,
      "text": "在Artificial Analysis的流逝语音榜单上",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07038392011935894,
      "no_speech_prob": 1.5198472688715547e-11,
      "compression_ratio": 1.008130081300813
    },
    {
      "start": 937.06,
      "end": 942.06,
      "text": "Muse Voice Transcribe的最终词错误率约3.1%",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07038392011935894,
      "no_speech_prob": 1.5198472688715547e-11,
      "compression_ratio": 1.008130081300813
    },
    {
      "start": 942.06,
      "end": 943.46,
      "text": "直接登上第一名",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07038392011935894,
      "no_speech_prob": 1.5198472688715547e-11,
      "compression_ratio": 1.008130081300813
    },
    {
      "start": 944.354,
      "end": 958.114,
      "text": "这个数字是什么概念?在流逝实时的场景下,3.1%已经非常接近人类专业数据员的水平,人类大概是2%到3%左右,相当于每讲100个字只错3个。",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08717218605247704,
      "no_speech_prob": 8.431735344649471e-12,
      "compression_ratio": 1.1295180722891567
    },
    {
      "start": 958.114,
      "end": 960.094,
      "text": "不过这里必须先打个预防针",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.060139172041236456,
      "no_speech_prob": 1.5210398912612888e-11,
      "compression_ratio": 1.2991202346041055
    },
    {
      "start": 960.094,
      "end": 962.894,
      "text": "这个成绩是英语流逝对话的测试结果",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.060139172041236456,
      "no_speech_prob": 1.5210398912612888e-11,
      "compression_ratio": 1.2991202346041055
    },
    {
      "start": 962.894,
      "end": 966.374,
      "text": "绝对不能直接套用到所有语言 所有场景",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.060139172041236456,
      "no_speech_prob": 1.5210398912612888e-11,
      "compression_ratio": 1.2991202346041055
    },
    {
      "start": 966.374,
      "end": 968.834,
      "text": "中文因为同音字多 声调复杂",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.060139172041236456,
      "no_speech_prob": 1.5210398912612888e-11,
      "compression_ratio": 1.2991202346041055
    },
    {
      "start": 968.834,
      "end": 973.214,
      "text": "同样的模型词错误率通常会比英语高2到3个百分点",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.060139172041236456,
      "no_speech_prob": 1.5210398912612888e-11,
      "compression_ratio": 1.2991202346041055
    },
    {
      "start": 973.214,
      "end": 977.294,
      "text": "如果是嘈杂环境 多人强化 或者带方言口音",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.060139172041236456,
      "no_speech_prob": 1.5210398912612888e-11,
      "compression_ratio": 1.2991202346041055
    },
    {
      "start": 977.294,
      "end": 979.014,
      "text": "表现还会再打折扣",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.060139172041236456,
      "no_speech_prob": 1.5210398912612888e-11,
      "compression_ratio": 1.2991202346041055
    },
    {
      "start": 979.014,
      "end": 982.834,
      "text": "官方也明确说 模型训练覆盖了70多种语言",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.060139172041236456,
      "no_speech_prob": 1.5210398912612888e-11,
      "compression_ratio": 1.2991202346041055
    },
    {
      "start": 982.834,
      "end": 985.734,
      "text": "其中只有25种经过充分验证",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.060139172041236456,
      "no_speech_prob": 1.5210398912612888e-11,
      "compression_ratio": 1.2991202346041055
    },
    {
      "start": 985.734,
      "end": 987.834,
      "text": "其他语言可能只是有训练过",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.060139172041236456,
      "no_speech_prob": 1.5210398912612888e-11,
      "compression_ratio": 1.2991202346041055
    },
    {
      "start": 987.834,
      "end": 990.234,
      "text": "但没有经过严格的效果评测",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05867964156130527,
      "no_speech_prob": 1.9066049086746695e-11,
      "compression_ratio": 1.2692307692307692
    },
    {
      "start": 990.234,
      "end": 992.134,
      "text": "除了基础的转写准确率",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05867964156130527,
      "no_speech_prob": 1.9066049086746695e-11,
      "compression_ratio": 1.2692307692307692
    },
    {
      "start": 992.134,
      "end": 994.194,
      "text": "还有几个功能特别值得一提",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05867964156130527,
      "no_speech_prob": 1.9066049086746695e-11,
      "compression_ratio": 1.2692307692307692
    },
    {
      "start": 994.194,
      "end": 996.714,
      "text": "首先是原生的说话人分离",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05867964156130527,
      "no_speech_prob": 1.9066049086746695e-11,
      "compression_ratio": 1.2692307692307692
    },
    {
      "start": 996.714,
      "end": 999.334,
      "text": "而且支持20人以上同时发言",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05867964156130527,
      "no_speech_prob": 1.9066049086746695e-11,
      "compression_ratio": 1.2692307692307692
    },
    {
      "start": 999.334,
      "end": 1002.094,
      "text": "传统的会议转写通常是先做语音识别",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05867964156130527,
      "no_speech_prob": 1.9066049086746695e-11,
      "compression_ratio": 1.2692307692307692
    },
    {
      "start": 1002.094,
      "end": 1004.014,
      "text": "再另外用一套声文系统",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05867964156130527,
      "no_speech_prob": 1.9066049086746695e-11,
      "compression_ratio": 1.2692307692307692
    },
    {
      "start": 1004.014,
      "end": 1005.714,
      "text": "判断每段话是谁说的",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05867964156130527,
      "no_speech_prob": 1.9066049086746695e-11,
      "compression_ratio": 1.2692307692307692
    },
    {
      "start": 1005.714,
      "end": 1007.194,
      "text": "两个系统的时间轴",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05867964156130527,
      "no_speech_prob": 1.9066049086746695e-11,
      "compression_ratio": 1.2692307692307692
    },
    {
      "start": 1007.194,
      "end": 1009.354,
      "text": "只要差一点点就会出现文字",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05867964156130527,
      "no_speech_prob": 1.9066049086746695e-11,
      "compression_ratio": 1.2692307692307692
    },
    {
      "start": 1009.354,
      "end": 1010.474,
      "text": "人认错的情况",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05867964156130527,
      "no_speech_prob": 1.9066049086746695e-11,
      "compression_ratio": 1.2692307692307692
    },
    {
      "start": 1010.474,
      "end": 1012.814,
      "text": "Meta是在同一个生成过程里",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05867964156130527,
      "no_speech_prob": 1.9066049086746695e-11,
      "compression_ratio": 1.2692307692307692
    },
    {
      "start": 1012.814,
      "end": 1015.674,
      "text": "加入说话人切换标记和身份标签",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05867964156130527,
      "no_speech_prob": 1.9066049086746695e-11,
      "compression_ratio": 1.2692307692307692
    },
    {
      "start": 1015.674,
      "end": 1018.434,
      "text": "检测到换人就先标一轮新发言",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07008492634558552,
      "no_speech_prob": 1.457392366510657e-11,
      "compression_ratio": 1.2928348909657321
    },
    {
      "start": 1018.434,
      "end": 1020.894,
      "text": "等听到足够声音再确定身份",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07008492634558552,
      "no_speech_prob": 1.457392366510657e-11,
      "compression_ratio": 1.2928348909657321
    },
    {
      "start": 1020.894,
      "end": 1023.014,
      "text": "时间轴的准确度会高很多",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07008492634558552,
      "no_speech_prob": 1.457392366510657e-11,
      "compression_ratio": 1.2928348909657321
    },
    {
      "start": 1023.014,
      "end": 1026.214,
      "text": "官方展示的长对话里包含多人插话",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07008492634558552,
      "no_speech_prob": 1.457392366510657e-11,
      "compression_ratio": 1.2928348909657321
    },
    {
      "start": 1026.214,
      "end": 1028.314,
      "text": "重叠发言 自然停顿",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07008492634558552,
      "no_speech_prob": 1.457392366510657e-11,
      "compression_ratio": 1.2928348909657321
    },
    {
      "start": 1028.314,
      "end": 1030.194,
      "text": "都能正确标记身份",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07008492634558552,
      "no_speech_prob": 1.457392366510657e-11,
      "compression_ratio": 1.2928348909657321
    },
    {
      "start": 1030.194,
      "end": 1033.734,
      "text": "这对于会议记录 访谈整理 客服记录来说",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07008492634558552,
      "no_speech_prob": 1.457392366510657e-11,
      "compression_ratio": 1.2928348909657321
    },
    {
      "start": 1033.734,
      "end": 1034.674,
      "text": "价值很高",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07008492634558552,
      "no_speech_prob": 1.457392366510657e-11,
      "compression_ratio": 1.2928348909657321
    },
    {
      "start": 1034.674,
      "end": 1038.214,
      "text": "因为后续的摘要系统拿到的不只是一整段文字",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07008492634558552,
      "no_speech_prob": 1.457392366510657e-11,
      "compression_ratio": 1.2928348909657321
    },
    {
      "start": 1038.214,
      "end": 1040.894,
      "text": "还知道谁提了问题 谁做了承诺",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07008492634558552,
      "no_speech_prob": 1.457392366510657e-11,
      "compression_ratio": 1.2928348909657321
    },
    {
      "start": 1040.894,
      "end": 1042.234,
      "text": "谁反对哪个方案",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07008492634558552,
      "no_speech_prob": 1.457392366510657e-11,
      "compression_ratio": 1.2928348909657321
    },
    {
      "start": 1042.234,
      "end": 1043.974,
      "text": "其次是端点侦测能力",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07008492634558552,
      "no_speech_prob": 1.457392366510657e-11,
      "compression_ratio": 1.2928348909657321
    },
    {
      "start": 1043.974,
      "end": 1047.474,
      "text": "人类对话本来就不会每次讲完都按个结束钮",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.048076316793128925,
      "no_speech_prob": 1.3421238079236453e-11,
      "compression_ratio": 1.2920353982300885
    },
    {
      "start": 1047.474,
      "end": 1050.794,
      "text": "系统要从语气停顿和上下文判断",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.048076316793128925,
      "no_speech_prob": 1.3421238079236453e-11,
      "compression_ratio": 1.2920353982300885
    },
    {
      "start": 1050.794,
      "end": 1053.474,
      "text": "对方只是喘口气还是真的说完了",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.048076316793128925,
      "no_speech_prob": 1.3421238079236453e-11,
      "compression_ratio": 1.2920353982300885
    },
    {
      "start": 1053.474,
      "end": 1055.014,
      "text": "如果端点判断不准",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.048076316793128925,
      "no_speech_prob": 1.3421238079236453e-11,
      "compression_ratio": 1.2920353982300885
    },
    {
      "start": 1055.014,
      "end": 1056.694,
      "text": "要么用户都讲完半天了",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.048076316793128925,
      "no_speech_prob": 1.3421238079236453e-11,
      "compression_ratio": 1.2920353982300885
    },
    {
      "start": 1056.694,
      "end": 1057.774,
      "text": "助手还没反应",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.048076316793128925,
      "no_speech_prob": 1.3421238079236453e-11,
      "compression_ratio": 1.2920353982300885
    },
    {
      "start": 1057.774,
      "end": 1059.234,
      "text": "要么用户还没说完",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.048076316793128925,
      "no_speech_prob": 1.3421238079236453e-11,
      "compression_ratio": 1.2920353982300885
    },
    {
      "start": 1059.234,
      "end": 1061.494,
      "text": "助手就急着插嘴体验都很差",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.048076316793128925,
      "no_speech_prob": 1.3421238079236453e-11,
      "compression_ratio": 1.2920353982300885
    },
    {
      "start": 1061.494,
      "end": 1062.814,
      "text": "Muse Voice Transcribe",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.048076316793128925,
      "no_speech_prob": 1.3421238079236453e-11,
      "compression_ratio": 1.2920353982300885
    },
    {
      "start": 1062.814,
      "end": 1064.654,
      "text": "专门针对这点做了优化",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.048076316793128925,
      "no_speech_prob": 1.3421238079236453e-11,
      "compression_ratio": 1.2920353982300885
    },
    {
      "start": 1064.654,
      "end": 1067.514,
      "text": "能更准确判断语音的开始和结束",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.048076316793128925,
      "no_speech_prob": 1.3421238079236453e-11,
      "compression_ratio": 1.2920353982300885
    },
    {
      "start": 1067.514,
      "end": 1070.014,
      "text": "这也是代理人能自然对话的前提",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.048076316793128925,
      "no_speech_prob": 1.3421238079236453e-11,
      "compression_ratio": 1.2920353982300885
    },
    {
      "start": 1070.014,
      "end": 1071.894,
      "text": "第三个很实用的功能是",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.048076316793128925,
      "no_speech_prob": 1.3421238079236453e-11,
      "compression_ratio": 1.2920353982300885
    },
    {
      "start": 1071.894,
      "end": 1074.594,
      "text": "句内和句间的多语言自由切换",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07389446140564594,
      "no_speech_prob": 1.670223161165385e-11,
      "compression_ratio": 1.2889518413597734
    },
    {
      "start": 1074.594,
      "end": 1076.134,
      "text": "还有上下文偏制",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07389446140564594,
      "no_speech_prob": 1.670223161165385e-11,
      "compression_ratio": 1.2889518413597734
    },
    {
      "start": 1076.134,
      "end": 1078.374,
      "text": "现实中很多双语使用者",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07389446140564594,
      "no_speech_prob": 1.670223161165385e-11,
      "compression_ratio": 1.2889518413597734
    },
    {
      "start": 1078.374,
      "end": 1080.574,
      "text": "不会先切语言模式再讲话",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07389446140564594,
      "no_speech_prob": 1.670223161165385e-11,
      "compression_ratio": 1.2889518413597734
    },
    {
      "start": 1080.574,
      "end": 1082.994,
      "text": "中文句子里夹个英文专有名词",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07389446140564594,
      "no_speech_prob": 1.670223161165385e-11,
      "compression_ratio": 1.2889518413597734
    },
    {
      "start": 1082.994,
      "end": 1085.274,
      "text": "项目代号是家常便饭",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07389446140564594,
      "no_speech_prob": 1.670223161165385e-11,
      "compression_ratio": 1.2889518413597734
    },
    {
      "start": 1085.274,
      "end": 1086.954,
      "text": "旧的系统很容易把英文词",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07389446140564594,
      "no_speech_prob": 1.670223161165385e-11,
      "compression_ratio": 1.2889518413597734
    },
    {
      "start": 1086.954,
      "end": 1088.994,
      "text": "应转成发音相近的中文字",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07389446140564594,
      "no_speech_prob": 1.670223161165385e-11,
      "compression_ratio": 1.2889518413597734
    },
    {
      "start": 1088.994,
      "end": 1091.494,
      "text": "或者切换语言之后直接丢了半句话",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07389446140564594,
      "no_speech_prob": 1.670223161165385e-11,
      "compression_ratio": 1.2889518413597734
    },
    {
      "start": 1091.494,
      "end": 1092.994,
      "text": "Muse Voice Transcribe",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07389446140564594,
      "no_speech_prob": 1.670223161165385e-11,
      "compression_ratio": 1.2889518413597734
    },
    {
      "start": 1092.994,
      "end": 1095.294,
      "text": "支持70多种语言的自由切换",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07389446140564594,
      "no_speech_prob": 1.670223161165385e-11,
      "compression_ratio": 1.2889518413597734
    },
    {
      "start": 1095.294,
      "end": 1098.254,
      "text": "还可以让用户提前输入常用的人名",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07389446140564594,
      "no_speech_prob": 1.670223161165385e-11,
      "compression_ratio": 1.2889518413597734
    },
    {
      "start": 1098.254,
      "end": 1099.734,
      "text": "公司名专业术语",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07389446140564594,
      "no_speech_prob": 1.670223161165385e-11,
      "compression_ratio": 1.2889518413597734
    },
    {
      "start": 1099.734,
      "end": 1101.714,
      "text": "模型碰到相近发音的时候",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07389446140564594,
      "no_speech_prob": 1.670223161165385e-11,
      "compression_ratio": 1.2889518413597734
    },
    {
      "start": 1101.714,
      "end": 1104.994,
      "text": "会优先选择这些符合上下文的词",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06164922349439944,
      "no_speech_prob": 4.1133912248580984e-11,
      "compression_ratio": 1.2349570200573066
    },
    {
      "start": 1104.994,
      "end": 1107.654,
      "text": "比如医生可以提前录入药品名称",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06164922349439944,
      "no_speech_prob": 4.1133912248580984e-11,
      "compression_ratio": 1.2349570200573066
    },
    {
      "start": 1107.654,
      "end": 1110.914,
      "text": "城市设计师可以加入函数名和项目代号",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06164922349439944,
      "no_speech_prob": 4.1133912248580984e-11,
      "compression_ratio": 1.2349570200573066
    },
    {
      "start": 1110.914,
      "end": 1113.594,
      "text": "企业用户可以上传内部数语表",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06164922349439944,
      "no_speech_prob": 4.1133912248580984e-11,
      "compression_ratio": 1.2349570200573066
    },
    {
      "start": 1113.594,
      "end": 1116.014,
      "text": "这样转写的准确率会大幅提升",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06164922349439944,
      "no_speech_prob": 4.1133912248580984e-11,
      "compression_ratio": 1.2349570200573066
    },
    {
      "start": 1116.014,
      "end": 1118.154,
      "text": "甚至可以直接进入工作流程",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06164922349439944,
      "no_speech_prob": 4.1133912248580984e-11,
      "compression_ratio": 1.2349570200573066
    },
    {
      "start": 1118.154,
      "end": 1119.714,
      "text": "不用再人工校对一遍",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06164922349439944,
      "no_speech_prob": 4.1133912248580984e-11,
      "compression_ratio": 1.2349570200573066
    },
    {
      "start": 1119.714,
      "end": 1122.054,
      "text": "这才是语音转写真正的价值",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06164922349439944,
      "no_speech_prob": 4.1133912248580984e-11,
      "compression_ratio": 1.2349570200573066
    },
    {
      "start": 1122.054,
      "end": 1124.254,
      "text": "不是把声音变成文字就好",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06164922349439944,
      "no_speech_prob": 4.1133912248580984e-11,
      "compression_ratio": 1.2349570200573066
    },
    {
      "start": 1124.254,
      "end": 1126.234,
      "text": "而是变成能直接用的文字",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06164922349439944,
      "no_speech_prob": 4.1133912248580984e-11,
      "compression_ratio": 1.2349570200573066
    },
    {
      "start": 1126.234,
      "end": 1127.874,
      "text": "目前Muse Voice Transcribe",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06164922349439944,
      "no_speech_prob": 4.1133912248580984e-11,
      "compression_ratio": 1.2349570200573066
    },
    {
      "start": 1127.874,
      "end": 1129.614,
      "text": "已经接入Meta Model API",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06164922349439944,
      "no_speech_prob": 4.1133912248580984e-11,
      "compression_ratio": 1.2349570200573066
    },
    {
      "start": 1129.614,
      "end": 1132.774,
      "text": "还有Meta AI4 Mac Muse Code Mac版本",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.053896017332334775,
      "no_speech_prob": 1.8852567076899085e-11,
      "compression_ratio": 1.2779369627507164
    },
    {
      "start": 1132.774,
      "end": 1135.494,
      "text": "用户可以按住FN键直接语音输入",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.053896017332334775,
      "no_speech_prob": 1.8852567076899085e-11,
      "compression_ratio": 1.2779369627507164
    },
    {
      "start": 1135.494,
      "end": 1138.314,
      "text": "把转写结果送到不同的应用程式里",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.053896017332334775,
      "no_speech_prob": 1.8852567076899085e-11,
      "compression_ratio": 1.2779369627507164
    },
    {
      "start": 1138.314,
      "end": 1140.834,
      "text": "不过有个隐私细节一定要讲清楚",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.053896017332334775,
      "no_speech_prob": 1.8852567076899085e-11,
      "compression_ratio": 1.2779369627507164
    },
    {
      "start": 1140.834,
      "end": 1144.734,
      "text": "官方网页演示的时候说麦克风音频只会用来生成转写",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.053896017332334775,
      "no_speech_prob": 1.8852567076899085e-11,
      "compression_ratio": 1.2779369627507164
    },
    {
      "start": 1144.734,
      "end": 1145.614,
      "text": "不会被保存",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.053896017332334775,
      "no_speech_prob": 1.8852567076899085e-11,
      "compression_ratio": 1.2779369627507164
    },
    {
      "start": 1145.614,
      "end": 1148.394,
      "text": "但这个承诺只限于网页演示场景",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.053896017332334775,
      "no_speech_prob": 1.8852567076899085e-11,
      "compression_ratio": 1.2779369627507164
    },
    {
      "start": 1148.394,
      "end": 1150.254,
      "text": "不等于所有API呼叫",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.053896017332334775,
      "no_speech_prob": 1.8852567076899085e-11,
      "compression_ratio": 1.2779369627507164
    },
    {
      "start": 1150.254,
      "end": 1153.014,
      "text": "所有产品部署都不会保存音频",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.053896017332334775,
      "no_speech_prob": 1.8852567076899085e-11,
      "compression_ratio": 1.2779369627507164
    },
    {
      "start": 1153.014,
      "end": 1155.354,
      "text": "如果你是透过API上传录音",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.053896017332334775,
      "no_speech_prob": 1.8852567076899085e-11,
      "compression_ratio": 1.2779369627507164
    },
    {
      "start": 1155.354,
      "end": 1156.694,
      "text": "Meta会不会保存",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.053896017332334775,
      "no_speech_prob": 1.8852567076899085e-11,
      "compression_ratio": 1.2779369627507164
    },
    {
      "start": 1156.694,
      "end": 1157.554,
      "text": "保存多久",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.053896017332334775,
      "no_speech_prob": 1.8852567076899085e-11,
      "compression_ratio": 1.2779369627507164
    },
    {
      "start": 1157.554,
      "end": 1159.254,
      "text": "会不会用来训练模型",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.053896017332334775,
      "no_speech_prob": 1.8852567076899085e-11,
      "compression_ratio": 1.2779369627507164
    },
    {
      "start": 1159.254,
      "end": 1162.774,
      "text": "都要看你签署的服务条款和数据处理协议",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06083359501578591,
      "no_speech_prob": 1.3903270028314818e-11,
      "compression_ratio": 1.2028571428571428
    },
    {
      "start": 1162.774,
      "end": 1164.714,
      "text": "不能因为网页演示不保存",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06083359501578591,
      "no_speech_prob": 1.3903270028314818e-11,
      "compression_ratio": 1.2028571428571428
    },
    {
      "start": 1164.714,
      "end": 1168.074,
      "text": "就理所当然认为所有场景都不保存",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06083359501578591,
      "no_speech_prob": 1.3903270028314818e-11,
      "compression_ratio": 1.2028571428571428
    },
    {
      "start": 1168.074,
      "end": 1170.994,
      "text": "这在个资合规上是非常危险的认知",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06083359501578591,
      "no_speech_prob": 1.3903270028314818e-11,
      "compression_ratio": 1.2028571428571428
    },
    {
      "start": 1170.994,
      "end": 1173.534,
      "text": "讲到这里就必须好好聊一下",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06083359501578591,
      "no_speech_prob": 1.3903270028314818e-11,
      "compression_ratio": 1.2028571428571428
    },
    {
      "start": 1173.534,
      "end": 1175.034,
      "text": "隐私与安全的边界",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06083359501578591,
      "no_speech_prob": 1.3903270028314818e-11,
      "compression_ratio": 1.2028571428571428
    },
    {
      "start": 1175.034,
      "end": 1177.354,
      "text": "这也是很多人看AI产品的时候",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06083359501578591,
      "no_speech_prob": 1.3903270028314818e-11,
      "compression_ratio": 1.2028571428571428
    },
    {
      "start": 1177.354,
      "end": 1180.374,
      "text": "最容易忽略但绝对最重要的部分",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06083359501578591,
      "no_speech_prob": 1.3903270028314818e-11,
      "compression_ratio": 1.2028571428571428
    },
    {
      "start": 1180.374,
      "end": 1182.854,
      "text": "先讲Muse Spark 1.3的安全设计",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06083359501578591,
      "no_speech_prob": 1.3903270028314818e-11,
      "compression_ratio": 1.2028571428571428
    },
    {
      "start": 1182.854,
      "end": 1185.254,
      "text": "官方说这次提升了对提示注入",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06083359501578591,
      "no_speech_prob": 1.3903270028314818e-11,
      "compression_ratio": 1.2028571428571428
    },
    {
      "start": 1185.254,
      "end": 1187.294,
      "text": "和对抗输入的抵抗能力",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06083359501578591,
      "no_speech_prob": 1.3903270028314818e-11,
      "compression_ratio": 1.2028571428571428
    },
    {
      "start": 1187.294,
      "end": 1189.994,
      "text": "也更能识别哪些操作是不可逆的",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05357116548885853,
      "no_speech_prob": 1.4571958223408288e-11,
      "compression_ratio": 1.2023809523809523
    },
    {
      "start": 1189.994,
      "end": 1191.054,
      "text": "提示注入是什么",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05357116548885853,
      "no_speech_prob": 1.4571958223408288e-11,
      "compression_ratio": 1.2023809523809523
    },
    {
      "start": 1191.054,
      "end": 1193.254,
      "text": "就是攻击者在输入内容里",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05357116548885853,
      "no_speech_prob": 1.4571958223408288e-11,
      "compression_ratio": 1.2023809523809523
    },
    {
      "start": 1193.254,
      "end": 1195.274,
      "text": "嵌入看不见的恶意指令",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05357116548885853,
      "no_speech_prob": 1.4571958223408288e-11,
      "compression_ratio": 1.2023809523809523
    },
    {
      "start": 1195.274,
      "end": 1197.314,
      "text": "比如藏在PDF档案里",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05357116548885853,
      "no_speech_prob": 1.4571958223408288e-11,
      "compression_ratio": 1.2023809523809523
    },
    {
      "start": 1197.314,
      "end": 1198.114,
      "text": "网页里",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05357116548885853,
      "no_speech_prob": 1.4571958223408288e-11,
      "compression_ratio": 1.2023809523809523
    },
    {
      "start": 1198.114,
      "end": 1200.834,
      "text": "诱导AI代理人忽略原本的规则",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05357116548885853,
      "no_speech_prob": 1.4571958223408288e-11,
      "compression_ratio": 1.2023809523809523
    },
    {
      "start": 1200.834,
      "end": 1202.554,
      "text": "去转发机密档案",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05357116548885853,
      "no_speech_prob": 1.4571958223408288e-11,
      "compression_ratio": 1.2023809523809523
    },
    {
      "start": 1202.554,
      "end": 1203.974,
      "text": "修改设定甚至转账",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05357116548885853,
      "no_speech_prob": 1.4571958223408288e-11,
      "compression_ratio": 1.2023809523809523
    },
    {
      "start": 1203.974,
      "end": 1206.434,
      "text": "2025年就曾经出现过",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05357116548885853,
      "no_speech_prob": 1.4571958223408288e-11,
      "compression_ratio": 1.2023809523809523
    },
    {
      "start": 1206.434,
      "end": 1207.714,
      "text": "类似的攻击案例",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05357116548885853,
      "no_speech_prob": 1.4571958223408288e-11,
      "compression_ratio": 1.2023809523809523
    },
    {
      "start": 1207.714,
      "end": 1210.754,
      "text": "让不少企业对AI代理人望而却步",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05357116548885853,
      "no_speech_prob": 1.4571958223408288e-11,
      "compression_ratio": 1.2023809523809523
    },
    {
      "start": 1210.754,
      "end": 1213.774,
      "text": "Spark 1.3针对这类攻击做了专门训练",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05357116548885853,
      "no_speech_prob": 1.4571958223408288e-11,
      "compression_ratio": 1.2023809523809523
    },
    {
      "start": 1213.774,
      "end": 1215.594,
      "text": "但这不代表它完全免疫",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05357116548885853,
      "no_speech_prob": 1.4571958223408288e-11,
      "compression_ratio": 1.2023809523809523
    },
    {
      "start": 1215.594,
      "end": 1219.874,
      "text": "目前没有任何模型能100%防御所有提示注入攻击",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05953136503387609,
      "no_speech_prob": 1.463024840164806e-11,
      "compression_ratio": 1.284829721362229
    },
    {
      "start": 1219.874,
      "end": 1221.594,
      "text": "识别不可逆操作也是一样",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05953136503387609,
      "no_speech_prob": 1.463024840164806e-11,
      "compression_ratio": 1.284829721362229
    },
    {
      "start": 1221.594,
      "end": 1223.394,
      "text": "模型能标出删除档案",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05953136503387609,
      "no_speech_prob": 1.463024840164806e-11,
      "compression_ratio": 1.284829721362229
    },
    {
      "start": 1223.394,
      "end": 1224.574,
      "text": "发送批量邮件",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05953136503387609,
      "no_speech_prob": 1.463024840164806e-11,
      "compression_ratio": 1.284829721362229
    },
    {
      "start": 1224.574,
      "end": 1226.934,
      "text": "修改数据库这些高风险动作",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05953136503387609,
      "no_speech_prob": 1.463024840164806e-11,
      "compression_ratio": 1.284829721362229
    },
    {
      "start": 1226.934,
      "end": 1228.334,
      "text": "要求用户确认",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05953136503387609,
      "no_speech_prob": 1.463024840164806e-11,
      "compression_ratio": 1.284829721362229
    },
    {
      "start": 1228.334,
      "end": 1231.114,
      "text": "但光有模型层面的防护远远不够",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05953136503387609,
      "no_speech_prob": 1.463024840164806e-11,
      "compression_ratio": 1.284829721362229
    },
    {
      "start": 1231.114,
      "end": 1233.814,
      "text": "官方自己也强调实际部署的时候",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05953136503387609,
      "no_speech_prob": 1.463024840164806e-11,
      "compression_ratio": 1.284829721362229
    },
    {
      "start": 1233.814,
      "end": 1236.374,
      "text": "必须同时做好全线隔离",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05953136503387609,
      "no_speech_prob": 1.463024840164806e-11,
      "compression_ratio": 1.284829721362229
    },
    {
      "start": 1236.374,
      "end": 1238.454,
      "text": "白名单和操作日志",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05953136503387609,
      "no_speech_prob": 1.463024840164806e-11,
      "compression_ratio": 1.284829721362229
    },
    {
      "start": 1238.454,
      "end": 1241.514,
      "text": "全线隔离就是给代理人最小必要的全线",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05953136503387609,
      "no_speech_prob": 1.463024840164806e-11,
      "compression_ratio": 1.284829721362229
    },
    {
      "start": 1241.514,
      "end": 1243.754,
      "text": "比如负责整理表格的助理",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.05953136503387609,
      "no_speech_prob": 1.463024840164806e-11,
      "compression_ratio": 1.284829721362229
    },
    {
      "start": 1243.754,
      "end": 1246.534,
      "text": "就不该有删除整个资料夹的权限",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.04462584628853747,
      "no_speech_prob": 1.316543228602196e-11,
      "compression_ratio": 1.2244897959183674
    },
    {
      "start": 1246.534,
      "end": 1250.654,
      "text": "白名单就是只允许模型呼叫预先核准的",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.04462584628853747,
      "no_speech_prob": 1.316543228602196e-11,
      "compression_ratio": 1.2244897959183674
    },
    {
      "start": 1250.654,
      "end": 1254.834,
      "text": "工具和API不能让它随便访问系统里的任何东西",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.04462584628853747,
      "no_speech_prob": 1.316543228602196e-11,
      "compression_ratio": 1.2244897959183674
    },
    {
      "start": 1254.834,
      "end": 1258.954,
      "text": "操作日志就是把代理人所有执行动作都记录下来",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.04462584628853747,
      "no_speech_prob": 1.316543228602196e-11,
      "compression_ratio": 1.2244897959183674
    },
    {
      "start": 1258.954,
      "end": 1261.134,
      "text": "方便事后审计和回溯",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.04462584628853747,
      "no_speech_prob": 1.316543228602196e-11,
      "compression_ratio": 1.2244897959183674
    },
    {
      "start": 1261.134,
      "end": 1264.334,
      "text": "这三层防护都是资安领域的基本常识",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.04462584628853747,
      "no_speech_prob": 1.316543228602196e-11,
      "compression_ratio": 1.2244897959183674
    },
    {
      "start": 1264.334,
      "end": 1267.754,
      "text": "却也是很多AI产品最容易省略的部分",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.04462584628853747,
      "no_speech_prob": 1.316543228602196e-11,
      "compression_ratio": 1.2244897959183674
    },
    {
      "start": 1267.754,
      "end": 1269.614,
      "text": "Meta愿意主动强调这点",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.04462584628853747,
      "no_speech_prob": 1.316543228602196e-11,
      "compression_ratio": 1.2244897959183674
    },
    {
      "start": 1269.614,
      "end": 1273.234,
      "text": "说明它是真的在考虑企业落地的实际需求",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.04462584628853747,
      "no_speech_prob": 1.316543228602196e-11,
      "compression_ratio": 1.2244897959183674
    },
    {
      "start": 1273.234,
      "end": 1275.894,
      "text": "而不是只做展示用的技术demo",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08127796839153956,
      "no_speech_prob": 1.457025819440183e-11,
      "compression_ratio": 1.260233918128655
    },
    {
      "start": 1275.894,
      "end": 1276.974,
      "text": "再回到语音部分",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08127796839153956,
      "no_speech_prob": 1.457025819440183e-11,
      "compression_ratio": 1.260233918128655
    },
    {
      "start": 1276.974,
      "end": 1279.874,
      "text": "当语音转写加上代理人执行",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08127796839153956,
      "no_speech_prob": 1.457025819440183e-11,
      "compression_ratio": 1.260233918128655
    },
    {
      "start": 1279.874,
      "end": 1283.534,
      "text": "再搭配Meta旗下的ray减班智慧眼镜",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08127796839153956,
      "no_speech_prob": 1.457025819440183e-11,
      "compression_ratio": 1.260233918128655
    },
    {
      "start": 1283.534,
      "end": 1285.214,
      "text": "想象空间确实很大",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08127796839153956,
      "no_speech_prob": 1.457025819440183e-11,
      "compression_ratio": 1.260233918128655
    },
    {
      "start": 1285.214,
      "end": 1288.694,
      "text": "手机上的语音助手要等你点开APP才能用",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08127796839153956,
      "no_speech_prob": 1.457025819440183e-11,
      "compression_ratio": 1.260233918128655
    },
    {
      "start": 1288.694,
      "end": 1290.854,
      "text": "但眼镜可以一直戴在脸上",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08127796839153956,
      "no_speech_prob": 1.457025819440183e-11,
      "compression_ratio": 1.260233918128655
    },
    {
      "start": 1290.854,
      "end": 1292.794,
      "text": "随时处在声音环境里",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08127796839153956,
      "no_speech_prob": 1.457025819440183e-11,
      "compression_ratio": 1.260233918128655
    },
    {
      "start": 1292.794,
      "end": 1294.574,
      "text": "如果他能听懂谁在讲话",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08127796839153956,
      "no_speech_prob": 1.457025819440183e-11,
      "compression_ratio": 1.260233918128655
    },
    {
      "start": 1294.574,
      "end": 1295.814,
      "text": "什么时候该接话",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08127796839153956,
      "no_speech_prob": 1.457025819440183e-11,
      "compression_ratio": 1.260233918128655
    },
    {
      "start": 1295.814,
      "end": 1297.134,
      "text": "什么时候不该插嘴",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08127796839153956,
      "no_speech_prob": 1.457025819440183e-11,
      "compression_ratio": 1.260233918128655
    },
    {
      "start": 1297.134,
      "end": 1299.954,
      "text": "就能在会议后自动整理每个人的行动项",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08127796839153956,
      "no_speech_prob": 1.457025819440183e-11,
      "compression_ratio": 1.260233918128655
    },
    {
      "start": 1299.954,
      "end": 1302.014,
      "text": "访谈的时候自动标记重点",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08127796839153956,
      "no_speech_prob": 1.457025819440183e-11,
      "compression_ratio": 1.260233918128655
    },
    {
      "start": 1302.014,
      "end": 1306.134,
      "text": "甚至日常聊天里提到的代办事项都能自动记下来",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06315357561056324,
      "no_speech_prob": 1.3426587966436365e-11,
      "compression_ratio": 1.2647058823529411
    },
    {
      "start": 1306.134,
      "end": 1308.634,
      "text": "但这也带来了更尖锐的隐私问题",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06315357561056324,
      "no_speech_prob": 1.3426587966436365e-11,
      "compression_ratio": 1.2647058823529411
    },
    {
      "start": 1308.634,
      "end": 1310.974,
      "text": "眼镜上的麦克风什么时候在听",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06315357561056324,
      "no_speech_prob": 1.3426587966436365e-11,
      "compression_ratio": 1.2647058823529411
    },
    {
      "start": 1310.974,
      "end": 1312.874,
      "text": "听到的内容传到哪里去了",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06315357561056324,
      "no_speech_prob": 1.3426587966436365e-11,
      "compression_ratio": 1.2647058823529411
    },
    {
      "start": 1312.874,
      "end": 1314.914,
      "text": "用户有没有明确的控制权",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06315357561056324,
      "no_speech_prob": 1.3426587966436365e-11,
      "compression_ratio": 1.2647058823529411
    },
    {
      "start": 1314.914,
      "end": 1317.234,
      "text": "旁边的人知不知道自己正在被录音",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06315357561056324,
      "no_speech_prob": 1.3426587966436365e-11,
      "compression_ratio": 1.2647058823529411
    },
    {
      "start": 1317.234,
      "end": 1320.234,
      "text": "这些问题目前都还没有完全公开的答案",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06315357561056324,
      "no_speech_prob": 1.3426587966436365e-11,
      "compression_ratio": 1.2647058823529411
    },
    {
      "start": 1320.234,
      "end": 1324.274,
      "text": "不过我们可以从Meta的双轨策略里推出一个很合理的假设",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06315357561056324,
      "no_speech_prob": 1.3426587966436365e-11,
      "compression_ratio": 1.2647058823529411
    },
    {
      "start": 1324.274,
      "end": 1326.954,
      "text": "未来轻量级的语音转写",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06315357561056324,
      "no_speech_prob": 1.3426587966436365e-11,
      "compression_ratio": 1.2647058823529411
    },
    {
      "start": 1326.954,
      "end": 1328.534,
      "text": "简单的指令理解",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06315357561056324,
      "no_speech_prob": 1.3426587966436365e-11,
      "compression_ratio": 1.2647058823529411
    },
    {
      "start": 1328.534,
      "end": 1332.294,
      "text": "很可能会直接跑在眼镜本地的端侧芯片上",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.045785757211538464,
      "no_speech_prob": 1.4972665468571122e-11,
      "compression_ratio": 1.2476780185758514
    },
    {
      "start": 1332.294,
      "end": 1333.534,
      "text": "不需要传到云端",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.045785757211538464,
      "no_speech_prob": 1.4972665468571122e-11,
      "compression_ratio": 1.2476780185758514
    },
    {
      "start": 1333.534,
      "end": 1335.294,
      "text": "敏感数据不会离开设备",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.045785757211538464,
      "no_speech_prob": 1.4972665468571122e-11,
      "compression_ratio": 1.2476780185758514
    },
    {
      "start": 1335.294,
      "end": 1337.034,
      "text": "大幅降低隐私风险",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.045785757211538464,
      "no_speech_prob": 1.4972665468571122e-11,
      "compression_ratio": 1.2476780185758514
    },
    {
      "start": 1337.034,
      "end": 1338.774,
      "text": "而复杂的常任务代理",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.045785757211538464,
      "no_speech_prob": 1.4972665468571122e-11,
      "compression_ratio": 1.2476780185758514
    },
    {
      "start": 1338.774,
      "end": 1339.934,
      "text": "多工具协作",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.045785757211538464,
      "no_speech_prob": 1.4972665468571122e-11,
      "compression_ratio": 1.2476780185758514
    },
    {
      "start": 1339.934,
      "end": 1342.694,
      "text": "再交给云端的Spark模型处理",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.045785757211538464,
      "no_speech_prob": 1.4972665468571122e-11,
      "compression_ratio": 1.2476780185758514
    },
    {
      "start": 1342.694,
      "end": 1345.634,
      "text": "刚好Meta有Glimmer这条本地开源路线",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.045785757211538464,
      "no_speech_prob": 1.4972665468571122e-11,
      "compression_ratio": 1.2476780185758514
    },
    {
      "start": 1345.634,
      "end": 1347.694,
      "text": "专门做轻量化端侧部署",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.045785757211538464,
      "no_speech_prob": 1.4972665468571122e-11,
      "compression_ratio": 1.2476780185758514
    },
    {
      "start": 1347.694,
      "end": 1350.634,
      "text": "刚好可以跟眼镜这类终端装置搭配",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.045785757211538464,
      "no_speech_prob": 1.4972665468571122e-11,
      "compression_ratio": 1.2476780185758514
    },
    {
      "start": 1350.634,
      "end": 1352.334,
      "text": "云端负责复杂思考",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.045785757211538464,
      "no_speech_prob": 1.4972665468571122e-11,
      "compression_ratio": 1.2476780185758514
    },
    {
      "start": 1352.334,
      "end": 1354.034,
      "text": "本地负责实时感知",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.045785757211538464,
      "no_speech_prob": 1.4972665468571122e-11,
      "compression_ratio": 1.2476780185758514
    },
    {
      "start": 1354.034,
      "end": 1356.274,
      "text": "既保证体验又兼顾隐私",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.045785757211538464,
      "no_speech_prob": 1.4972665468571122e-11,
      "compression_ratio": 1.2476780185758514
    },
    {
      "start": 1356.274,
      "end": 1360.194,
      "text": "这很可能就是Meta下一代AI入口的完整架构",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06715974142385084,
      "no_speech_prob": 1.3155427268374265e-11,
      "compression_ratio": 1.2332361516034986
    },
    {
      "start": 1360.194,
      "end": 1362.834,
      "text": "当然这只是基于现有产品线的推测",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06715974142385084,
      "no_speech_prob": 1.3155427268374265e-11,
      "compression_ratio": 1.2332361516034986
    },
    {
      "start": 1362.834,
      "end": 1364.754,
      "text": "官方还没有证实这个方向",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06715974142385084,
      "no_speech_prob": 1.3155427268374265e-11,
      "compression_ratio": 1.2332361516034986
    },
    {
      "start": 1364.754,
      "end": 1366.794,
      "text": "讲完了技术亮点和想象空间",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06715974142385084,
      "no_speech_prob": 1.3155427268374265e-11,
      "compression_ratio": 1.2332361516034986
    },
    {
      "start": 1366.794,
      "end": 1368.234,
      "text": "我们也得冷静下来",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06715974142385084,
      "no_speech_prob": 1.3155427268374265e-11,
      "compression_ratio": 1.2332361516034986
    },
    {
      "start": 1368.234,
      "end": 1371.154,
      "text": "聊聊这次发布的口径边界和实际风险",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06715974142385084,
      "no_speech_prob": 1.3155427268374265e-11,
      "compression_ratio": 1.2332361516034986
    },
    {
      "start": 1371.154,
      "end": 1374.114,
      "text": "首先Max版本目前只是有线预览",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06715974142385084,
      "no_speech_prob": 1.3155427268374265e-11,
      "compression_ratio": 1.2332361516034986
    },
    {
      "start": 1374.114,
      "end": 1376.514,
      "text": "普通开发者能用的只有Shi版本",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06715974142385084,
      "no_speech_prob": 1.3155427268374265e-11,
      "compression_ratio": 1.2332361516034986
    },
    {
      "start": 1376.514,
      "end": 1378.834,
      "text": "网上所有关于Max的评测数据",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06715974142385084,
      "no_speech_prob": 1.3155427268374265e-11,
      "compression_ratio": 1.2332361516034986
    },
    {
      "start": 1378.834,
      "end": 1381.034,
      "text": "都来自少数拿到权限的机构",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06715974142385084,
      "no_speech_prob": 1.3155427268374265e-11,
      "compression_ratio": 1.2332361516034986
    },
    {
      "start": 1381.034,
      "end": 1383.074,
      "text": "一般用户现在根本用不到",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06715974142385084,
      "no_speech_prob": 1.3155427268374265e-11,
      "compression_ratio": 1.2332361516034986
    },
    {
      "start": 1383.074,
      "end": 1387.054,
      "text": "所以62分那个数字只能当成技术上限的参考",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.0769476431387442,
      "no_speech_prob": 9.557438174212507e-12,
      "compression_ratio": 1.2298387096774193
    },
    {
      "start": 1387.054,
      "end": 1390.094,
      "text": "不能作为采购或者使用决策的依据",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.0769476431387442,
      "no_speech_prob": 9.557438174212507e-12,
      "compression_ratio": 1.2298387096774193
    },
    {
      "start": 1390.094,
      "end": 1392.414,
      "text": "其次所有的基准测试数据",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.0769476431387442,
      "no_speech_prob": 9.557438174212507e-12,
      "compression_ratio": 1.2298387096774193
    },
    {
      "start": 1392.414,
      "end": 1395.414,
      "text": "都是在标准化的干净环境里跑出来的",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.0769476431387442,
      "no_speech_prob": 9.557438174212507e-12,
      "compression_ratio": 1.2298387096774193
    },
    {
      "start": 1395.414,
      "end": 1399.494,
      "text": "真实世界的场景只会更复杂更混乱",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.0769476431387442,
      "no_speech_prob": 9.557438174212507e-12,
      "compression_ratio": 1.2298387096774193
    },
    {
      "start": 1399.494,
      "end": 1402.494,
      "text": "比如银行代理测试里的数据都是结构化的",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.0769476431387442,
      "no_speech_prob": 9.557438174212507e-12,
      "compression_ratio": 1.2298387096774193
    },
    {
      "start": 1402.494,
      "end": 1404.154,
      "text": "格式统一全线清晰",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.0769476431387442,
      "no_speech_prob": 9.557438174212507e-12,
      "compression_ratio": 1.2298387096774193
    },
    {
      "start": 1405.087,
      "end": 1411.527,
      "text": "但真实的企业里,数据可能分散在十几个老旧系统里,格式乱七八糟、权限层层卡关。",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07632808362023305,
      "no_speech_prob": 8.692954342470749e-12,
      "compression_ratio": 1.3642384105960266
    },
    {
      "start": 1411.527,
      "end": 1417.207,
      "text": "你在测试里拿到47%的通过率,到了真实环境可能连30%都不到。",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07632808362023305,
      "no_speech_prob": 8.692954342470749e-12,
      "compression_ratio": 1.3642384105960266
    },
    {
      "start": 1417.207,
      "end": 1423.787,
      "text": "这不是Meta的问题 是所有AI模型评测的普遍现象 跑分永远不等于真实生产力",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.053156744350086556,
      "no_speech_prob": 1.3475446453137252e-11,
      "compression_ratio": 1.2521246458923512
    },
    {
      "start": 1423.787,
      "end": 1426.327,
      "text": "第三 关于开源的期待也要放低",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.053156744350086556,
      "no_speech_prob": 1.3475446453137252e-11,
      "compression_ratio": 1.2521246458923512
    },
    {
      "start": 1426.327,
      "end": 1434.287,
      "text": "Spark 1.3目前是完全避原的 官方只说未来有开放权重的规划 但没有时间表 也没有说授权条款",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.053156744350086556,
      "no_speech_prob": 1.3475446453137252e-11,
      "compression_ratio": 1.2521246458923512
    },
    {
      "start": 1434.287,
      "end": 1441.247,
      "text": "参考Meta之前Lama系列的节奏 通常是新一代出来之后才会把上一代的权重逐步开放",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.053156744350086556,
      "no_speech_prob": 1.3475446453137252e-11,
      "compression_ratio": 1.2521246458923512
    },
    {
      "start": 1441.247,
      "end": 1446.807,
      "text": "而且早期都是非商业授权 要等到更后面才会用更宽松的授权",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.053156744350086556,
      "no_speech_prob": 1.3475446453137252e-11,
      "compression_ratio": 1.2521246458923512
    },
    {
      "start": 1446.807,
      "end": 1450.007,
      "text": "所以不要期待短期内就能在自己的电脑上",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06551507167044403,
      "no_speech_prob": 1.2034759473700252e-11,
      "compression_ratio": 1.1727272727272726
    },
    {
      "start": 1450.007,
      "end": 1452.987,
      "text": "跑Spark 1.3核心的代理人能力",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06551507167044403,
      "no_speech_prob": 1.2034759473700252e-11,
      "compression_ratio": 1.1727272727272726
    },
    {
      "start": 1452.987,
      "end": 1456.127,
      "text": "Meta肯定会先握在手里当商业护城河",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06551507167044403,
      "no_speech_prob": 1.2034759473700252e-11,
      "compression_ratio": 1.1727272727272726
    },
    {
      "start": 1456.127,
      "end": 1458.087,
      "text": "最后回到产业竞争的层面",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06551507167044403,
      "no_speech_prob": 1.2034759473700252e-11,
      "compression_ratio": 1.1727272727272726
    },
    {
      "start": 1458.087,
      "end": 1460.447,
      "text": "Meta这次的更新确实扎实",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06551507167044403,
      "no_speech_prob": 1.2034759473700252e-11,
      "compression_ratio": 1.1727272727272726
    },
    {
      "start": 1460.447,
      "end": 1465.027,
      "text": "在代理人可靠性和流逝语音转写上都往前走了一步",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06551507167044403,
      "no_speech_prob": 1.2034759473700252e-11,
      "compression_ratio": 1.1727272727272726
    },
    {
      "start": 1465.027,
      "end": 1467.327,
      "text": "成本控制也做得很有竞争力",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06551507167044403,
      "no_speech_prob": 1.2034759473700252e-11,
      "compression_ratio": 1.1727272727272726
    },
    {
      "start": 1467.327,
      "end": 1469.227,
      "text": "但要说已经赢了还太早",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06551507167044403,
      "no_speech_prob": 1.2034759473700252e-11,
      "compression_ratio": 1.1727272727272726
    },
    {
      "start": 1469.227,
      "end": 1473.127,
      "text": "OpenEye的GPT系列代理人能力同样在快速迭代",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06551507167044403,
      "no_speech_prob": 1.2034759473700252e-11,
      "compression_ratio": 1.1727272727272726
    },
    {
      "start": 1473.127,
      "end": 1477.847,
      "text": "谷歌的Gemini在多模态和端侧整合上有自己的优势",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.056126269301927156,
      "no_speech_prob": 1.1483192868810832e-11,
      "compression_ratio": 1.1858407079646018
    },
    {
      "start": 1477.847,
      "end": 1482.527,
      "text": "Anthropic在长上下文和安全合规上依然领先",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.056126269301927156,
      "no_speech_prob": 1.1483192868810832e-11,
      "compression_ratio": 1.1858407079646018
    },
    {
      "start": 1482.527,
      "end": 1486.927,
      "text": "还有Eleven Labs这类专业语音公司也在往代理人方向走",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.056126269301927156,
      "no_speech_prob": 1.1483192868810832e-11,
      "compression_ratio": 1.1858407079646018
    },
    {
      "start": 1486.927,
      "end": 1492.547,
      "text": "Meta现在的领先更像是某几个细分场景的小幅度领先",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.056126269301927156,
      "no_speech_prob": 1.1483192868810832e-11,
      "compression_ratio": 1.1858407079646018
    },
    {
      "start": 1492.547,
      "end": 1495.207,
      "text": "还远没有到拉开差距的程度",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.056126269301927156,
      "no_speech_prob": 1.1483192868810832e-11,
      "compression_ratio": 1.1858407079646018
    },
    {
      "start": 1495.207,
      "end": 1498.527,
      "text": "而且真正的竞争从来都不是比模型本身",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.056126269301927156,
      "no_speech_prob": 1.1483192868810832e-11,
      "compression_ratio": 1.1858407079646018
    },
    {
      "start": 1498.527,
      "end": 1502.807,
      "text": "而是比谁能先把技术变成用户愿意天天用的产品",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.056126269301927156,
      "no_speech_prob": 1.1483192868810832e-11,
      "compression_ratio": 1.1858407079646018
    },
    {
      "start": 1502.807,
      "end": 1504.927,
      "text": "谁能解决好隐私安全",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07239272643108757,
      "no_speech_prob": 1.5740009989828607e-11,
      "compression_ratio": 1.304093567251462
    },
    {
      "start": 1504.927,
      "end": 1506.967,
      "text": "权限这些麻烦的落地问题",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07239272643108757,
      "no_speech_prob": 1.5740009989828607e-11,
      "compression_ratio": 1.304093567251462
    },
    {
      "start": 1506.967,
      "end": 1508.247,
      "text": "模型再聪明",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07239272643108757,
      "no_speech_prob": 1.5740009989828607e-11,
      "compression_ratio": 1.304093567251462
    },
    {
      "start": 1508.247,
      "end": 1510.887,
      "text": "如果用户不敢把重要的工作交给他",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07239272643108757,
      "no_speech_prob": 1.5740009989828607e-11,
      "compression_ratio": 1.304093567251462
    },
    {
      "start": 1510.887,
      "end": 1512.587,
      "text": "那也只是个高级玩具",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07239272643108757,
      "no_speech_prob": 1.5740009989828607e-11,
      "compression_ratio": 1.304093567251462
    },
    {
      "start": 1512.587,
      "end": 1513.167,
      "text": "讲到这里",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07239272643108757,
      "no_speech_prob": 1.5740009989828607e-11,
      "compression_ratio": 1.304093567251462
    },
    {
      "start": 1513.167,
      "end": 1516.007,
      "text": "我们把这次Meta两款产品的技术细节",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07239272643108757,
      "no_speech_prob": 1.5740009989828607e-11,
      "compression_ratio": 1.304093567251462
    },
    {
      "start": 1516.007,
      "end": 1516.947,
      "text": "评测数据",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07239272643108757,
      "no_speech_prob": 1.5740009989828607e-11,
      "compression_ratio": 1.304093567251462
    },
    {
      "start": 1516.947,
      "end": 1518.107,
      "text": "产品线定位",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07239272643108757,
      "no_speech_prob": 1.5740009989828607e-11,
      "compression_ratio": 1.304093567251462
    },
    {
      "start": 1518.107,
      "end": 1519.027,
      "text": "隐私风险",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07239272643108757,
      "no_speech_prob": 1.5740009989828607e-11,
      "compression_ratio": 1.304093567251462
    },
    {
      "start": 1519.027,
      "end": 1521.187,
      "text": "还有产业格局全部拆完了",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07239272643108757,
      "no_speech_prob": 1.5740009989828607e-11,
      "compression_ratio": 1.304093567251462
    },
    {
      "start": 1521.187,
      "end": 1523.667,
      "text": "你会发现每一个亮眼的数据背后",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07239272643108757,
      "no_speech_prob": 1.5740009989828607e-11,
      "compression_ratio": 1.304093567251462
    },
    {
      "start": 1523.667,
      "end": 1525.307,
      "text": "都有一个限定条件",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07239272643108757,
      "no_speech_prob": 1.5740009989828607e-11,
      "compression_ratio": 1.304093567251462
    },
    {
      "start": 1525.307,
      "end": 1527.067,
      "text": "每一项技术突破的旁边",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07239272643108757,
      "no_speech_prob": 1.5740009989828607e-11,
      "compression_ratio": 1.304093567251462
    },
    {
      "start": 1527.067,
      "end": 1529.227,
      "text": "都站着一个带解决的问题",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07239272643108757,
      "no_speech_prob": 1.5740009989828607e-11,
      "compression_ratio": 1.304093567251462
    },
    {
      "start": 1529.227,
      "end": 1530.947,
      "text": "这不是说Meta做得不好",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07239272643108757,
      "no_speech_prob": 1.5740009989828607e-11,
      "compression_ratio": 1.304093567251462
    },
    {
      "start": 1530.947,
      "end": 1532.047,
      "text": "恰恰相反",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07239272643108757,
      "no_speech_prob": 1.5740009989828607e-11,
      "compression_ratio": 1.304093567251462
    },
    {
      "start": 1532.047,
      "end": 1534.567,
      "text": "只有当你看清了所有的边界和限制",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06817786077435097,
      "no_speech_prob": 1.5449882692641914e-11,
      "compression_ratio": 1.16
    },
    {
      "start": 1534.567,
      "end": 1536.587,
      "text": "才能真正评估它的价值",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06817786077435097,
      "no_speech_prob": 1.5449882692641914e-11,
      "compression_ratio": 1.16
    },
    {
      "start": 1536.587,
      "end": 1541.227,
      "text": "Muse Spark 1.3确实把常任务代理人的可靠性往前推了一步",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06817786077435097,
      "no_speech_prob": 1.5449882692641914e-11,
      "compression_ratio": 1.16
    },
    {
      "start": 1541.227,
      "end": 1543.027,
      "text": "成本优势也很明显",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06817786077435097,
      "no_speech_prob": 1.5449882692641914e-11,
      "compression_ratio": 1.16
    },
    {
      "start": 1543.027,
      "end": 1546.167,
      "text": "Muse Voice Transcribe在英语流逝转写上",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06817786077435097,
      "no_speech_prob": 1.5449882692641914e-11,
      "compression_ratio": 1.16
    },
    {
      "start": 1546.167,
      "end": 1548.187,
      "text": "做到了业界顶尖",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06817786077435097,
      "no_speech_prob": 1.5449882692641914e-11,
      "compression_ratio": 1.16
    },
    {
      "start": 1548.187,
      "end": 1550.827,
      "text": "准确率和延迟的平衡做得可圈可点",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06817786077435097,
      "no_speech_prob": 1.5449882692641914e-11,
      "compression_ratio": 1.16
    },
    {
      "start": 1550.827,
      "end": 1553.427,
      "text": "但它们都还只是阶段性的进展",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06817786077435097,
      "no_speech_prob": 1.5449882692641914e-11,
      "compression_ratio": 1.16
    },
    {
      "start": 1553.427,
      "end": 1556.067,
      "text": "离大家想象中什么都能做的AI助理",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06817786077435097,
      "no_speech_prob": 1.5449882692641914e-11,
      "compression_ratio": 1.16
    },
    {
      "start": 1556.067,
      "end": 1557.647,
      "text": "还有非常远的距离",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06817786077435097,
      "no_speech_prob": 1.5449882692641914e-11,
      "compression_ratio": 1.16
    },
    {
      "start": 1557.647,
      "end": 1559.707,
      "text": "2026年的AI产业",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06817786077435097,
      "no_speech_prob": 1.5449882692641914e-11,
      "compression_ratio": 1.16
    },
    {
      "start": 1559.707,
      "end": 1562.607,
      "text": "其实正在经历一个很重要的转向",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.05707484797427529,
      "no_speech_prob": 1.3311634780577286e-11,
      "compression_ratio": 1.3322683706070289
    },
    {
      "start": 1562.607,
      "end": 1565.447,
      "text": "前几年大家都在比谁的模型更大",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.05707484797427529,
      "no_speech_prob": 1.3311634780577286e-11,
      "compression_ratio": 1.3322683706070289
    },
    {
      "start": 1565.447,
      "end": 1566.747,
      "text": "谁的跑分更高",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.05707484797427529,
      "no_speech_prob": 1.3311634780577286e-11,
      "compression_ratio": 1.3322683706070289
    },
    {
      "start": 1566.747,
      "end": 1568.627,
      "text": "谁的牛皮吹得更响",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.05707484797427529,
      "no_speech_prob": 1.3311634780577286e-11,
      "compression_ratio": 1.3322683706070289
    },
    {
      "start": 1568.627,
      "end": 1572.087,
      "text": "但从今年开始竞争慢慢落到了实处",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.05707484797427529,
      "no_speech_prob": 1.3311634780577286e-11,
      "compression_ratio": 1.3322683706070289
    },
    {
      "start": 1572.087,
      "end": 1574.167,
      "text": "比的是谁能把长任务做稳",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.05707484797427529,
      "no_speech_prob": 1.3311634780577286e-11,
      "compression_ratio": 1.3322683706070289
    },
    {
      "start": 1574.167,
      "end": 1575.927,
      "text": "谁能把真实声音听懂",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.05707484797427529,
      "no_speech_prob": 1.3311634780577286e-11,
      "compression_ratio": 1.3322683706070289
    },
    {
      "start": 1575.927,
      "end": 1577.727,
      "text": "谁能让企业敢放心用",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.05707484797427529,
      "no_speech_prob": 1.3311634780577286e-11,
      "compression_ratio": 1.3322683706070289
    },
    {
      "start": 1577.727,
      "end": 1580.227,
      "text": "谁能把隐私控制在合理范围",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.05707484797427529,
      "no_speech_prob": 1.3311634780577286e-11,
      "compression_ratio": 1.3322683706070289
    },
    {
      "start": 1580.227,
      "end": 1582.847,
      "text": "那些看起来不怎么耀眼的细节优化",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.05707484797427529,
      "no_speech_prob": 1.3311634780577286e-11,
      "compression_ratio": 1.3322683706070289
    },
    {
      "start": 1582.847,
      "end": 1587.407,
      "text": "反而决定了产品能不能真正走进日常生活和工作流",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.05707484797427529,
      "no_speech_prob": 1.3311634780577286e-11,
      "compression_ratio": 1.3322683706070289
    },
    {
      "start": 1587.407,
      "end": 1588.267,
      "text": "就像造车",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.05707484797427529,
      "no_speech_prob": 1.3311634780577286e-11,
      "compression_ratio": 1.3322683706070289
    },
    {
      "start": 1588.267,
      "end": 1596.447,
      "text": "早期大家都比马力 比急速 等到市场成熟了 比的是隔音 是油耗 是故障率 是乘坐舒适度",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06071184888298129,
      "no_speech_prob": 1.650789401008712e-11,
      "compression_ratio": 1.2394822006472492
    },
    {
      "start": 1596.447,
      "end": 1599.987,
      "text": "这些才是决定你愿不愿意每天开的关",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06071184888298129,
      "no_speech_prob": 1.650789401008712e-11,
      "compression_ratio": 1.2394822006472492
    },
    {
      "start": 1599.987,
      "end": 1605.347,
      "text": "见如果你喜欢这种不吹不黑 把数据和限制都讲清楚的深度科技解析",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06071184888298129,
      "no_speech_prob": 1.650789401008712e-11,
      "compression_ratio": 1.2394822006472492
    },
    {
      "start": 1605.347,
      "end": 1608.947,
      "text": "欢迎订阅这个频道 并且打开通知小铃铛",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06071184888298129,
      "no_speech_prob": 1.650789401008712e-11,
      "compression_ratio": 1.2394822006472492
    },
    {
      "start": 1608.947,
      "end": 1614.787,
      "text": "我们会持续追踪AI代理人 语音交互 端策AI这些领域的最新进展",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.06071184888298129,
      "no_speech_prob": 1.650789401008712e-11,
      "compression_ratio": 1.2394822006472492
    },
    {
      "start": 1614.787,
      "end": 1617.307,
      "text": "帮你过滤掉夸大的营销话术",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07295373507908412,
      "no_speech_prob": 1.346761244191974e-11,
      "compression_ratio": 1.1407035175879396
    },
    {
      "start": 1617.307,
      "end": 1619.727,
      "text": "留下真正值得关注的技术变化",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07295373507908412,
      "no_speech_prob": 1.346761244191974e-11,
      "compression_ratio": 1.1407035175879396
    },
    {
      "start": 1619.727,
      "end": 1621.327,
      "text": "也欢迎在留言区告诉我",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07295373507908412,
      "no_speech_prob": 1.346761244191974e-11,
      "compression_ratio": 1.1407035175879396
    },
    {
      "start": 1621.327,
      "end": 1624.287,
      "text": "你最期待AI帮你完成什么样的任务",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07295373507908412,
      "no_speech_prob": 1.346761244191974e-11,
      "compression_ratio": 1.1407035175879396
    },
    {
      "start": 1624.287,
      "end": 1625.927,
      "text": "是整理会议记录",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07295373507908412,
      "no_speech_prob": 1.346761244191974e-11,
      "compression_ratio": 1.1407035175879396
    },
    {
      "start": 1625.927,
      "end": 1627.887,
      "text": "还是处理繁琐的文书工作",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07295373507908412,
      "no_speech_prob": 1.346761244191974e-11,
      "compression_ratio": 1.1407035175879396
    },
    {
      "start": 1627.887,
      "end": 1629.647,
      "text": "我们下期影片再见",
      "chunk": 3,
      "language": "zh",
      "avg_logprob": -0.07295373507908412,
      "no_speech_prob": 1.346761244191974e-11,
      "compression_ratio": 1.1407035175879396
    }
  ],
  "segmented_transcription": {
    "status": "SEGMENTED_TRANSCRIPT_OK",
    "created_at": "2026-09-14T19:32:51",
    "manifest": "/Users/sagawa/AI_WORK/video_notes/url/20260914_193149__Meta_真的超車了_實測_Muse_Spark_1.3_與_Voice_Transcribe_真能幹活還是行銷話術/segmented_work/chunk_manifest.json",
    "audio": "/Users/sagawa/AI_WORK/video_notes/url/20260914_193149__Meta_真的超車了_實測_Muse_Spark_1.3_與_Voice_Transcribe_真能幹活還是行銷話術/audio_16k_mono.wav",
    "audio_sha256": "f32c0765cd932c2b00635b89730034ab49cac40d596830f7279a73e810458731",
    "audio_duration": 1629.855063,
    "model": "/Users/sagawa/AI_WORK/cache/huggingface/hub/models--mlx-community--whisper-large-v3-turbo/snapshots/a4aaeec0636e6fef84abdcbe3544cb2bf7e9f6fb",
    "language_hint": "zh",
    "condition_on_previous_text": false,
    "chunk_count": 4,
    "resumed_chunks": 0,
    "merged_segment_count": 655,
    "cleaned_segment_count": 655,
    "removed_hallucination_segment_count": 0,
    "final_segment_end": 1629.647,
    "global_repeat_run": 1,
    "global_repeat_text": "你有没有过这种很无力的经验?",
    "tail_repeat_run": 1,
    "tail_repeat_text": "做到了业界顶尖",
    "hallucination_region_count": 0,
    "hallucination_regions": [],
    "qc_ok": true,
    "qc_errors": []
  }
}