{
  "text": "9月15日迪奥哥·阿尔梅达在X上发了一条帖子。第一句话就很抓人。他说,参与创造XGP之后,我一直在问自己,为什么这些已经超过人类水平的聊天模型,没有带来AGI呢?这个问题从他嘴里说出来,确实有点特别。阿尔梅达是Intract GPT的作者之一,也参与过GPT4的RLHF和指令遵循的相关工作。\n简单来说\n今天我们已经习以为常的这件事\n就是你对AI说一句话\n它能够理解你想干什么\n然后按照你的要求回答\n它就是这个早期参与\n把这条路线做出来的人之一\n不过参与创造ChatGPT\n是它自己的表述\n并不是OpenAI给它的\nChatGPT联合发明人头衔\n但是ChatGPT出现以后\n它反而开始怀疑一件事了\n就是模型已经越来越会说话了\n那为什么我们期待中的大规模自动化\n却没有以同样的速度发生呢\n于是过去两年\n阿尔梅达一直在隐身做另一件东西\n直到这几天\n他和自己的公司TypeSafe AI\n才把它公开出来\n这个模型叫Jeff\n你知道吗\nTypeSafe给出的数字非常夸张\n说它20到200倍更快\n40到400倍更便宜\n不过我看到这里的时候\n真正让我觉得有意思的还不是这些数字\n而是Jav做了一件非常反常识的事情\n就是他不会说话\n真的不会\nJav他不跟你聊天\n不写文章\n也不生成自由文本\n你把一段对话\n一个agent的当前的状态\n或者一堆数据交给他\n然后问他\n这个请求安全吗\n下一步应该调用哪个工具呢\n这几个方案选哪一个呢\n这之后\nJav不会先写一段分析\n也不需要慢慢吐出一串杰森\n他会直接给你决定\n以及他对这个决定有多大把握\n欢迎回到All Insight\n在继续今天的视频之前\n我想先感谢一直支持这个频道的朋友\n我知道还有不少观看我视频的朋友\n还没有订阅我的频道\n那如果你觉得我的内容对你有所帮助\n欢迎点赞分享和订阅\n也可以加入频道会员\n每个月一杯咖啡的钱\n不仅是对我的支持\n也会帮助我继续制作更多关于AI和未来技术方向的深度内容\n好的 接下来我们进入今天的主题\n这件事听起来好像只是少生成了一些文字\n但放到今天的agent里面\n你会发现问题其实没有那么简单\n现在一个agent每走一步可能都要做很多判断\n比如下一步该调用哪个工具\n这个结果能不能相信\n要不要继续\n需不需要升级到更强的模型\n很多时候我们真正需要的\n其实只是一个是或者不是\nABC选哪个\n但今天最常见的做法\n是把整个上下文交给一个大语言模型\n让它先生成文字或者结构化输出\n然后程序再从这些输出里\n拿出最后那个决定\n所以阿尔梅达这次实际上是在问\n如果软件真正需要的只是一个决定\n那为什么一定要让AI先说一段话呢\nJav发布以后\nX上很快出现了两种不同的声音\n第一种反应非常直接\n就是等等这不就是分类器吗\n独立AI研究者Alexis Gallagher就说\n他觉得Jav很酷\n但更有意思的是大家对Jav的反应\n因为这似乎说明很多人已经忘了\n锦编码器分类器已经存在很多年了\n而且一直很好用\n像Modern Bird这样的模型\n本来就可以快速和廉价的完成结构化判断\nJav也许更好\n而且不需要针对每一个任务重新训练\n但这个类别本身并不新\n另一边XBOM的AI研究员Lan Butler马上反驳说\n问题恰恰就在不需要重新训练\n如果今天突然给你一个非常奇怪\n而且类别又特别多的决策任务\n你手里甚至还没有数据集\n那传统分类器怎么办\n而Jav的卖点就是直接用自然语言告诉他\n现在要判断什么\n更有意思的是\n已经有人不吵了\n他们直接把Jab塞进agent里面开始跑\n有人拿它做模型路由\n让Jab决定什么时候该用便宜模型\n什么时候升级到更强的模型\n还有人拿它做Computer Use的下一步决策\n还有一个叫Observed开发者\n测试了一整晚跑了500多次API\n总共才花了大约3美分\n一次同时多问20个问题\n延迟也只增加了47毫秒\n他还抓住了一次针对他自己agent的假审批注入\n然后他发现Jeff有一次打错了\n而Jeff给那个错误答案给出的信心是1.00\n这就有意思了\n一边是这不就是我们用了很多年的分类器吗\n而另一边却已经有人开始把他塞进agent的决策循环\n官方说他的概率经过较准\n但在真实测试里却已经出现了\n答错了但信心百分之百的情况\n所以这期视频我真正想弄清楚的\n不是Jav到底有没有快200倍\n而是说一个参与创造ChattGPT的人\n在模型越来越会说话之后\n为什么会花两年的时间\n重新做一个不会说话的AI呢\n以及他是不是碰到了今天A成的系统里\n一个更有意思的问题\n就是我们真的需要让大模型来做每一个决定吗\n那么一个不生成自由文本的AI\n到底是怎么工作的呢\nTypeSafe对Jav的描述其实很简洁\n翻译过来的意思就是\n输入可以很复杂\n但输出不再是一串自由生成的文字\n而是提前定义好类型的概率决策\n比如一个客服agent的正在处理退款请求\n你可以把完整的对话记录\n用户账户状态\n以及最近的付款记录一起交给Jav\n然后同时问他\n用户现在是不是在要求退款呢\n有没有欺诈风险呢\n需不需要转人工呢\n如果要继续处理\n下一步应该执行哪个动作呢\n当然普通大模型也能回答这些问题\n但它通常还是要生成\n哪怕最后输出的是JESN\n本质上也是一个Token接一个Token的\n把字符串写出来\n但JESN走的是另一条路\n他把这些问题直接变成了几个\n类型确定的决策\n而且可以一次并行回答\nTypeSafe目前把这种输出\n分成三种基本形式\n第一种叫Nual\n它处理是不是这一类的判断\n比如这个请求有没有越狱风险\n那么它返回的就不是一句\n我认为没有\n而是说这个判断对应的概率\n第二种叫Choice\n也就是从提前给定的选项里做选择\n比如一个Agent手里有八个工具\n现在应该调用哪一个呢\nJive直接在这八个选项里做判断\n并给出相应的概率\n第三种叫Score\n用来处理打分和排序\n比如一个任务的优先级有多高\n或者一个结果有多值得信任\nJF在开始计算之前\n输出空间就已经被软件定义好了\n它可以判断\n但它不能临场发明一个新的答案\n这里还有一个很容易被忽略的区别\n就是这些判断不一定要一个接一个的做\nType-Safe官方展示的Workflow\n本身就会把一个复杂的软件状态\n拆成很多独立的问题\n然后一次并行询问\n比如他们公开的客服工作流\n第一步就会同时读取用户意图\n情绪紧急程度以及欺诈风险等多个信号\n在后面的代码决定\n退款冻结银行卡转人工还是继续处理\n这时候我们再回头看TypeSafe给这种模型起的名字\n就比较容易理解了\n他们叫它System-1 Module\n这个名字来自丹尼尔·卡尼曼对System-1和System-2的区别\n他认为System-1负责快速和直觉式的判断\nSystem-2才负责缓慢和复杂的推理\n那Type-C想做的不是让Jeff取代所有大语言模型\n而是把软件系统里大量这种快速判断单独抽出来\n阿尔梅达后来把这个速度说得更直接\n他说他们把今天几条不同的训练路线\n看成是在解决不同的问题\n像RLVR就更适合可以验证答案的\nResonings和Proves\n而他们希望RLCD去解决的是自动控制\n也就是让模型的决定\n真正能够被软件直接使用\n这也是为什么Type-Safe\n一直把Jeff的典型用途写成\n分类 路由 打分 提取和分支判断\n他们甚至用了一个狠形象的说法\n是聪明的一笔句\n听到这里\nJav好像已经也没有那么神秘了\n他不是一个被砍掉聊天界面的GPT\n也不是为了写出更好的答案\n而是他想把做决定本身\n从语言生成里单独拆出来\n变成软件可以直接调语的一种智能原语\n但是这也就回到了开场那个质疑\n就是这不就是分类器吗\n以太坊基军会AI的负责人\nDavid Krapis的说法就更直接了\n他认为把Jav和通用ML放在一起比较\n很容易产生误导\n因为Jav根本不生成自由文本\n从功能上看\n它更接近一个非常聪明的工具\n或者一个分类器\n而不是GPT Cloud这种通用模型的直接竞争者\n有人继续往下拆\nRobert世文透客把Jav的几个卖点\n一条一条列了出来\n像是类型受限的输出\n我们已经可以用受限解码做\n概率可以从LogProps里面拿\n概率校准也早就有Temperature Scaling\n什么时候放行\n什么时候升级\n本来就可以在应用代表里面写Threathood\n所以他最后的评价其实很有意思\n他认为这些统计学上的零件\n没有哪一个是突然从天上掉下来的\n真正可能不同的是TypeSafe把校准放进了训练目标\n然后让API从一开始返回的就是Decision\n而不是一段等待程序再次解析的文字\n换句话说\n就不是发明了一堆从来没有存在过的零件\n而是少兵了几层东西\n这时候Jav就不是分类器吗\n这个问题就开始变得没那么简单了\n因为TypeSafe真正想挑战的\n并不是传统分类器做不到的分类\n而是它的使用方式\n假设你今天做一个金融agent\n你只需要判断的是\n这笔交易有没有欺诈风险\n那你当然可以收集数据标注样本\n训练一个专门的分类器\n如果业务一直不变\n这就可能比调用GF更便宜更快\n甚至更可靠\n但现实里的agent很少永远只问同一个问题\n今天你要判断的是欺诈\n明天可能增加是否需要人工复合\n后天可能又增加\n应该交给哪个风险团队\n换一个产品\n问题标签和业务规则可能全部都变了\n如果每增加一种判断\n都要重新积累数据\n训练和部署一个模型\n那传统分类器的成本\n就开始从推理阶段\n转移到了开发和维护阶段\n而Jeff想做的是把\n大语言模型最有价值的一部分能力保留下来\n就是你可以直接用自然语言告诉模型\n现在要判断什么\n那我们目前还没有看到\nJeff和一个认真微调过的\nMartin Burrton\n专门的小型分类器\n或者一个使用受限解码的小语言模型\n在同样任务同样数据同样准确率要求下\n做一次真正公平的正面对比\nTypeSafe真正需要证明的是\n当任务不断变化\n而你又不想每次重新训练一个模型的时候\nJab能不能同时保住分类器的速度和成本\n又同时保住大模型那种拿自然语言\n就能直接定义新任务的灵活性\n而这还只是第一层\n因为Type Save对Jav最大胆的主张\n其实不是快也不是便宜\n甚至不是自然语言可编程\n而是他认为\nJav不只会给你一个决定\n他还知道自己这个决定到底有多可靠\n这就进入了整个Jav的故事里\n最关键也是目前最神秘的一块\nRLCD\n前面讲到Jav会给每一个决策附上概率\n但这里真正重要的\n不是他能不能输出一个0.73或者0.91\n而是说这些数字到底有没有统计意义\n这就是RLCD想解决的问题\n它的全称是校准决策强化学习\n所谓校准\n我们可以先用一个很简单的例子来理解\n就是如果一个模型做了很多次判断\n我们把它所有认为有70%把握的判断放在一起\n那么一个校准的好的模型\n长期来看应该有大约70%是正确的\n他说90%就应该接近90%\n说50%就真的意味着\n这件事差不多只有一半的把握\n这和准确率其实是两回事\n一个模型完全可能总体准确率很高\n但他根本不知道自己什么时候会错\n而另一个模型也许平均准确率没有高出多少\n但他在容易出错的时候\n真的会把自己的概率降下来\n对于聊天机器人来说\n这个区别可能没有那么明显\n但对于agent来说\n这就直接决定了一件事\n就是这个概率能不能拿来控制系统\n比如一个agent正在决定\n要不要自动批准退款\n债务判断可以批准\n概率99.8%\n如果经过大量测试之后\n你真的知道这种99.8%值得性\n那么harness就可以设置一个预值\n超过这个概率自动执行\n但如果只有70%\n就升级给更强更贵的模型\n如果再低一点就交给人工或者verify\n这个时候CoundEfence就不再是给人看的一个装饰数字\n它变成了Agent的内部的一种控制信号\n就是什么时候继续\n什么时候升级\n什么时候验证\n什么时候停下来\n前面那个满分信心却答错的例子\n追问的也正是这件事\n就是这些概率究竟有多可信呢\n而RLCD真正生成改变的是另一件事\n就是从训练阶段开始\n就让模型学习一个决定应该有多确定\n让概率本身也成为训练目标的一部分\n但是,我目前還沒有看到TypeSafe公開完整的論文,具體的訓練流程和損失函數,也沒有足夠完整的第三方校準基準。\n所以RLCD是否真的比现有校准方法更好\n这一部分还要等更多的证据\n这里也刚好可以解释TypeSafe另一个\n很容易让人误解的说法\n就是Jav不能产生幻觉\n如果你只允许Jav在批准和拒绝之间选择\n他确实不可能突然回答\n也许也不会捏造第三个不存在的函数\n更不会生成一段格式错误程序\n根本解析不了的Jason\n它的输出空间在计算之前\n就已经被系统类型限制好了\n但是它完全可能选择批准\n而正确答案其实是拒绝\n所以更准确的说\nJav解决的是一种非常具体的幻觉\n它不会生成类型之外的答案\n但仍然可能在合法答案里面选错\n不过呢\n真正决定Jav能不能进入高风险自动化流程的\n还是后一个问题\n就是它选错的时候\n自己到底知不知道\n这也是为什么我觉得\nRLCD反而比Jav那个200倍更快的数字\n更值得关注\n如果一个模型只是快而且便宜\n我们其实已经有很多办法可以做到了\n但如果它能在保持这种速度和成本的同时\n给软件提供一个真正可信的不确定性信号\n那它在A层里面的位置就完全不一样了\n当然接下来还有一个跟现实的问题\n就是这些概率决策究竟能省下多少时间和成本呢\n现在我们终于可以回到\nJF发布时最新眼球的那组数字\n20到200倍更快\n40到400倍更便宜\nTypeSafe自己公布的Workflow Benchmark里面\n最高端的数字达到了193.6倍更快\n444.6倍更便宜\n这个数字第一眼看起来很离谱\n但离着JF前面的设计之后\n你会发现它其实并不神秘\n因为Jav和普通LM做计算的方式\n从一开始就不一样\n假设一个agent现在需要做20个判断\n比如用户是不是在退款\n有没有欺诈风险\n需不需要人工介入\n需要调用哪个工具\n结果要不要再验证一次\n阿尔梅达回应过一个很直接的问题\n就是为什么不干脆用受限解码呢\n他的回答是\n这层约束仍然叠在LM上\n字符串生成的开销还在\n他还提出了一更尖锐的判断\n就是模型本来就拿不准的时候\n再去限制他能选哪些输出\n反而会让模型变笨\n债务直接省掉了字符串生成这一步\n他不需要把决定翻译成一段语言\n再让软件把语言重新翻译成决定\n他的输出空间已经提前确定\n而且多个问题可以在同一次询问里并行计算\nTypeSafe公布的端到端延迟\n大约是70到500毫秒\n输入价格是每百万Token4.2美分\n输出则没有单独计费\n所以从物理账上看\nJav比大元模型快很多便宜很多\n是完全说得通的\n但我们真正需要问的是另一件事\n就是他们比较的是不是同一种工作呢\nTypeSafe的Workflow Benchmark\n不是让Jav和GPT Cloud\n各自写一篇文章\n也不是让他们解决一道开放式的数学题\n而是专门选择软件自动化里面的System-1任务\n对照的是同样接入这套工作流的\n按要求返回结果化结果的通用LM\n比如处理一张发票\n系统真正需要的是这几个决定\n付款、赞款、要求补文件、送人工审批\n还是怀疑欺诈\n又比如一个客服agent完成了一轮任务\n系统要判断的是自动关闭、人工复合、优化复合\n还是直接通知值班人员\n这些任务天然就非常适合被拆成很多窄判断\n然后再用代码组合\n而这正好是Jav的主场\nTypeSave甚至在自己的发布文章里主动承认\n他们展示的side-by-side demo\n使用了比较短和比较密集的输入\n这会让Jav看起来更有优势\n193.6倍和444.6倍\n来自他们设计的工作流程评估\n他们认为这应该处在现实收益的高端\n而且这些workflow是公司自己模型能力团队的人做的\n所以他们也明确写了一句\n可能存在偏见\n这一点反而很重要\n因为他告诉我们200倍不是一个模型智力倍率\n他真正表达的是\n帮任务本来就是大量窄决策\n而你还让一个通用生成模型不断生成字符串的时候\n你可能正在为很多根本不需要的能力付钱\nTypeSafe的工作流程评估\n其实还暴露了另外一个很有意思的结果\n就是同一个模型把任务拆成workflow\n让代码负责确定性的规则\n让模型只负责那些真正需要判断的部分的话\n那么在这些测试里\n就比把整套规则塞进一个巨大pront里\n一次性解决更准确更快和更便宜\n所以这笔账不只是和模型本身有关\n也和任务怎么拆\n软件怎么调用它有关\n那如果把这个思路放进一个完整的agent\n分工又会变成什么样呢\n讲到这里\n我觉得Jab最有意思的地方\n已经不是Jab本身了\n而是一个真正长期运行的agent\n到底应该把多少事情交给同一个大模型呢\n一个agent可能连续工作几十分钟\n甚至几个小时\n中间经历几十几百个决策点\n真正需要复杂推理的\n也许只有其中几步\n剩下大量的工作\n其实只是判断走哪条路\n调用哪个工具\n这个结果能不能接受\n需不需要再检查一次\n而Jav指向的是另一种架构\n真正开放复杂\n需要理解模糊目标\n长程规划和深度推理的问题\n仍然交给前沿模型\n如果是确定性的规则\n比如金额超过多少\n必须人工审批的\n就直接交给代码\n如果是大量高频窄而不断变化的判断\n就可以交给像Jav这种决策模型\n结果是不是满足某个明确标准\n可以再交给Verify\n需要长期保存的信息进入Memory\n需要真正改变外部世界的动作\n就交给Tools\n最后真正重要的\n可能不再是其中某一个模型\n而是外面的Harness\n因为Harness决定谁在什么时候被调用\n什么结果可以直接通过\n什么结果必须升级\n什么任务值得花几十倍的成本\n去唤醒最强的模型\n当软件真正需要的只是一个决定\n我们为什么一定要让AI先开口说话呢\n这里是All Insight\n感谢收看\n我们下期再见",
  "language": "zh",
  "duration": 1341.486438,
  "segments": [
    {
      "start": 0.0,
      "end": 27.58,
      "text": "9月15日迪奥哥·阿尔梅达在X上发了一条帖子。第一句话就很抓人。他说,参与创造XGP之后,我一直在问自己,为什么这些已经超过人类水平的聊天模型,没有带来AGI呢?这个问题从他嘴里说出来,确实有点特别。阿尔梅达是Intract GPT的作者之一,也参与过GPT4的RLHF和指令遵循的相关工作。",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12109144582044358,
      "no_speech_prob": 8.76492628476555e-12,
      "compression_ratio": 1.1370716510903427
    },
    {
      "start": 27.58,
      "end": 28.98,
      "text": "简单来说",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08632145812179869,
      "no_speech_prob": 1.3463728396057029e-11,
      "compression_ratio": 1.2294520547945205
    },
    {
      "start": 28.98,
      "end": 31.9,
      "text": "今天我们已经习以为常的这件事",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08632145812179869,
      "no_speech_prob": 1.3463728396057029e-11,
      "compression_ratio": 1.2294520547945205
    },
    {
      "start": 31.9,
      "end": 34.02,
      "text": "就是你对AI说一句话",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08632145812179869,
      "no_speech_prob": 1.3463728396057029e-11,
      "compression_ratio": 1.2294520547945205
    },
    {
      "start": 34.02,
      "end": 35.86,
      "text": "它能够理解你想干什么",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08632145812179869,
      "no_speech_prob": 1.3463728396057029e-11,
      "compression_ratio": 1.2294520547945205
    },
    {
      "start": 35.86,
      "end": 38.12,
      "text": "然后按照你的要求回答",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08632145812179869,
      "no_speech_prob": 1.3463728396057029e-11,
      "compression_ratio": 1.2294520547945205
    },
    {
      "start": 38.12,
      "end": 40.18,
      "text": "它就是这个早期参与",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08632145812179869,
      "no_speech_prob": 1.3463728396057029e-11,
      "compression_ratio": 1.2294520547945205
    },
    {
      "start": 40.18,
      "end": 42.72,
      "text": "把这条路线做出来的人之一",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08632145812179869,
      "no_speech_prob": 1.3463728396057029e-11,
      "compression_ratio": 1.2294520547945205
    },
    {
      "start": 42.72,
      "end": 45.1,
      "text": "不过参与创造ChatGPT",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08632145812179869,
      "no_speech_prob": 1.3463728396057029e-11,
      "compression_ratio": 1.2294520547945205
    },
    {
      "start": 45.1,
      "end": 46.3,
      "text": "是它自己的表述",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08632145812179869,
      "no_speech_prob": 1.3463728396057029e-11,
      "compression_ratio": 1.2294520547945205
    },
    {
      "start": 46.3,
      "end": 48.22,
      "text": "并不是OpenAI给它的",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08632145812179869,
      "no_speech_prob": 1.3463728396057029e-11,
      "compression_ratio": 1.2294520547945205
    },
    {
      "start": 48.22,
      "end": 50.5,
      "text": "ChatGPT联合发明人头衔",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08632145812179869,
      "no_speech_prob": 1.3463728396057029e-11,
      "compression_ratio": 1.2294520547945205
    },
    {
      "start": 50.5,
      "end": 53.74,
      "text": "但是ChatGPT出现以后",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08632145812179869,
      "no_speech_prob": 1.3463728396057029e-11,
      "compression_ratio": 1.2294520547945205
    },
    {
      "start": 53.74,
      "end": 55.78,
      "text": "它反而开始怀疑一件事了",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08632145812179869,
      "no_speech_prob": 1.3463728396057029e-11,
      "compression_ratio": 1.2294520547945205
    },
    {
      "start": 55.78,
      "end": 58.68,
      "text": "就是模型已经越来越会说话了",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12080804148027974,
      "no_speech_prob": 1.3521438309294087e-11,
      "compression_ratio": 1.205521472392638
    },
    {
      "start": 58.68,
      "end": 61.88,
      "text": "那为什么我们期待中的大规模自动化",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12080804148027974,
      "no_speech_prob": 1.3521438309294087e-11,
      "compression_ratio": 1.205521472392638
    },
    {
      "start": 61.88,
      "end": 64.04,
      "text": "却没有以同样的速度发生呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12080804148027974,
      "no_speech_prob": 1.3521438309294087e-11,
      "compression_ratio": 1.205521472392638
    },
    {
      "start": 64.04,
      "end": 65.8,
      "text": "于是过去两年",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12080804148027974,
      "no_speech_prob": 1.3521438309294087e-11,
      "compression_ratio": 1.205521472392638
    },
    {
      "start": 65.8,
      "end": 68.52,
      "text": "阿尔梅达一直在隐身做另一件东西",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12080804148027974,
      "no_speech_prob": 1.3521438309294087e-11,
      "compression_ratio": 1.205521472392638
    },
    {
      "start": 68.52,
      "end": 69.62,
      "text": "直到这几天",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12080804148027974,
      "no_speech_prob": 1.3521438309294087e-11,
      "compression_ratio": 1.205521472392638
    },
    {
      "start": 69.62,
      "end": 71.9,
      "text": "他和自己的公司TypeSafe AI",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12080804148027974,
      "no_speech_prob": 1.3521438309294087e-11,
      "compression_ratio": 1.205521472392638
    },
    {
      "start": 71.9,
      "end": 73.34,
      "text": "才把它公开出来",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12080804148027974,
      "no_speech_prob": 1.3521438309294087e-11,
      "compression_ratio": 1.205521472392638
    },
    {
      "start": 73.34,
      "end": 75.1,
      "text": "这个模型叫Jeff",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12080804148027974,
      "no_speech_prob": 1.3521438309294087e-11,
      "compression_ratio": 1.205521472392638
    },
    {
      "start": 75.1,
      "end": 76.38,
      "text": "你知道吗",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12080804148027974,
      "no_speech_prob": 1.3521438309294087e-11,
      "compression_ratio": 1.205521472392638
    },
    {
      "start": 76.38,
      "end": 79.06,
      "text": "TypeSafe给出的数字非常夸张",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12080804148027974,
      "no_speech_prob": 1.3521438309294087e-11,
      "compression_ratio": 1.205521472392638
    },
    {
      "start": 79.06,
      "end": 81.72,
      "text": "说它20到200倍更快",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12080804148027974,
      "no_speech_prob": 1.3521438309294087e-11,
      "compression_ratio": 1.205521472392638
    },
    {
      "start": 81.72,
      "end": 83.62,
      "text": "40到400倍更便宜",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12080804148027974,
      "no_speech_prob": 1.3521438309294087e-11,
      "compression_ratio": 1.205521472392638
    },
    {
      "start": 83.62,
      "end": 85.56,
      "text": "不过我看到这里的时候",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12080804148027974,
      "no_speech_prob": 1.3521438309294087e-11,
      "compression_ratio": 1.205521472392638
    },
    {
      "start": 85.56,
      "end": 88.7,
      "text": "真正让我觉得有意思的还不是这些数字",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.11226752381887495,
      "no_speech_prob": 1.4400296059757789e-11,
      "compression_ratio": 1.286644951140065
    },
    {
      "start": 88.7,
      "end": 91.86,
      "text": "而是Jav做了一件非常反常识的事情",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.11226752381887495,
      "no_speech_prob": 1.4400296059757789e-11,
      "compression_ratio": 1.286644951140065
    },
    {
      "start": 91.86,
      "end": 93.72,
      "text": "就是他不会说话",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.11226752381887495,
      "no_speech_prob": 1.4400296059757789e-11,
      "compression_ratio": 1.286644951140065
    },
    {
      "start": 93.72,
      "end": 94.74,
      "text": "真的不会",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.11226752381887495,
      "no_speech_prob": 1.4400296059757789e-11,
      "compression_ratio": 1.286644951140065
    },
    {
      "start": 94.74,
      "end": 97.06,
      "text": "Jav他不跟你聊天",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.11226752381887495,
      "no_speech_prob": 1.4400296059757789e-11,
      "compression_ratio": 1.286644951140065
    },
    {
      "start": 97.06,
      "end": 97.92,
      "text": "不写文章",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.11226752381887495,
      "no_speech_prob": 1.4400296059757789e-11,
      "compression_ratio": 1.286644951140065
    },
    {
      "start": 97.92,
      "end": 99.82,
      "text": "也不生成自由文本",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.11226752381887495,
      "no_speech_prob": 1.4400296059757789e-11,
      "compression_ratio": 1.286644951140065
    },
    {
      "start": 99.82,
      "end": 100.96,
      "text": "你把一段对话",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.11226752381887495,
      "no_speech_prob": 1.4400296059757789e-11,
      "compression_ratio": 1.286644951140065
    },
    {
      "start": 100.96,
      "end": 102.5,
      "text": "一个agent的当前的状态",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.11226752381887495,
      "no_speech_prob": 1.4400296059757789e-11,
      "compression_ratio": 1.286644951140065
    },
    {
      "start": 102.5,
      "end": 104.42,
      "text": "或者一堆数据交给他",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.11226752381887495,
      "no_speech_prob": 1.4400296059757789e-11,
      "compression_ratio": 1.286644951140065
    },
    {
      "start": 104.42,
      "end": 105.26,
      "text": "然后问他",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.11226752381887495,
      "no_speech_prob": 1.4400296059757789e-11,
      "compression_ratio": 1.286644951140065
    },
    {
      "start": 105.26,
      "end": 106.68,
      "text": "这个请求安全吗",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.11226752381887495,
      "no_speech_prob": 1.4400296059757789e-11,
      "compression_ratio": 1.286644951140065
    },
    {
      "start": 106.68,
      "end": 108.96,
      "text": "下一步应该调用哪个工具呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.11226752381887495,
      "no_speech_prob": 1.4400296059757789e-11,
      "compression_ratio": 1.286644951140065
    },
    {
      "start": 108.96,
      "end": 110.82,
      "text": "这几个方案选哪一个呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.11226752381887495,
      "no_speech_prob": 1.4400296059757789e-11,
      "compression_ratio": 1.286644951140065
    },
    {
      "start": 110.82,
      "end": 111.9,
      "text": "这之后",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.11226752381887495,
      "no_speech_prob": 1.4400296059757789e-11,
      "compression_ratio": 1.286644951140065
    },
    {
      "start": 111.9,
      "end": 113.9,
      "text": "Jav不会先写一段分析",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.11226752381887495,
      "no_speech_prob": 1.4400296059757789e-11,
      "compression_ratio": 1.286644951140065
    },
    {
      "start": 113.9,
      "end": 116.24,
      "text": "也不需要慢慢吐出一串杰森",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08654226230669625,
      "no_speech_prob": 1.5145171575192684e-11,
      "compression_ratio": 1.3198757763975155
    },
    {
      "start": 116.24,
      "end": 117.98,
      "text": "他会直接给你决定",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08654226230669625,
      "no_speech_prob": 1.5145171575192684e-11,
      "compression_ratio": 1.3198757763975155
    },
    {
      "start": 117.98,
      "end": 120.54,
      "text": "以及他对这个决定有多大把握",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08654226230669625,
      "no_speech_prob": 1.5145171575192684e-11,
      "compression_ratio": 1.3198757763975155
    },
    {
      "start": 120.54,
      "end": 122.38,
      "text": "欢迎回到All Insight",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08654226230669625,
      "no_speech_prob": 1.5145171575192684e-11,
      "compression_ratio": 1.3198757763975155
    },
    {
      "start": 122.38,
      "end": 124.2,
      "text": "在继续今天的视频之前",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08654226230669625,
      "no_speech_prob": 1.5145171575192684e-11,
      "compression_ratio": 1.3198757763975155
    },
    {
      "start": 124.2,
      "end": 126.84,
      "text": "我想先感谢一直支持这个频道的朋友",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08654226230669625,
      "no_speech_prob": 1.5145171575192684e-11,
      "compression_ratio": 1.3198757763975155
    },
    {
      "start": 126.84,
      "end": 129.64,
      "text": "我知道还有不少观看我视频的朋友",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08654226230669625,
      "no_speech_prob": 1.5145171575192684e-11,
      "compression_ratio": 1.3198757763975155
    },
    {
      "start": 129.64,
      "end": 130.86,
      "text": "还没有订阅我的频道",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08654226230669625,
      "no_speech_prob": 1.5145171575192684e-11,
      "compression_ratio": 1.3198757763975155
    },
    {
      "start": 130.86,
      "end": 133.74,
      "text": "那如果你觉得我的内容对你有所帮助",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08654226230669625,
      "no_speech_prob": 1.5145171575192684e-11,
      "compression_ratio": 1.3198757763975155
    },
    {
      "start": 133.74,
      "end": 135.8,
      "text": "欢迎点赞分享和订阅",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08654226230669625,
      "no_speech_prob": 1.5145171575192684e-11,
      "compression_ratio": 1.3198757763975155
    },
    {
      "start": 135.8,
      "end": 137.48,
      "text": "也可以加入频道会员",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08654226230669625,
      "no_speech_prob": 1.5145171575192684e-11,
      "compression_ratio": 1.3198757763975155
    },
    {
      "start": 137.48,
      "end": 139.44,
      "text": "每个月一杯咖啡的钱",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08654226230669625,
      "no_speech_prob": 1.5145171575192684e-11,
      "compression_ratio": 1.3198757763975155
    },
    {
      "start": 139.44,
      "end": 140.96,
      "text": "不仅是对我的支持",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.08654226230669625,
      "no_speech_prob": 1.5145171575192684e-11,
      "compression_ratio": 1.3198757763975155
    },
    {
      "start": 140.96,
      "end": 146.62,
      "text": "也会帮助我继续制作更多关于AI和未来技术方向的深度内容",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.0787290382385254,
      "no_speech_prob": 1.2661522871826314e-11,
      "compression_ratio": 1.2388059701492538
    },
    {
      "start": 146.62,
      "end": 149.34,
      "text": "好的 接下来我们进入今天的主题",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.0787290382385254,
      "no_speech_prob": 1.2661522871826314e-11,
      "compression_ratio": 1.2388059701492538
    },
    {
      "start": 149.34,
      "end": 153.26,
      "text": "这件事听起来好像只是少生成了一些文字",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.0787290382385254,
      "no_speech_prob": 1.2661522871826314e-11,
      "compression_ratio": 1.2388059701492538
    },
    {
      "start": 153.26,
      "end": 155.32,
      "text": "但放到今天的agent里面",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.0787290382385254,
      "no_speech_prob": 1.2661522871826314e-11,
      "compression_ratio": 1.2388059701492538
    },
    {
      "start": 155.32,
      "end": 158.08,
      "text": "你会发现问题其实没有那么简单",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.0787290382385254,
      "no_speech_prob": 1.2661522871826314e-11,
      "compression_ratio": 1.2388059701492538
    },
    {
      "start": 158.08,
      "end": 162.6,
      "text": "现在一个agent每走一步可能都要做很多判断",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.0787290382385254,
      "no_speech_prob": 1.2661522871826314e-11,
      "compression_ratio": 1.2388059701492538
    },
    {
      "start": 162.6,
      "end": 165.18,
      "text": "比如下一步该调用哪个工具",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.0787290382385254,
      "no_speech_prob": 1.2661522871826314e-11,
      "compression_ratio": 1.2388059701492538
    },
    {
      "start": 165.18,
      "end": 167.26,
      "text": "这个结果能不能相信",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.0787290382385254,
      "no_speech_prob": 1.2661522871826314e-11,
      "compression_ratio": 1.2388059701492538
    },
    {
      "start": 167.26,
      "end": 168.58,
      "text": "要不要继续",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.0787290382385254,
      "no_speech_prob": 1.2661522871826314e-11,
      "compression_ratio": 1.2388059701492538
    },
    {
      "start": 168.58,
      "end": 170.68,
      "text": "需不需要升级到更强的模型",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.0787290382385254,
      "no_speech_prob": 1.2661522871826314e-11,
      "compression_ratio": 1.2388059701492538
    },
    {
      "start": 170.68,
      "end": 173.6,
      "text": "很多时候我们真正需要的",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07160012038437637,
      "no_speech_prob": 1.4067139814111229e-11,
      "compression_ratio": 1.3344947735191637
    },
    {
      "start": 173.6,
      "end": 175.92,
      "text": "其实只是一个是或者不是",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07160012038437637,
      "no_speech_prob": 1.4067139814111229e-11,
      "compression_ratio": 1.3344947735191637
    },
    {
      "start": 175.92,
      "end": 177.4,
      "text": "ABC选哪个",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07160012038437637,
      "no_speech_prob": 1.4067139814111229e-11,
      "compression_ratio": 1.3344947735191637
    },
    {
      "start": 177.4,
      "end": 179.5,
      "text": "但今天最常见的做法",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07160012038437637,
      "no_speech_prob": 1.4067139814111229e-11,
      "compression_ratio": 1.3344947735191637
    },
    {
      "start": 179.5,
      "end": 182.46,
      "text": "是把整个上下文交给一个大语言模型",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07160012038437637,
      "no_speech_prob": 1.4067139814111229e-11,
      "compression_ratio": 1.3344947735191637
    },
    {
      "start": 182.46,
      "end": 185.7,
      "text": "让它先生成文字或者结构化输出",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07160012038437637,
      "no_speech_prob": 1.4067139814111229e-11,
      "compression_ratio": 1.3344947735191637
    },
    {
      "start": 185.7,
      "end": 188.5,
      "text": "然后程序再从这些输出里",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07160012038437637,
      "no_speech_prob": 1.4067139814111229e-11,
      "compression_ratio": 1.3344947735191637
    },
    {
      "start": 188.5,
      "end": 190.16,
      "text": "拿出最后那个决定",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07160012038437637,
      "no_speech_prob": 1.4067139814111229e-11,
      "compression_ratio": 1.3344947735191637
    },
    {
      "start": 190.16,
      "end": 193.98,
      "text": "所以阿尔梅达这次实际上是在问",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07160012038437637,
      "no_speech_prob": 1.4067139814111229e-11,
      "compression_ratio": 1.3344947735191637
    },
    {
      "start": 193.98,
      "end": 196.98,
      "text": "如果软件真正需要的只是一个决定",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07160012038437637,
      "no_speech_prob": 1.4067139814111229e-11,
      "compression_ratio": 1.3344947735191637
    },
    {
      "start": 196.98,
      "end": 199.98,
      "text": "那为什么一定要让AI先说一段话呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07160012038437637,
      "no_speech_prob": 1.4067139814111229e-11,
      "compression_ratio": 1.3344947735191637
    },
    {
      "start": 199.98,
      "end": 201.5,
      "text": "Jav发布以后",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12986414945578273,
      "no_speech_prob": 1.4566655173742227e-11,
      "compression_ratio": 1.2035398230088497
    },
    {
      "start": 201.5,
      "end": 205.14,
      "text": "X上很快出现了两种不同的声音",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12986414945578273,
      "no_speech_prob": 1.4566655173742227e-11,
      "compression_ratio": 1.2035398230088497
    },
    {
      "start": 205.14,
      "end": 206.76,
      "text": "第一种反应非常直接",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12986414945578273,
      "no_speech_prob": 1.4566655173742227e-11,
      "compression_ratio": 1.2035398230088497
    },
    {
      "start": 206.76,
      "end": 209.0,
      "text": "就是等等这不就是分类器吗",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12986414945578273,
      "no_speech_prob": 1.4566655173742227e-11,
      "compression_ratio": 1.2035398230088497
    },
    {
      "start": 209.0,
      "end": 211.66,
      "text": "独立AI研究者Alexis Gallagher就说",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12986414945578273,
      "no_speech_prob": 1.4566655173742227e-11,
      "compression_ratio": 1.2035398230088497
    },
    {
      "start": 211.66,
      "end": 213.26,
      "text": "他觉得Jav很酷",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12986414945578273,
      "no_speech_prob": 1.4566655173742227e-11,
      "compression_ratio": 1.2035398230088497
    },
    {
      "start": 213.26,
      "end": 216.14,
      "text": "但更有意思的是大家对Jav的反应",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12986414945578273,
      "no_speech_prob": 1.4566655173742227e-11,
      "compression_ratio": 1.2035398230088497
    },
    {
      "start": 216.14,
      "end": 219.32,
      "text": "因为这似乎说明很多人已经忘了",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12986414945578273,
      "no_speech_prob": 1.4566655173742227e-11,
      "compression_ratio": 1.2035398230088497
    },
    {
      "start": 219.32,
      "end": 222.42,
      "text": "锦编码器分类器已经存在很多年了",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12986414945578273,
      "no_speech_prob": 1.4566655173742227e-11,
      "compression_ratio": 1.2035398230088497
    },
    {
      "start": 222.42,
      "end": 223.9,
      "text": "而且一直很好用",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12986414945578273,
      "no_speech_prob": 1.4566655173742227e-11,
      "compression_ratio": 1.2035398230088497
    },
    {
      "start": 223.9,
      "end": 226.02,
      "text": "像Modern Bird这样的模型",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12986414945578273,
      "no_speech_prob": 1.4566655173742227e-11,
      "compression_ratio": 1.2035398230088497
    },
    {
      "start": 226.02,
      "end": 229.84,
      "text": "本来就可以快速和廉价的完成结构化判断",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.12986414945578273,
      "no_speech_prob": 1.4566655173742227e-11,
      "compression_ratio": 1.2035398230088497
    },
    {
      "start": 229.84,
      "end": 231.54,
      "text": "Jav也许更好",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1056989338583575,
      "no_speech_prob": 1.3787465093229034e-11,
      "compression_ratio": 1.2103658536585367
    },
    {
      "start": 231.54,
      "end": 234.56,
      "text": "而且不需要针对每一个任务重新训练",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1056989338583575,
      "no_speech_prob": 1.3787465093229034e-11,
      "compression_ratio": 1.2103658536585367
    },
    {
      "start": 234.56,
      "end": 236.96,
      "text": "但这个类别本身并不新",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1056989338583575,
      "no_speech_prob": 1.3787465093229034e-11,
      "compression_ratio": 1.2103658536585367
    },
    {
      "start": 236.96,
      "end": 241.4,
      "text": "另一边XBOM的AI研究员Lan Butler马上反驳说",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1056989338583575,
      "no_speech_prob": 1.3787465093229034e-11,
      "compression_ratio": 1.2103658536585367
    },
    {
      "start": 241.4,
      "end": 244.0,
      "text": "问题恰恰就在不需要重新训练",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1056989338583575,
      "no_speech_prob": 1.3787465093229034e-11,
      "compression_ratio": 1.2103658536585367
    },
    {
      "start": 244.0,
      "end": 246.82,
      "text": "如果今天突然给你一个非常奇怪",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1056989338583575,
      "no_speech_prob": 1.3787465093229034e-11,
      "compression_ratio": 1.2103658536585367
    },
    {
      "start": 246.82,
      "end": 249.22,
      "text": "而且类别又特别多的决策任务",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1056989338583575,
      "no_speech_prob": 1.3787465093229034e-11,
      "compression_ratio": 1.2103658536585367
    },
    {
      "start": 249.22,
      "end": 251.06,
      "text": "你手里甚至还没有数据集",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1056989338583575,
      "no_speech_prob": 1.3787465093229034e-11,
      "compression_ratio": 1.2103658536585367
    },
    {
      "start": 251.06,
      "end": 253.38,
      "text": "那传统分类器怎么办",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1056989338583575,
      "no_speech_prob": 1.3787465093229034e-11,
      "compression_ratio": 1.2103658536585367
    },
    {
      "start": 253.38,
      "end": 256.9,
      "text": "而Jav的卖点就是直接用自然语言告诉他",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1056989338583575,
      "no_speech_prob": 1.3787465093229034e-11,
      "compression_ratio": 1.2103658536585367
    },
    {
      "start": 256.9,
      "end": 258.2,
      "text": "现在要判断什么",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1056989338583575,
      "no_speech_prob": 1.3787465093229034e-11,
      "compression_ratio": 1.2103658536585367
    },
    {
      "start": 258.2,
      "end": 259.84,
      "text": "更有意思的是",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09643690145699083,
      "no_speech_prob": 1.1934822054249228e-11,
      "compression_ratio": 1.2074303405572755
    },
    {
      "start": 259.84,
      "end": 261.5,
      "text": "已经有人不吵了",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09643690145699083,
      "no_speech_prob": 1.1934822054249228e-11,
      "compression_ratio": 1.2074303405572755
    },
    {
      "start": 261.5,
      "end": 264.46,
      "text": "他们直接把Jab塞进agent里面开始跑",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09643690145699083,
      "no_speech_prob": 1.1934822054249228e-11,
      "compression_ratio": 1.2074303405572755
    },
    {
      "start": 264.46,
      "end": 266.6,
      "text": "有人拿它做模型路由",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09643690145699083,
      "no_speech_prob": 1.1934822054249228e-11,
      "compression_ratio": 1.2074303405572755
    },
    {
      "start": 266.6,
      "end": 269.48,
      "text": "让Jab决定什么时候该用便宜模型",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09643690145699083,
      "no_speech_prob": 1.1934822054249228e-11,
      "compression_ratio": 1.2074303405572755
    },
    {
      "start": 269.48,
      "end": 271.62,
      "text": "什么时候升级到更强的模型",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09643690145699083,
      "no_speech_prob": 1.1934822054249228e-11,
      "compression_ratio": 1.2074303405572755
    },
    {
      "start": 271.62,
      "end": 274.86,
      "text": "还有人拿它做Computer Use的下一步决策",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09643690145699083,
      "no_speech_prob": 1.1934822054249228e-11,
      "compression_ratio": 1.2074303405572755
    },
    {
      "start": 274.86,
      "end": 276.78,
      "text": "还有一个叫Observed开发者",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09643690145699083,
      "no_speech_prob": 1.1934822054249228e-11,
      "compression_ratio": 1.2074303405572755
    },
    {
      "start": 276.78,
      "end": 280.06,
      "text": "测试了一整晚跑了500多次API",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09643690145699083,
      "no_speech_prob": 1.1934822054249228e-11,
      "compression_ratio": 1.2074303405572755
    },
    {
      "start": 280.06,
      "end": 282.24,
      "text": "总共才花了大约3美分",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09643690145699083,
      "no_speech_prob": 1.1934822054249228e-11,
      "compression_ratio": 1.2074303405572755
    },
    {
      "start": 282.24,
      "end": 284.64,
      "text": "一次同时多问20个问题",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09643690145699083,
      "no_speech_prob": 1.1934822054249228e-11,
      "compression_ratio": 1.2074303405572755
    },
    {
      "start": 284.64,
      "end": 287.34,
      "text": "延迟也只增加了47毫秒",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09643690145699083,
      "no_speech_prob": 1.1934822054249228e-11,
      "compression_ratio": 1.2074303405572755
    },
    {
      "start": 287.34,
      "end": 291.78,
      "text": "他还抓住了一次针对他自己agent的假审批注入",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10219100853065391,
      "no_speech_prob": 1.2943673043108728e-11,
      "compression_ratio": 1.2806451612903227
    },
    {
      "start": 291.78,
      "end": 294.86,
      "text": "然后他发现Jeff有一次打错了",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10219100853065391,
      "no_speech_prob": 1.2943673043108728e-11,
      "compression_ratio": 1.2806451612903227
    },
    {
      "start": 294.86,
      "end": 298.22,
      "text": "而Jeff给那个错误答案给出的信心是1.00",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10219100853065391,
      "no_speech_prob": 1.2943673043108728e-11,
      "compression_ratio": 1.2806451612903227
    },
    {
      "start": 298.22,
      "end": 299.92,
      "text": "这就有意思了",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10219100853065391,
      "no_speech_prob": 1.2943673043108728e-11,
      "compression_ratio": 1.2806451612903227
    },
    {
      "start": 299.92,
      "end": 303.16,
      "text": "一边是这不就是我们用了很多年的分类器吗",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10219100853065391,
      "no_speech_prob": 1.2943673043108728e-11,
      "compression_ratio": 1.2806451612903227
    },
    {
      "start": 303.16,
      "end": 308.18,
      "text": "而另一边却已经有人开始把他塞进agent的决策循环",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10219100853065391,
      "no_speech_prob": 1.2943673043108728e-11,
      "compression_ratio": 1.2806451612903227
    },
    {
      "start": 308.18,
      "end": 310.7,
      "text": "官方说他的概率经过较准",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10219100853065391,
      "no_speech_prob": 1.2943673043108728e-11,
      "compression_ratio": 1.2806451612903227
    },
    {
      "start": 310.7,
      "end": 313.38,
      "text": "但在真实测试里却已经出现了",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10219100853065391,
      "no_speech_prob": 1.2943673043108728e-11,
      "compression_ratio": 1.2806451612903227
    },
    {
      "start": 313.38,
      "end": 316.06,
      "text": "答错了但信心百分之百的情况",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10219100853065391,
      "no_speech_prob": 1.2943673043108728e-11,
      "compression_ratio": 1.2806451612903227
    },
    {
      "start": 316.06,
      "end": 320.02,
      "text": "所以这期视频我真正想弄清楚的",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14895453923185106,
      "no_speech_prob": 1.2269812769971633e-11,
      "compression_ratio": 1.2769230769230768
    },
    {
      "start": 320.02,
      "end": 322.36,
      "text": "不是Jav到底有没有快200倍",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14895453923185106,
      "no_speech_prob": 1.2269812769971633e-11,
      "compression_ratio": 1.2769230769230768
    },
    {
      "start": 322.36,
      "end": 325.74,
      "text": "而是说一个参与创造ChattGPT的人",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14895453923185106,
      "no_speech_prob": 1.2269812769971633e-11,
      "compression_ratio": 1.2769230769230768
    },
    {
      "start": 325.74,
      "end": 327.94,
      "text": "在模型越来越会说话之后",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14895453923185106,
      "no_speech_prob": 1.2269812769971633e-11,
      "compression_ratio": 1.2769230769230768
    },
    {
      "start": 327.94,
      "end": 329.76,
      "text": "为什么会花两年的时间",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14895453923185106,
      "no_speech_prob": 1.2269812769971633e-11,
      "compression_ratio": 1.2769230769230768
    },
    {
      "start": 329.76,
      "end": 331.96,
      "text": "重新做一个不会说话的AI呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14895453923185106,
      "no_speech_prob": 1.2269812769971633e-11,
      "compression_ratio": 1.2769230769230768
    },
    {
      "start": 331.96,
      "end": 335.22,
      "text": "以及他是不是碰到了今天A成的系统里",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14895453923185106,
      "no_speech_prob": 1.2269812769971633e-11,
      "compression_ratio": 1.2769230769230768
    },
    {
      "start": 335.22,
      "end": 336.68,
      "text": "一个更有意思的问题",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14895453923185106,
      "no_speech_prob": 1.2269812769971633e-11,
      "compression_ratio": 1.2769230769230768
    },
    {
      "start": 336.68,
      "end": 340.48,
      "text": "就是我们真的需要让大模型来做每一个决定吗",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14895453923185106,
      "no_speech_prob": 1.2269812769971633e-11,
      "compression_ratio": 1.2769230769230768
    },
    {
      "start": 340.48,
      "end": 343.4,
      "text": "那么一个不生成自由文本的AI",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14895453923185106,
      "no_speech_prob": 1.2269812769971633e-11,
      "compression_ratio": 1.2769230769230768
    },
    {
      "start": 343.4,
      "end": 345.24,
      "text": "到底是怎么工作的呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14895453923185106,
      "no_speech_prob": 1.2269812769971633e-11,
      "compression_ratio": 1.2769230769230768
    },
    {
      "start": 345.24,
      "end": 348.3,
      "text": "TypeSafe对Jav的描述其实很简洁",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07746307099063093,
      "no_speech_prob": 1.46757380553586e-11,
      "compression_ratio": 1.2642642642642643
    },
    {
      "start": 348.3,
      "end": 349.98,
      "text": "翻译过来的意思就是",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07746307099063093,
      "no_speech_prob": 1.46757380553586e-11,
      "compression_ratio": 1.2642642642642643
    },
    {
      "start": 349.98,
      "end": 351.58,
      "text": "输入可以很复杂",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07746307099063093,
      "no_speech_prob": 1.46757380553586e-11,
      "compression_ratio": 1.2642642642642643
    },
    {
      "start": 351.58,
      "end": 354.58,
      "text": "但输出不再是一串自由生成的文字",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07746307099063093,
      "no_speech_prob": 1.46757380553586e-11,
      "compression_ratio": 1.2642642642642643
    },
    {
      "start": 354.58,
      "end": 358.04,
      "text": "而是提前定义好类型的概率决策",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07746307099063093,
      "no_speech_prob": 1.46757380553586e-11,
      "compression_ratio": 1.2642642642642643
    },
    {
      "start": 358.04,
      "end": 361.52,
      "text": "比如一个客服agent的正在处理退款请求",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07746307099063093,
      "no_speech_prob": 1.46757380553586e-11,
      "compression_ratio": 1.2642642642642643
    },
    {
      "start": 361.52,
      "end": 363.96,
      "text": "你可以把完整的对话记录",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07746307099063093,
      "no_speech_prob": 1.46757380553586e-11,
      "compression_ratio": 1.2642642642642643
    },
    {
      "start": 363.96,
      "end": 365.22,
      "text": "用户账户状态",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07746307099063093,
      "no_speech_prob": 1.46757380553586e-11,
      "compression_ratio": 1.2642642642642643
    },
    {
      "start": 365.22,
      "end": 367.96,
      "text": "以及最近的付款记录一起交给Jav",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07746307099063093,
      "no_speech_prob": 1.46757380553586e-11,
      "compression_ratio": 1.2642642642642643
    },
    {
      "start": 367.96,
      "end": 369.14,
      "text": "然后同时问他",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07746307099063093,
      "no_speech_prob": 1.46757380553586e-11,
      "compression_ratio": 1.2642642642642643
    },
    {
      "start": 369.14,
      "end": 371.68,
      "text": "用户现在是不是在要求退款呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07746307099063093,
      "no_speech_prob": 1.46757380553586e-11,
      "compression_ratio": 1.2642642642642643
    },
    {
      "start": 371.68,
      "end": 373.32,
      "text": "有没有欺诈风险呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07746307099063093,
      "no_speech_prob": 1.46757380553586e-11,
      "compression_ratio": 1.2642642642642643
    },
    {
      "start": 373.32,
      "end": 374.76,
      "text": "需不需要转人工呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07746307099063093,
      "no_speech_prob": 1.46757380553586e-11,
      "compression_ratio": 1.2642642642642643
    },
    {
      "start": 374.76,
      "end": 376.46,
      "text": "如果要继续处理",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1348577474618887,
      "no_speech_prob": 1.3208185413449147e-11,
      "compression_ratio": 1.251700680272109
    },
    {
      "start": 376.46,
      "end": 378.62,
      "text": "下一步应该执行哪个动作呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1348577474618887,
      "no_speech_prob": 1.3208185413449147e-11,
      "compression_ratio": 1.251700680272109
    },
    {
      "start": 378.62,
      "end": 382.66,
      "text": "当然普通大模型也能回答这些问题",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1348577474618887,
      "no_speech_prob": 1.3208185413449147e-11,
      "compression_ratio": 1.251700680272109
    },
    {
      "start": 382.66,
      "end": 384.76,
      "text": "但它通常还是要生成",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1348577474618887,
      "no_speech_prob": 1.3208185413449147e-11,
      "compression_ratio": 1.251700680272109
    },
    {
      "start": 384.76,
      "end": 387.14,
      "text": "哪怕最后输出的是JESN",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1348577474618887,
      "no_speech_prob": 1.3208185413449147e-11,
      "compression_ratio": 1.251700680272109
    },
    {
      "start": 387.14,
      "end": 389.78,
      "text": "本质上也是一个Token接一个Token的",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1348577474618887,
      "no_speech_prob": 1.3208185413449147e-11,
      "compression_ratio": 1.251700680272109
    },
    {
      "start": 389.78,
      "end": 391.32,
      "text": "把字符串写出来",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1348577474618887,
      "no_speech_prob": 1.3208185413449147e-11,
      "compression_ratio": 1.251700680272109
    },
    {
      "start": 391.32,
      "end": 393.46,
      "text": "但JESN走的是另一条路",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1348577474618887,
      "no_speech_prob": 1.3208185413449147e-11,
      "compression_ratio": 1.251700680272109
    },
    {
      "start": 393.46,
      "end": 396.26,
      "text": "他把这些问题直接变成了几个",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1348577474618887,
      "no_speech_prob": 1.3208185413449147e-11,
      "compression_ratio": 1.251700680272109
    },
    {
      "start": 396.26,
      "end": 397.82,
      "text": "类型确定的决策",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1348577474618887,
      "no_speech_prob": 1.3208185413449147e-11,
      "compression_ratio": 1.251700680272109
    },
    {
      "start": 397.82,
      "end": 400.36,
      "text": "而且可以一次并行回答",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1348577474618887,
      "no_speech_prob": 1.3208185413449147e-11,
      "compression_ratio": 1.251700680272109
    },
    {
      "start": 400.36,
      "end": 403.02,
      "text": "TypeSafe目前把这种输出",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1348577474618887,
      "no_speech_prob": 1.3208185413449147e-11,
      "compression_ratio": 1.251700680272109
    },
    {
      "start": 403.02,
      "end": 405.02,
      "text": "分成三种基本形式",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1278513515696806,
      "no_speech_prob": 1.78267383493802e-11,
      "compression_ratio": 1.3300970873786409
    },
    {
      "start": 405.02,
      "end": 406.34,
      "text": "第一种叫Nual",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1278513515696806,
      "no_speech_prob": 1.78267383493802e-11,
      "compression_ratio": 1.3300970873786409
    },
    {
      "start": 406.34,
      "end": 408.42,
      "text": "它处理是不是这一类的判断",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1278513515696806,
      "no_speech_prob": 1.78267383493802e-11,
      "compression_ratio": 1.3300970873786409
    },
    {
      "start": 408.42,
      "end": 411.3,
      "text": "比如这个请求有没有越狱风险",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1278513515696806,
      "no_speech_prob": 1.78267383493802e-11,
      "compression_ratio": 1.3300970873786409
    },
    {
      "start": 411.3,
      "end": 413.1,
      "text": "那么它返回的就不是一句",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1278513515696806,
      "no_speech_prob": 1.78267383493802e-11,
      "compression_ratio": 1.3300970873786409
    },
    {
      "start": 413.1,
      "end": 413.98,
      "text": "我认为没有",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1278513515696806,
      "no_speech_prob": 1.78267383493802e-11,
      "compression_ratio": 1.3300970873786409
    },
    {
      "start": 413.98,
      "end": 416.5,
      "text": "而是说这个判断对应的概率",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1278513515696806,
      "no_speech_prob": 1.78267383493802e-11,
      "compression_ratio": 1.3300970873786409
    },
    {
      "start": 416.5,
      "end": 418.0,
      "text": "第二种叫Choice",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1278513515696806,
      "no_speech_prob": 1.78267383493802e-11,
      "compression_ratio": 1.3300970873786409
    },
    {
      "start": 418.0,
      "end": 421.64,
      "text": "也就是从提前给定的选项里做选择",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1278513515696806,
      "no_speech_prob": 1.78267383493802e-11,
      "compression_ratio": 1.3300970873786409
    },
    {
      "start": 421.64,
      "end": 424.06,
      "text": "比如一个Agent手里有八个工具",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1278513515696806,
      "no_speech_prob": 1.78267383493802e-11,
      "compression_ratio": 1.3300970873786409
    },
    {
      "start": 424.06,
      "end": 425.8,
      "text": "现在应该调用哪一个呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1278513515696806,
      "no_speech_prob": 1.78267383493802e-11,
      "compression_ratio": 1.3300970873786409
    },
    {
      "start": 425.8,
      "end": 428.4,
      "text": "Jive直接在这八个选项里做判断",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1278513515696806,
      "no_speech_prob": 1.78267383493802e-11,
      "compression_ratio": 1.3300970873786409
    },
    {
      "start": 428.4,
      "end": 430.08,
      "text": "并给出相应的概率",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1278513515696806,
      "no_speech_prob": 1.78267383493802e-11,
      "compression_ratio": 1.3300970873786409
    },
    {
      "start": 430.08,
      "end": 431.6,
      "text": "第三种叫Score",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1278513515696806,
      "no_speech_prob": 1.78267383493802e-11,
      "compression_ratio": 1.3300970873786409
    },
    {
      "start": 431.6,
      "end": 433.58,
      "text": "用来处理打分和排序",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.050853729248046875,
      "no_speech_prob": 7.648058401865665e-12,
      "compression_ratio": 0.9807692307692307
    },
    {
      "start": 433.58,
      "end": 435.96,
      "text": "比如一个任务的优先级有多高",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.050853729248046875,
      "no_speech_prob": 7.648058401865665e-12,
      "compression_ratio": 0.9807692307692307
    },
    {
      "start": 435.96,
      "end": 438.26,
      "text": "或者一个结果有多值得信任",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.050853729248046875,
      "no_speech_prob": 7.648058401865665e-12,
      "compression_ratio": 0.9807692307692307
    },
    {
      "start": 453.246,
      "end": 455.546,
      "text": "JF在开始计算之前",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10647632598876954,
      "no_speech_prob": 1.3010984650785318e-11,
      "compression_ratio": 1.2147435897435896
    },
    {
      "start": 455.546,
      "end": 458.766,
      "text": "输出空间就已经被软件定义好了",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10647632598876954,
      "no_speech_prob": 1.3010984650785318e-11,
      "compression_ratio": 1.2147435897435896
    },
    {
      "start": 458.766,
      "end": 460.146,
      "text": "它可以判断",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10647632598876954,
      "no_speech_prob": 1.3010984650785318e-11,
      "compression_ratio": 1.2147435897435896
    },
    {
      "start": 460.146,
      "end": 463.046,
      "text": "但它不能临场发明一个新的答案",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10647632598876954,
      "no_speech_prob": 1.3010984650785318e-11,
      "compression_ratio": 1.2147435897435896
    },
    {
      "start": 463.046,
      "end": 465.126,
      "text": "这里还有一个很容易被忽略的区别",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10647632598876954,
      "no_speech_prob": 1.3010984650785318e-11,
      "compression_ratio": 1.2147435897435896
    },
    {
      "start": 465.126,
      "end": 469.126,
      "text": "就是这些判断不一定要一个接一个的做",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10647632598876954,
      "no_speech_prob": 1.3010984650785318e-11,
      "compression_ratio": 1.2147435897435896
    },
    {
      "start": 469.126,
      "end": 471.526,
      "text": "Type-Safe官方展示的Workflow",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10647632598876954,
      "no_speech_prob": 1.3010984650785318e-11,
      "compression_ratio": 1.2147435897435896
    },
    {
      "start": 471.526,
      "end": 474.406,
      "text": "本身就会把一个复杂的软件状态",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10647632598876954,
      "no_speech_prob": 1.3010984650785318e-11,
      "compression_ratio": 1.2147435897435896
    },
    {
      "start": 474.406,
      "end": 476.066,
      "text": "拆成很多独立的问题",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10647632598876954,
      "no_speech_prob": 1.3010984650785318e-11,
      "compression_ratio": 1.2147435897435896
    },
    {
      "start": 476.066,
      "end": 477.926,
      "text": "然后一次并行询问",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10647632598876954,
      "no_speech_prob": 1.3010984650785318e-11,
      "compression_ratio": 1.2147435897435896
    },
    {
      "start": 477.926,
      "end": 480.306,
      "text": "比如他们公开的客服工作流",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10647632598876954,
      "no_speech_prob": 1.3010984650785318e-11,
      "compression_ratio": 1.2147435897435896
    },
    {
      "start": 480.306,
      "end": 483.286,
      "text": "第一步就会同时读取用户意图",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10723382039786944,
      "no_speech_prob": 1.475515716553577e-11,
      "compression_ratio": 1.131578947368421
    },
    {
      "start": 483.286,
      "end": 487.406,
      "text": "情绪紧急程度以及欺诈风险等多个信号",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10723382039786944,
      "no_speech_prob": 1.475515716553577e-11,
      "compression_ratio": 1.131578947368421
    },
    {
      "start": 487.406,
      "end": 489.126,
      "text": "在后面的代码决定",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10723382039786944,
      "no_speech_prob": 1.475515716553577e-11,
      "compression_ratio": 1.131578947368421
    },
    {
      "start": 489.126,
      "end": 493.126,
      "text": "退款冻结银行卡转人工还是继续处理",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10723382039786944,
      "no_speech_prob": 1.475515716553577e-11,
      "compression_ratio": 1.131578947368421
    },
    {
      "start": 493.126,
      "end": 497.546,
      "text": "这时候我们再回头看TypeSafe给这种模型起的名字",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10723382039786944,
      "no_speech_prob": 1.475515716553577e-11,
      "compression_ratio": 1.131578947368421
    },
    {
      "start": 497.546,
      "end": 499.346,
      "text": "就比较容易理解了",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10723382039786944,
      "no_speech_prob": 1.475515716553577e-11,
      "compression_ratio": 1.131578947368421
    },
    {
      "start": 499.346,
      "end": 501.286,
      "text": "他们叫它System-1 Module",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10723382039786944,
      "no_speech_prob": 1.475515716553577e-11,
      "compression_ratio": 1.131578947368421
    },
    {
      "start": 501.286,
      "end": 506.246,
      "text": "这个名字来自丹尼尔·卡尼曼对System-1和System-2的区别",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10723382039786944,
      "no_speech_prob": 1.475515716553577e-11,
      "compression_ratio": 1.131578947368421
    },
    {
      "start": 506.246,
      "end": 510.646,
      "text": "他认为System-1负责快速和直觉式的判断",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010150625899031,
      "no_speech_prob": 1.2063954869800941e-11,
      "compression_ratio": 1.1758620689655173
    },
    {
      "start": 510.646,
      "end": 513.946,
      "text": "System-2才负责缓慢和复杂的推理",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010150625899031,
      "no_speech_prob": 1.2063954869800941e-11,
      "compression_ratio": 1.1758620689655173
    },
    {
      "start": 513.946,
      "end": 518.806,
      "text": "那Type-C想做的不是让Jeff取代所有大语言模型",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010150625899031,
      "no_speech_prob": 1.2063954869800941e-11,
      "compression_ratio": 1.1758620689655173
    },
    {
      "start": 518.806,
      "end": 524.206,
      "text": "而是把软件系统里大量这种快速判断单独抽出来",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010150625899031,
      "no_speech_prob": 1.2063954869800941e-11,
      "compression_ratio": 1.1758620689655173
    },
    {
      "start": 524.206,
      "end": 527.386,
      "text": "阿尔梅达后来把这个速度说得更直接",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010150625899031,
      "no_speech_prob": 1.2063954869800941e-11,
      "compression_ratio": 1.1758620689655173
    },
    {
      "start": 527.386,
      "end": 530.726,
      "text": "他说他们把今天几条不同的训练路线",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010150625899031,
      "no_speech_prob": 1.2063954869800941e-11,
      "compression_ratio": 1.1758620689655173
    },
    {
      "start": 530.726,
      "end": 532.906,
      "text": "看成是在解决不同的问题",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09010150625899031,
      "no_speech_prob": 1.2063954869800941e-11,
      "compression_ratio": 1.1758620689655173
    },
    {
      "start": 532.906,
      "end": 536.646,
      "text": "像RLVR就更适合可以验证答案的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13628989703034702,
      "no_speech_prob": 1.1208395322981346e-11,
      "compression_ratio": 1.125
    },
    {
      "start": 536.646,
      "end": 538.046,
      "text": "Resonings和Proves",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13628989703034702,
      "no_speech_prob": 1.1208395322981346e-11,
      "compression_ratio": 1.125
    },
    {
      "start": 538.046,
      "end": 542.926,
      "text": "而他们希望RLCD去解决的是自动控制",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13628989703034702,
      "no_speech_prob": 1.1208395322981346e-11,
      "compression_ratio": 1.125
    },
    {
      "start": 542.926,
      "end": 544.986,
      "text": "也就是让模型的决定",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13628989703034702,
      "no_speech_prob": 1.1208395322981346e-11,
      "compression_ratio": 1.125
    },
    {
      "start": 544.986,
      "end": 547.766,
      "text": "真正能够被软件直接使用",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13628989703034702,
      "no_speech_prob": 1.1208395322981346e-11,
      "compression_ratio": 1.125
    },
    {
      "start": 547.766,
      "end": 549.366,
      "text": "这也是为什么Type-Safe",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13628989703034702,
      "no_speech_prob": 1.1208395322981346e-11,
      "compression_ratio": 1.125
    },
    {
      "start": 549.366,
      "end": 551.626,
      "text": "一直把Jeff的典型用途写成",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13628989703034702,
      "no_speech_prob": 1.1208395322981346e-11,
      "compression_ratio": 1.125
    },
    {
      "start": 551.626,
      "end": 555.346,
      "text": "分类 路由 打分 提取和分支判断",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13628989703034702,
      "no_speech_prob": 1.1208395322981346e-11,
      "compression_ratio": 1.125
    },
    {
      "start": 555.346,
      "end": 557.606,
      "text": "他们甚至用了一个狠形象的说法",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13628989703034702,
      "no_speech_prob": 1.1208395322981346e-11,
      "compression_ratio": 1.125
    },
    {
      "start": 557.606,
      "end": 559.326,
      "text": "是聪明的一笔句",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13628989703034702,
      "no_speech_prob": 1.1208395322981346e-11,
      "compression_ratio": 1.125
    },
    {
      "start": 559.326,
      "end": 561.146,
      "text": "听到这里",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13628989703034702,
      "no_speech_prob": 1.1208395322981346e-11,
      "compression_ratio": 1.125
    },
    {
      "start": 561.146,
      "end": 563.506,
      "text": "Jav好像已经也没有那么神秘了",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11989893411335192,
      "no_speech_prob": 1.538397881306608e-11,
      "compression_ratio": 1.196774193548387
    },
    {
      "start": 563.506,
      "end": 566.666,
      "text": "他不是一个被砍掉聊天界面的GPT",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11989893411335192,
      "no_speech_prob": 1.538397881306608e-11,
      "compression_ratio": 1.196774193548387
    },
    {
      "start": 566.666,
      "end": 569.346,
      "text": "也不是为了写出更好的答案",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11989893411335192,
      "no_speech_prob": 1.538397881306608e-11,
      "compression_ratio": 1.196774193548387
    },
    {
      "start": 569.346,
      "end": 571.906,
      "text": "而是他想把做决定本身",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11989893411335192,
      "no_speech_prob": 1.538397881306608e-11,
      "compression_ratio": 1.196774193548387
    },
    {
      "start": 571.906,
      "end": 574.146,
      "text": "从语言生成里单独拆出来",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11989893411335192,
      "no_speech_prob": 1.538397881306608e-11,
      "compression_ratio": 1.196774193548387
    },
    {
      "start": 574.146,
      "end": 577.726,
      "text": "变成软件可以直接调语的一种智能原语",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11989893411335192,
      "no_speech_prob": 1.538397881306608e-11,
      "compression_ratio": 1.196774193548387
    },
    {
      "start": 577.726,
      "end": 581.406,
      "text": "但是这也就回到了开场那个质疑",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11989893411335192,
      "no_speech_prob": 1.538397881306608e-11,
      "compression_ratio": 1.196774193548387
    },
    {
      "start": 581.406,
      "end": 583.346,
      "text": "就是这不就是分类器吗",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11989893411335192,
      "no_speech_prob": 1.538397881306608e-11,
      "compression_ratio": 1.196774193548387
    },
    {
      "start": 583.346,
      "end": 585.906,
      "text": "以太坊基军会AI的负责人",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11989893411335192,
      "no_speech_prob": 1.538397881306608e-11,
      "compression_ratio": 1.196774193548387
    },
    {
      "start": 585.906,
      "end": 588.326,
      "text": "David Krapis的说法就更直接了",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11989893411335192,
      "no_speech_prob": 1.538397881306608e-11,
      "compression_ratio": 1.196774193548387
    },
    {
      "start": 588.326,
      "end": 592.006,
      "text": "他认为把Jav和通用ML放在一起比较",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10058372643343202,
      "no_speech_prob": 1.3210544637376476e-11,
      "compression_ratio": 1.1930379746835442
    },
    {
      "start": 592.006,
      "end": 593.306,
      "text": "很容易产生误导",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10058372643343202,
      "no_speech_prob": 1.3210544637376476e-11,
      "compression_ratio": 1.1930379746835442
    },
    {
      "start": 593.306,
      "end": 596.326,
      "text": "因为Jav根本不生成自由文本",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10058372643343202,
      "no_speech_prob": 1.3210544637376476e-11,
      "compression_ratio": 1.1930379746835442
    },
    {
      "start": 596.326,
      "end": 597.666,
      "text": "从功能上看",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10058372643343202,
      "no_speech_prob": 1.3210544637376476e-11,
      "compression_ratio": 1.1930379746835442
    },
    {
      "start": 597.666,
      "end": 600.066,
      "text": "它更接近一个非常聪明的工具",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10058372643343202,
      "no_speech_prob": 1.3210544637376476e-11,
      "compression_ratio": 1.1930379746835442
    },
    {
      "start": 600.066,
      "end": 601.306,
      "text": "或者一个分类器",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10058372643343202,
      "no_speech_prob": 1.3210544637376476e-11,
      "compression_ratio": 1.1930379746835442
    },
    {
      "start": 601.306,
      "end": 605.526,
      "text": "而不是GPT Cloud这种通用模型的直接竞争者",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10058372643343202,
      "no_speech_prob": 1.3210544637376476e-11,
      "compression_ratio": 1.1930379746835442
    },
    {
      "start": 605.526,
      "end": 607.346,
      "text": "有人继续往下拆",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10058372643343202,
      "no_speech_prob": 1.3210544637376476e-11,
      "compression_ratio": 1.1930379746835442
    },
    {
      "start": 607.346,
      "end": 610.466,
      "text": "Robert世文透客把Jav的几个卖点",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10058372643343202,
      "no_speech_prob": 1.3210544637376476e-11,
      "compression_ratio": 1.1930379746835442
    },
    {
      "start": 610.466,
      "end": 611.926,
      "text": "一条一条列了出来",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10058372643343202,
      "no_speech_prob": 1.3210544637376476e-11,
      "compression_ratio": 1.1930379746835442
    },
    {
      "start": 611.926,
      "end": 614.086,
      "text": "像是类型受限的输出",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10058372643343202,
      "no_speech_prob": 1.3210544637376476e-11,
      "compression_ratio": 1.1930379746835442
    },
    {
      "start": 614.086,
      "end": 616.546,
      "text": "我们已经可以用受限解码做",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10058372643343202,
      "no_speech_prob": 1.3210544637376476e-11,
      "compression_ratio": 1.1930379746835442
    },
    {
      "start": 616.546,
      "end": 619.026,
      "text": "概率可以从LogProps里面拿",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11474531060023023,
      "no_speech_prob": 1.3416833616330948e-11,
      "compression_ratio": 1.1842900302114803
    },
    {
      "start": 619.026,
      "end": 622.466,
      "text": "概率校准也早就有Temperature Scaling",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11474531060023023,
      "no_speech_prob": 1.3416833616330948e-11,
      "compression_ratio": 1.1842900302114803
    },
    {
      "start": 622.466,
      "end": 623.686,
      "text": "什么时候放行",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11474531060023023,
      "no_speech_prob": 1.3416833616330948e-11,
      "compression_ratio": 1.1842900302114803
    },
    {
      "start": 623.686,
      "end": 624.626,
      "text": "什么时候升级",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11474531060023023,
      "no_speech_prob": 1.3416833616330948e-11,
      "compression_ratio": 1.1842900302114803
    },
    {
      "start": 624.626,
      "end": 627.806,
      "text": "本来就可以在应用代表里面写Threathood",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11474531060023023,
      "no_speech_prob": 1.3416833616330948e-11,
      "compression_ratio": 1.1842900302114803
    },
    {
      "start": 627.806,
      "end": 631.066,
      "text": "所以他最后的评价其实很有意思",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11474531060023023,
      "no_speech_prob": 1.3416833616330948e-11,
      "compression_ratio": 1.1842900302114803
    },
    {
      "start": 631.066,
      "end": 633.566,
      "text": "他认为这些统计学上的零件",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11474531060023023,
      "no_speech_prob": 1.3416833616330948e-11,
      "compression_ratio": 1.1842900302114803
    },
    {
      "start": 633.566,
      "end": 636.166,
      "text": "没有哪一个是突然从天上掉下来的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11474531060023023,
      "no_speech_prob": 1.3416833616330948e-11,
      "compression_ratio": 1.1842900302114803
    },
    {
      "start": 636.166,
      "end": 641.306,
      "text": "真正可能不同的是TypeSafe把校准放进了训练目标",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11474531060023023,
      "no_speech_prob": 1.3416833616330948e-11,
      "compression_ratio": 1.1842900302114803
    },
    {
      "start": 641.306,
      "end": 644.966,
      "text": "然后让API从一开始返回的就是Decision",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11474531060023023,
      "no_speech_prob": 1.3416833616330948e-11,
      "compression_ratio": 1.1842900302114803
    },
    {
      "start": 644.966,
      "end": 648.186,
      "text": "而不是一段等待程序再次解析的文字",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10132473568583644,
      "no_speech_prob": 1.2909852006859346e-11,
      "compression_ratio": 1.2623456790123457
    },
    {
      "start": 648.186,
      "end": 649.146,
      "text": "换句话说",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10132473568583644,
      "no_speech_prob": 1.2909852006859346e-11,
      "compression_ratio": 1.2623456790123457
    },
    {
      "start": 649.146,
      "end": 652.306,
      "text": "就不是发明了一堆从来没有存在过的零件",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10132473568583644,
      "no_speech_prob": 1.2909852006859346e-11,
      "compression_ratio": 1.2623456790123457
    },
    {
      "start": 652.306,
      "end": 654.546,
      "text": "而是少兵了几层东西",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10132473568583644,
      "no_speech_prob": 1.2909852006859346e-11,
      "compression_ratio": 1.2623456790123457
    },
    {
      "start": 654.546,
      "end": 656.746,
      "text": "这时候Jav就不是分类器吗",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10132473568583644,
      "no_speech_prob": 1.2909852006859346e-11,
      "compression_ratio": 1.2623456790123457
    },
    {
      "start": 656.746,
      "end": 659.146,
      "text": "这个问题就开始变得没那么简单了",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10132473568583644,
      "no_speech_prob": 1.2909852006859346e-11,
      "compression_ratio": 1.2623456790123457
    },
    {
      "start": 659.146,
      "end": 662.006,
      "text": "因为TypeSafe真正想挑战的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10132473568583644,
      "no_speech_prob": 1.2909852006859346e-11,
      "compression_ratio": 1.2623456790123457
    },
    {
      "start": 662.006,
      "end": 664.866,
      "text": "并不是传统分类器做不到的分类",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10132473568583644,
      "no_speech_prob": 1.2909852006859346e-11,
      "compression_ratio": 1.2623456790123457
    },
    {
      "start": 664.866,
      "end": 666.426,
      "text": "而是它的使用方式",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10132473568583644,
      "no_speech_prob": 1.2909852006859346e-11,
      "compression_ratio": 1.2623456790123457
    },
    {
      "start": 666.426,
      "end": 669.766,
      "text": "假设你今天做一个金融agent",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10132473568583644,
      "no_speech_prob": 1.2909852006859346e-11,
      "compression_ratio": 1.2623456790123457
    },
    {
      "start": 669.766,
      "end": 671.366,
      "text": "你只需要判断的是",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10132473568583644,
      "no_speech_prob": 1.2909852006859346e-11,
      "compression_ratio": 1.2623456790123457
    },
    {
      "start": 671.366,
      "end": 673.626,
      "text": "这笔交易有没有欺诈风险",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10132473568583644,
      "no_speech_prob": 1.2909852006859346e-11,
      "compression_ratio": 1.2623456790123457
    },
    {
      "start": 673.626,
      "end": 677.166,
      "text": "那你当然可以收集数据标注样本",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0648686220479566,
      "no_speech_prob": 1.2942849916819377e-11,
      "compression_ratio": 1.2763157894736843
    },
    {
      "start": 677.166,
      "end": 678.726,
      "text": "训练一个专门的分类器",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0648686220479566,
      "no_speech_prob": 1.2942849916819377e-11,
      "compression_ratio": 1.2763157894736843
    },
    {
      "start": 678.726,
      "end": 680.486,
      "text": "如果业务一直不变",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0648686220479566,
      "no_speech_prob": 1.2942849916819377e-11,
      "compression_ratio": 1.2763157894736843
    },
    {
      "start": 680.486,
      "end": 683.226,
      "text": "这就可能比调用GF更便宜更快",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0648686220479566,
      "no_speech_prob": 1.2942849916819377e-11,
      "compression_ratio": 1.2763157894736843
    },
    {
      "start": 683.226,
      "end": 684.366,
      "text": "甚至更可靠",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0648686220479566,
      "no_speech_prob": 1.2942849916819377e-11,
      "compression_ratio": 1.2763157894736843
    },
    {
      "start": 684.366,
      "end": 688.206,
      "text": "但现实里的agent很少永远只问同一个问题",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0648686220479566,
      "no_speech_prob": 1.2942849916819377e-11,
      "compression_ratio": 1.2763157894736843
    },
    {
      "start": 688.206,
      "end": 690.346,
      "text": "今天你要判断的是欺诈",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0648686220479566,
      "no_speech_prob": 1.2942849916819377e-11,
      "compression_ratio": 1.2763157894736843
    },
    {
      "start": 690.346,
      "end": 693.086,
      "text": "明天可能增加是否需要人工复合",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0648686220479566,
      "no_speech_prob": 1.2942849916819377e-11,
      "compression_ratio": 1.2763157894736843
    },
    {
      "start": 693.086,
      "end": 694.826,
      "text": "后天可能又增加",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0648686220479566,
      "no_speech_prob": 1.2942849916819377e-11,
      "compression_ratio": 1.2763157894736843
    },
    {
      "start": 694.826,
      "end": 696.906,
      "text": "应该交给哪个风险团队",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0648686220479566,
      "no_speech_prob": 1.2942849916819377e-11,
      "compression_ratio": 1.2763157894736843
    },
    {
      "start": 696.906,
      "end": 698.026,
      "text": "换一个产品",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0648686220479566,
      "no_speech_prob": 1.2942849916819377e-11,
      "compression_ratio": 1.2763157894736843
    },
    {
      "start": 698.026,
      "end": 702.006,
      "text": "问题标签和业务规则可能全部都变了",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0648686220479566,
      "no_speech_prob": 1.2942849916819377e-11,
      "compression_ratio": 1.2763157894736843
    },
    {
      "start": 702.006,
      "end": 703.866,
      "text": "如果每增加一种判断",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.12092739650181361,
      "no_speech_prob": 1.3743011069433653e-11,
      "compression_ratio": 1.2393939393939395
    },
    {
      "start": 703.866,
      "end": 705.486,
      "text": "都要重新积累数据",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.12092739650181361,
      "no_speech_prob": 1.3743011069433653e-11,
      "compression_ratio": 1.2393939393939395
    },
    {
      "start": 705.486,
      "end": 707.046,
      "text": "训练和部署一个模型",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.12092739650181361,
      "no_speech_prob": 1.3743011069433653e-11,
      "compression_ratio": 1.2393939393939395
    },
    {
      "start": 707.046,
      "end": 709.206,
      "text": "那传统分类器的成本",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.12092739650181361,
      "no_speech_prob": 1.3743011069433653e-11,
      "compression_ratio": 1.2393939393939395
    },
    {
      "start": 709.206,
      "end": 711.066,
      "text": "就开始从推理阶段",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.12092739650181361,
      "no_speech_prob": 1.3743011069433653e-11,
      "compression_ratio": 1.2393939393939395
    },
    {
      "start": 711.066,
      "end": 713.526,
      "text": "转移到了开发和维护阶段",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.12092739650181361,
      "no_speech_prob": 1.3743011069433653e-11,
      "compression_ratio": 1.2393939393939395
    },
    {
      "start": 713.526,
      "end": 715.046,
      "text": "而Jeff想做的是把",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.12092739650181361,
      "no_speech_prob": 1.3743011069433653e-11,
      "compression_ratio": 1.2393939393939395
    },
    {
      "start": 715.046,
      "end": 718.566,
      "text": "大语言模型最有价值的一部分能力保留下来",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.12092739650181361,
      "no_speech_prob": 1.3743011069433653e-11,
      "compression_ratio": 1.2393939393939395
    },
    {
      "start": 718.566,
      "end": 721.306,
      "text": "就是你可以直接用自然语言告诉模型",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.12092739650181361,
      "no_speech_prob": 1.3743011069433653e-11,
      "compression_ratio": 1.2393939393939395
    },
    {
      "start": 721.306,
      "end": 722.866,
      "text": "现在要判断什么",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.12092739650181361,
      "no_speech_prob": 1.3743011069433653e-11,
      "compression_ratio": 1.2393939393939395
    },
    {
      "start": 722.866,
      "end": 726.546,
      "text": "那我们目前还没有看到",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.12092739650181361,
      "no_speech_prob": 1.3743011069433653e-11,
      "compression_ratio": 1.2393939393939395
    },
    {
      "start": 726.546,
      "end": 728.926,
      "text": "Jeff和一个认真微调过的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.12092739650181361,
      "no_speech_prob": 1.3743011069433653e-11,
      "compression_ratio": 1.2393939393939395
    },
    {
      "start": 728.926,
      "end": 729.786,
      "text": "Martin Burrton",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.12092739650181361,
      "no_speech_prob": 1.3743011069433653e-11,
      "compression_ratio": 1.2393939393939395
    },
    {
      "start": 729.786,
      "end": 731.826,
      "text": "专门的小型分类器",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.12092739650181361,
      "no_speech_prob": 1.3743011069433653e-11,
      "compression_ratio": 1.2393939393939395
    },
    {
      "start": 731.826,
      "end": 735.086,
      "text": "或者一个使用受限解码的小语言模型",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07180780317725205,
      "no_speech_prob": 1.4059521775966477e-11,
      "compression_ratio": 1.3104575163398693
    },
    {
      "start": 735.086,
      "end": 739.026,
      "text": "在同样任务同样数据同样准确率要求下",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07180780317725205,
      "no_speech_prob": 1.4059521775966477e-11,
      "compression_ratio": 1.3104575163398693
    },
    {
      "start": 739.026,
      "end": 741.586,
      "text": "做一次真正公平的正面对比",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07180780317725205,
      "no_speech_prob": 1.4059521775966477e-11,
      "compression_ratio": 1.3104575163398693
    },
    {
      "start": 741.586,
      "end": 744.026,
      "text": "TypeSafe真正需要证明的是",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07180780317725205,
      "no_speech_prob": 1.4059521775966477e-11,
      "compression_ratio": 1.3104575163398693
    },
    {
      "start": 744.026,
      "end": 745.746,
      "text": "当任务不断变化",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07180780317725205,
      "no_speech_prob": 1.4059521775966477e-11,
      "compression_ratio": 1.3104575163398693
    },
    {
      "start": 745.746,
      "end": 748.766,
      "text": "而你又不想每次重新训练一个模型的时候",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07180780317725205,
      "no_speech_prob": 1.4059521775966477e-11,
      "compression_ratio": 1.3104575163398693
    },
    {
      "start": 748.766,
      "end": 752.466,
      "text": "Jab能不能同时保住分类器的速度和成本",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07180780317725205,
      "no_speech_prob": 1.4059521775966477e-11,
      "compression_ratio": 1.3104575163398693
    },
    {
      "start": 752.466,
      "end": 755.886,
      "text": "又同时保住大模型那种拿自然语言",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07180780317725205,
      "no_speech_prob": 1.4059521775966477e-11,
      "compression_ratio": 1.3104575163398693
    },
    {
      "start": 755.886,
      "end": 758.186,
      "text": "就能直接定义新任务的灵活性",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07180780317725205,
      "no_speech_prob": 1.4059521775966477e-11,
      "compression_ratio": 1.3104575163398693
    },
    {
      "start": 758.186,
      "end": 760.126,
      "text": "而这还只是第一层",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07180780317725205,
      "no_speech_prob": 1.4059521775966477e-11,
      "compression_ratio": 1.3104575163398693
    },
    {
      "start": 760.126,
      "end": 762.826,
      "text": "因为Type Save对Jav最大胆的主张",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09684740822270231,
      "no_speech_prob": 1.1928397505855948e-11,
      "compression_ratio": 1.2341772151898733
    },
    {
      "start": 762.826,
      "end": 765.006,
      "text": "其实不是快也不是便宜",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09684740822270231,
      "no_speech_prob": 1.1928397505855948e-11,
      "compression_ratio": 1.2341772151898733
    },
    {
      "start": 765.006,
      "end": 767.166,
      "text": "甚至不是自然语言可编程",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09684740822270231,
      "no_speech_prob": 1.1928397505855948e-11,
      "compression_ratio": 1.2341772151898733
    },
    {
      "start": 767.166,
      "end": 768.606,
      "text": "而是他认为",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09684740822270231,
      "no_speech_prob": 1.1928397505855948e-11,
      "compression_ratio": 1.2341772151898733
    },
    {
      "start": 768.606,
      "end": 770.766,
      "text": "Jav不只会给你一个决定",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09684740822270231,
      "no_speech_prob": 1.1928397505855948e-11,
      "compression_ratio": 1.2341772151898733
    },
    {
      "start": 770.766,
      "end": 774.486,
      "text": "他还知道自己这个决定到底有多可靠",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09684740822270231,
      "no_speech_prob": 1.1928397505855948e-11,
      "compression_ratio": 1.2341772151898733
    },
    {
      "start": 774.486,
      "end": 776.666,
      "text": "这就进入了整个Jav的故事里",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09684740822270231,
      "no_speech_prob": 1.1928397505855948e-11,
      "compression_ratio": 1.2341772151898733
    },
    {
      "start": 776.666,
      "end": 779.126,
      "text": "最关键也是目前最神秘的一块",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09684740822270231,
      "no_speech_prob": 1.1928397505855948e-11,
      "compression_ratio": 1.2341772151898733
    },
    {
      "start": 779.126,
      "end": 780.206,
      "text": "RLCD",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09684740822270231,
      "no_speech_prob": 1.1928397505855948e-11,
      "compression_ratio": 1.2341772151898733
    },
    {
      "start": 780.206,
      "end": 784.146,
      "text": "前面讲到Jav会给每一个决策附上概率",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09684740822270231,
      "no_speech_prob": 1.1928397505855948e-11,
      "compression_ratio": 1.2341772151898733
    },
    {
      "start": 784.146,
      "end": 785.926,
      "text": "但这里真正重要的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09684740822270231,
      "no_speech_prob": 1.1928397505855948e-11,
      "compression_ratio": 1.2341772151898733
    },
    {
      "start": 785.926,
      "end": 789.526,
      "text": "不是他能不能输出一个0.73或者0.91",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.09684740822270231,
      "no_speech_prob": 1.1928397505855948e-11,
      "compression_ratio": 1.2341772151898733
    },
    {
      "start": 789.526,
      "end": 793.126,
      "text": "而是说这些数字到底有没有统计意义",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07337602418044518,
      "no_speech_prob": 1.3152636098301418e-11,
      "compression_ratio": 1.2831541218637992
    },
    {
      "start": 793.126,
      "end": 795.606,
      "text": "这就是RLCD想解决的问题",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07337602418044518,
      "no_speech_prob": 1.3152636098301418e-11,
      "compression_ratio": 1.2831541218637992
    },
    {
      "start": 795.606,
      "end": 799.306,
      "text": "它的全称是校准决策强化学习",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07337602418044518,
      "no_speech_prob": 1.3152636098301418e-11,
      "compression_ratio": 1.2831541218637992
    },
    {
      "start": 799.306,
      "end": 800.286,
      "text": "所谓校准",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07337602418044518,
      "no_speech_prob": 1.3152636098301418e-11,
      "compression_ratio": 1.2831541218637992
    },
    {
      "start": 800.286,
      "end": 802.806,
      "text": "我们可以先用一个很简单的例子来理解",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07337602418044518,
      "no_speech_prob": 1.3152636098301418e-11,
      "compression_ratio": 1.2831541218637992
    },
    {
      "start": 802.806,
      "end": 806.446,
      "text": "就是如果一个模型做了很多次判断",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07337602418044518,
      "no_speech_prob": 1.3152636098301418e-11,
      "compression_ratio": 1.2831541218637992
    },
    {
      "start": 806.446,
      "end": 810.806,
      "text": "我们把它所有认为有70%把握的判断放在一起",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07337602418044518,
      "no_speech_prob": 1.3152636098301418e-11,
      "compression_ratio": 1.2831541218637992
    },
    {
      "start": 810.806,
      "end": 813.046,
      "text": "那么一个校准的好的模型",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07337602418044518,
      "no_speech_prob": 1.3152636098301418e-11,
      "compression_ratio": 1.2831541218637992
    },
    {
      "start": 813.046,
      "end": 816.766,
      "text": "长期来看应该有大约70%是正确的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.07337602418044518,
      "no_speech_prob": 1.3152636098301418e-11,
      "compression_ratio": 1.2831541218637992
    },
    {
      "start": 816.766,
      "end": 819.806,
      "text": "他说90%就应该接近90%",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08424971854850037,
      "no_speech_prob": 1.3147322640294501e-11,
      "compression_ratio": 1.3664383561643836
    },
    {
      "start": 819.806,
      "end": 822.406,
      "text": "说50%就真的意味着",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08424971854850037,
      "no_speech_prob": 1.3147322640294501e-11,
      "compression_ratio": 1.3664383561643836
    },
    {
      "start": 822.406,
      "end": 824.866,
      "text": "这件事差不多只有一半的把握",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08424971854850037,
      "no_speech_prob": 1.3147322640294501e-11,
      "compression_ratio": 1.3664383561643836
    },
    {
      "start": 824.866,
      "end": 827.246,
      "text": "这和准确率其实是两回事",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08424971854850037,
      "no_speech_prob": 1.3147322640294501e-11,
      "compression_ratio": 1.3664383561643836
    },
    {
      "start": 827.246,
      "end": 830.846,
      "text": "一个模型完全可能总体准确率很高",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08424971854850037,
      "no_speech_prob": 1.3147322640294501e-11,
      "compression_ratio": 1.3664383561643836
    },
    {
      "start": 830.846,
      "end": 833.406,
      "text": "但他根本不知道自己什么时候会错",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08424971854850037,
      "no_speech_prob": 1.3147322640294501e-11,
      "compression_ratio": 1.3664383561643836
    },
    {
      "start": 833.406,
      "end": 837.306,
      "text": "而另一个模型也许平均准确率没有高出多少",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08424971854850037,
      "no_speech_prob": 1.3147322640294501e-11,
      "compression_ratio": 1.3664383561643836
    },
    {
      "start": 837.306,
      "end": 839.306,
      "text": "但他在容易出错的时候",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08424971854850037,
      "no_speech_prob": 1.3147322640294501e-11,
      "compression_ratio": 1.3664383561643836
    },
    {
      "start": 839.306,
      "end": 841.486,
      "text": "真的会把自己的概率降下来",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08424971854850037,
      "no_speech_prob": 1.3147322640294501e-11,
      "compression_ratio": 1.3664383561643836
    },
    {
      "start": 841.486,
      "end": 843.406,
      "text": "对于聊天机器人来说",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08424971854850037,
      "no_speech_prob": 1.3147322640294501e-11,
      "compression_ratio": 1.3664383561643836
    },
    {
      "start": 843.406,
      "end": 845.666,
      "text": "这个区别可能没有那么明显",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08424971854850037,
      "no_speech_prob": 1.3147322640294501e-11,
      "compression_ratio": 1.3664383561643836
    },
    {
      "start": 845.666,
      "end": 847.206,
      "text": "但对于agent来说",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08377061999818813,
      "no_speech_prob": 1.2358440659721026e-11,
      "compression_ratio": 1.2525252525252526
    },
    {
      "start": 847.206,
      "end": 848.966,
      "text": "这就直接决定了一件事",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08377061999818813,
      "no_speech_prob": 1.2358440659721026e-11,
      "compression_ratio": 1.2525252525252526
    },
    {
      "start": 848.966,
      "end": 852.526,
      "text": "就是这个概率能不能拿来控制系统",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08377061999818813,
      "no_speech_prob": 1.2358440659721026e-11,
      "compression_ratio": 1.2525252525252526
    },
    {
      "start": 852.526,
      "end": 854.266,
      "text": "比如一个agent正在决定",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08377061999818813,
      "no_speech_prob": 1.2358440659721026e-11,
      "compression_ratio": 1.2525252525252526
    },
    {
      "start": 854.266,
      "end": 856.266,
      "text": "要不要自动批准退款",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08377061999818813,
      "no_speech_prob": 1.2358440659721026e-11,
      "compression_ratio": 1.2525252525252526
    },
    {
      "start": 856.266,
      "end": 858.106,
      "text": "债务判断可以批准",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08377061999818813,
      "no_speech_prob": 1.2358440659721026e-11,
      "compression_ratio": 1.2525252525252526
    },
    {
      "start": 858.106,
      "end": 860.066,
      "text": "概率99.8%",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08377061999818813,
      "no_speech_prob": 1.2358440659721026e-11,
      "compression_ratio": 1.2525252525252526
    },
    {
      "start": 860.066,
      "end": 862.406,
      "text": "如果经过大量测试之后",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08377061999818813,
      "no_speech_prob": 1.2358440659721026e-11,
      "compression_ratio": 1.2525252525252526
    },
    {
      "start": 862.406,
      "end": 865.666,
      "text": "你真的知道这种99.8%值得性",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08377061999818813,
      "no_speech_prob": 1.2358440659721026e-11,
      "compression_ratio": 1.2525252525252526
    },
    {
      "start": 865.666,
      "end": 868.346,
      "text": "那么harness就可以设置一个预值",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08377061999818813,
      "no_speech_prob": 1.2358440659721026e-11,
      "compression_ratio": 1.2525252525252526
    },
    {
      "start": 868.346,
      "end": 870.506,
      "text": "超过这个概率自动执行",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08377061999818813,
      "no_speech_prob": 1.2358440659721026e-11,
      "compression_ratio": 1.2525252525252526
    },
    {
      "start": 870.506,
      "end": 872.786,
      "text": "但如果只有70%",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08377061999818813,
      "no_speech_prob": 1.2358440659721026e-11,
      "compression_ratio": 1.2525252525252526
    },
    {
      "start": 872.786,
      "end": 875.166,
      "text": "就升级给更强更贵的模型",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08377061999818813,
      "no_speech_prob": 1.2358440659721026e-11,
      "compression_ratio": 1.2525252525252526
    },
    {
      "start": 875.166,
      "end": 878.506,
      "text": "如果再低一点就交给人工或者verify",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.15101958268525584,
      "no_speech_prob": 1.4443218322723883e-11,
      "compression_ratio": 1.2880794701986755
    },
    {
      "start": 878.506,
      "end": 883.066,
      "text": "这个时候CoundEfence就不再是给人看的一个装饰数字",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.15101958268525584,
      "no_speech_prob": 1.4443218322723883e-11,
      "compression_ratio": 1.2880794701986755
    },
    {
      "start": 883.066,
      "end": 886.646,
      "text": "它变成了Agent的内部的一种控制信号",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.15101958268525584,
      "no_speech_prob": 1.4443218322723883e-11,
      "compression_ratio": 1.2880794701986755
    },
    {
      "start": 886.646,
      "end": 888.566,
      "text": "就是什么时候继续",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.15101958268525584,
      "no_speech_prob": 1.4443218322723883e-11,
      "compression_ratio": 1.2880794701986755
    },
    {
      "start": 888.566,
      "end": 889.606,
      "text": "什么时候升级",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.15101958268525584,
      "no_speech_prob": 1.4443218322723883e-11,
      "compression_ratio": 1.2880794701986755
    },
    {
      "start": 889.606,
      "end": 890.766,
      "text": "什么时候验证",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.15101958268525584,
      "no_speech_prob": 1.4443218322723883e-11,
      "compression_ratio": 1.2880794701986755
    },
    {
      "start": 890.766,
      "end": 892.066,
      "text": "什么时候停下来",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.15101958268525584,
      "no_speech_prob": 1.4443218322723883e-11,
      "compression_ratio": 1.2880794701986755
    },
    {
      "start": 892.066,
      "end": 895.106,
      "text": "前面那个满分信心却答错的例子",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.15101958268525584,
      "no_speech_prob": 1.4443218322723883e-11,
      "compression_ratio": 1.2880794701986755
    },
    {
      "start": 895.106,
      "end": 896.706,
      "text": "追问的也正是这件事",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.15101958268525584,
      "no_speech_prob": 1.4443218322723883e-11,
      "compression_ratio": 1.2880794701986755
    },
    {
      "start": 896.706,
      "end": 899.866,
      "text": "就是这些概率究竟有多可信呢",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.15101958268525584,
      "no_speech_prob": 1.4443218322723883e-11,
      "compression_ratio": 1.2880794701986755
    },
    {
      "start": 899.866,
      "end": 903.406,
      "text": "而RLCD真正生成改变的是另一件事",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.15101958268525584,
      "no_speech_prob": 1.4443218322723883e-11,
      "compression_ratio": 1.2880794701986755
    },
    {
      "start": 903.406,
      "end": 905.666,
      "text": "就是从训练阶段开始",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0577942560303886,
      "no_speech_prob": 7.954285667632899e-12,
      "compression_ratio": 1.0
    },
    {
      "start": 905.666,
      "end": 908.966,
      "text": "就让模型学习一个决定应该有多确定",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0577942560303886,
      "no_speech_prob": 7.954285667632899e-12,
      "compression_ratio": 1.0
    },
    {
      "start": 908.966,
      "end": 912.446,
      "text": "让概率本身也成为训练目标的一部分",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.0577942560303886,
      "no_speech_prob": 7.954285667632899e-12,
      "compression_ratio": 1.0
    },
    {
      "start": 912.892,
      "end": 923.892,
      "text": "但是,我目前還沒有看到TypeSafe公開完整的論文,具體的訓練流程和損失函數,也沒有足夠完整的第三方校準基準。",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.15514488557798672,
      "no_speech_prob": 8.825419561819814e-12,
      "compression_ratio": 1.1764705882352942
    },
    {
      "start": 923.892,
      "end": 928.552,
      "text": "所以RLCD是否真的比现有校准方法更好",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08066615575476538,
      "no_speech_prob": 1.2020963217895808e-11,
      "compression_ratio": 1.1949685534591195
    },
    {
      "start": 928.552,
      "end": 930.952,
      "text": "这一部分还要等更多的证据",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08066615575476538,
      "no_speech_prob": 1.2020963217895808e-11,
      "compression_ratio": 1.1949685534591195
    },
    {
      "start": 930.952,
      "end": 933.752,
      "text": "这里也刚好可以解释TypeSafe另一个",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08066615575476538,
      "no_speech_prob": 1.2020963217895808e-11,
      "compression_ratio": 1.1949685534591195
    },
    {
      "start": 933.752,
      "end": 935.472,
      "text": "很容易让人误解的说法",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08066615575476538,
      "no_speech_prob": 1.2020963217895808e-11,
      "compression_ratio": 1.1949685534591195
    },
    {
      "start": 935.472,
      "end": 938.212,
      "text": "就是Jav不能产生幻觉",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08066615575476538,
      "no_speech_prob": 1.2020963217895808e-11,
      "compression_ratio": 1.1949685534591195
    },
    {
      "start": 938.212,
      "end": 942.852,
      "text": "如果你只允许Jav在批准和拒绝之间选择",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08066615575476538,
      "no_speech_prob": 1.2020963217895808e-11,
      "compression_ratio": 1.1949685534591195
    },
    {
      "start": 942.852,
      "end": 945.252,
      "text": "他确实不可能突然回答",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08066615575476538,
      "no_speech_prob": 1.2020963217895808e-11,
      "compression_ratio": 1.1949685534591195
    },
    {
      "start": 945.252,
      "end": 948.332,
      "text": "也许也不会捏造第三个不存在的函数",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08066615575476538,
      "no_speech_prob": 1.2020963217895808e-11,
      "compression_ratio": 1.1949685534591195
    },
    {
      "start": 948.332,
      "end": 951.012,
      "text": "更不会生成一段格式错误程序",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08066615575476538,
      "no_speech_prob": 1.2020963217895808e-11,
      "compression_ratio": 1.1949685534591195
    },
    {
      "start": 951.012,
      "end": 952.572,
      "text": "根本解析不了的Jason",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08066615575476538,
      "no_speech_prob": 1.2020963217895808e-11,
      "compression_ratio": 1.1949685534591195
    },
    {
      "start": 952.572,
      "end": 955.472,
      "text": "它的输出空间在计算之前",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09537408306340503,
      "no_speech_prob": 1.305707712090376e-11,
      "compression_ratio": 1.2842809364548495
    },
    {
      "start": 955.472,
      "end": 958.072,
      "text": "就已经被系统类型限制好了",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09537408306340503,
      "no_speech_prob": 1.305707712090376e-11,
      "compression_ratio": 1.2842809364548495
    },
    {
      "start": 958.072,
      "end": 961.132,
      "text": "但是它完全可能选择批准",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09537408306340503,
      "no_speech_prob": 1.305707712090376e-11,
      "compression_ratio": 1.2842809364548495
    },
    {
      "start": 961.132,
      "end": 963.252,
      "text": "而正确答案其实是拒绝",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09537408306340503,
      "no_speech_prob": 1.305707712090376e-11,
      "compression_ratio": 1.2842809364548495
    },
    {
      "start": 963.252,
      "end": 965.012,
      "text": "所以更准确的说",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09537408306340503,
      "no_speech_prob": 1.305707712090376e-11,
      "compression_ratio": 1.2842809364548495
    },
    {
      "start": 965.012,
      "end": 967.572,
      "text": "Jav解决的是一种非常具体的幻觉",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09537408306340503,
      "no_speech_prob": 1.305707712090376e-11,
      "compression_ratio": 1.2842809364548495
    },
    {
      "start": 967.572,
      "end": 969.892,
      "text": "它不会生成类型之外的答案",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09537408306340503,
      "no_speech_prob": 1.305707712090376e-11,
      "compression_ratio": 1.2842809364548495
    },
    {
      "start": 969.892,
      "end": 973.252,
      "text": "但仍然可能在合法答案里面选错",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09537408306340503,
      "no_speech_prob": 1.305707712090376e-11,
      "compression_ratio": 1.2842809364548495
    },
    {
      "start": 973.252,
      "end": 974.872,
      "text": "不过呢",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09537408306340503,
      "no_speech_prob": 1.305707712090376e-11,
      "compression_ratio": 1.2842809364548495
    },
    {
      "start": 974.872,
      "end": 978.412,
      "text": "真正决定Jav能不能进入高风险自动化流程的",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09537408306340503,
      "no_speech_prob": 1.305707712090376e-11,
      "compression_ratio": 1.2842809364548495
    },
    {
      "start": 978.412,
      "end": 979.672,
      "text": "还是后一个问题",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09537408306340503,
      "no_speech_prob": 1.305707712090376e-11,
      "compression_ratio": 1.2842809364548495
    },
    {
      "start": 979.672,
      "end": 981.292,
      "text": "就是它选错的时候",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09537408306340503,
      "no_speech_prob": 1.305707712090376e-11,
      "compression_ratio": 1.2842809364548495
    },
    {
      "start": 981.292,
      "end": 982.812,
      "text": "自己到底知不知道",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10096795823838976,
      "no_speech_prob": 1.3520718399051557e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 982.812,
      "end": 984.652,
      "text": "这也是为什么我觉得",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10096795823838976,
      "no_speech_prob": 1.3520718399051557e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 984.652,
      "end": 988.792,
      "text": "RLCD反而比Jav那个200倍更快的数字",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10096795823838976,
      "no_speech_prob": 1.3520718399051557e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 988.792,
      "end": 989.632,
      "text": "更值得关注",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10096795823838976,
      "no_speech_prob": 1.3520718399051557e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 989.632,
      "end": 992.572,
      "text": "如果一个模型只是快而且便宜",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10096795823838976,
      "no_speech_prob": 1.3520718399051557e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 992.572,
      "end": 995.232,
      "text": "我们其实已经有很多办法可以做到了",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10096795823838976,
      "no_speech_prob": 1.3520718399051557e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 995.232,
      "end": 999.272,
      "text": "但如果它能在保持这种速度和成本的同时",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10096795823838976,
      "no_speech_prob": 1.3520718399051557e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 999.272,
      "end": 1003.612,
      "text": "给软件提供一个真正可信的不确定性信号",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10096795823838976,
      "no_speech_prob": 1.3520718399051557e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 1003.612,
      "end": 1006.552,
      "text": "那它在A层里面的位置就完全不一样了",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10096795823838976,
      "no_speech_prob": 1.3520718399051557e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 1006.552,
      "end": 1009.332,
      "text": "当然接下来还有一个跟现实的问题",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10096795823838976,
      "no_speech_prob": 1.3520718399051557e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 1009.332,
      "end": 1013.932,
      "text": "就是这些概率决策究竟能省下多少时间和成本呢",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.1034565544128418,
      "no_speech_prob": 1.524146434062068e-11,
      "compression_ratio": 1.1428571428571428
    },
    {
      "start": 1013.932,
      "end": 1016.272,
      "text": "现在我们终于可以回到",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.1034565544128418,
      "no_speech_prob": 1.524146434062068e-11,
      "compression_ratio": 1.1428571428571428
    },
    {
      "start": 1016.272,
      "end": 1018.992,
      "text": "JF发布时最新眼球的那组数字",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.1034565544128418,
      "no_speech_prob": 1.524146434062068e-11,
      "compression_ratio": 1.1428571428571428
    },
    {
      "start": 1018.992,
      "end": 1021.152,
      "text": "20到200倍更快",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.1034565544128418,
      "no_speech_prob": 1.524146434062068e-11,
      "compression_ratio": 1.1428571428571428
    },
    {
      "start": 1021.152,
      "end": 1023.492,
      "text": "40到400倍更便宜",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.1034565544128418,
      "no_speech_prob": 1.524146434062068e-11,
      "compression_ratio": 1.1428571428571428
    },
    {
      "start": 1023.492,
      "end": 1026.532,
      "text": "TypeSafe自己公布的Workflow Benchmark里面",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.1034565544128418,
      "no_speech_prob": 1.524146434062068e-11,
      "compression_ratio": 1.1428571428571428
    },
    {
      "start": 1026.532,
      "end": 1030.312,
      "text": "最高端的数字达到了193.6倍更快",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.1034565544128418,
      "no_speech_prob": 1.524146434062068e-11,
      "compression_ratio": 1.1428571428571428
    },
    {
      "start": 1030.312,
      "end": 1033.032,
      "text": "444.6倍更便宜",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.1034565544128418,
      "no_speech_prob": 1.524146434062068e-11,
      "compression_ratio": 1.1428571428571428
    },
    {
      "start": 1033.032,
      "end": 1035.452,
      "text": "这个数字第一眼看起来很离谱",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.1034565544128418,
      "no_speech_prob": 1.524146434062068e-11,
      "compression_ratio": 1.1428571428571428
    },
    {
      "start": 1035.452,
      "end": 1037.832,
      "text": "但离着JF前面的设计之后",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.1034565544128418,
      "no_speech_prob": 1.524146434062068e-11,
      "compression_ratio": 1.1428571428571428
    },
    {
      "start": 1037.832,
      "end": 1039.972,
      "text": "你会发现它其实并不神秘",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10925357313041227,
      "no_speech_prob": 1.4120786137605812e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 1039.972,
      "end": 1043.372,
      "text": "因为Jav和普通LM做计算的方式",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10925357313041227,
      "no_speech_prob": 1.4120786137605812e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 1043.372,
      "end": 1045.072,
      "text": "从一开始就不一样",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10925357313041227,
      "no_speech_prob": 1.4120786137605812e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 1045.072,
      "end": 1048.332,
      "text": "假设一个agent现在需要做20个判断",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10925357313041227,
      "no_speech_prob": 1.4120786137605812e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 1048.332,
      "end": 1050.652,
      "text": "比如用户是不是在退款",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10925357313041227,
      "no_speech_prob": 1.4120786137605812e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 1050.652,
      "end": 1052.112,
      "text": "有没有欺诈风险",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10925357313041227,
      "no_speech_prob": 1.4120786137605812e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 1052.112,
      "end": 1053.532,
      "text": "需不需要人工介入",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10925357313041227,
      "no_speech_prob": 1.4120786137605812e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 1053.532,
      "end": 1055.032,
      "text": "需要调用哪个工具",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10925357313041227,
      "no_speech_prob": 1.4120786137605812e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 1055.032,
      "end": 1057.252,
      "text": "结果要不要再验证一次",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10925357313041227,
      "no_speech_prob": 1.4120786137605812e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 1057.252,
      "end": 1060.372,
      "text": "阿尔梅达回应过一个很直接的问题",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10925357313041227,
      "no_speech_prob": 1.4120786137605812e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 1060.372,
      "end": 1063.072,
      "text": "就是为什么不干脆用受限解码呢",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10925357313041227,
      "no_speech_prob": 1.4120786137605812e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 1063.072,
      "end": 1064.312,
      "text": "他的回答是",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10925357313041227,
      "no_speech_prob": 1.4120786137605812e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 1064.312,
      "end": 1067.072,
      "text": "这层约束仍然叠在LM上",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.10925357313041227,
      "no_speech_prob": 1.4120786137605812e-11,
      "compression_ratio": 1.228395061728395
    },
    {
      "start": 1067.072,
      "end": 1069.332,
      "text": "字符串生成的开销还在",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09138413911224694,
      "no_speech_prob": 1.4911225032498976e-11,
      "compression_ratio": 1.3123028391167193
    },
    {
      "start": 1069.332,
      "end": 1071.592,
      "text": "他还提出了一更尖锐的判断",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09138413911224694,
      "no_speech_prob": 1.4911225032498976e-11,
      "compression_ratio": 1.3123028391167193
    },
    {
      "start": 1071.592,
      "end": 1074.172,
      "text": "就是模型本来就拿不准的时候",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09138413911224694,
      "no_speech_prob": 1.4911225032498976e-11,
      "compression_ratio": 1.3123028391167193
    },
    {
      "start": 1074.172,
      "end": 1076.552,
      "text": "再去限制他能选哪些输出",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09138413911224694,
      "no_speech_prob": 1.4911225032498976e-11,
      "compression_ratio": 1.3123028391167193
    },
    {
      "start": 1076.552,
      "end": 1078.152,
      "text": "反而会让模型变笨",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09138413911224694,
      "no_speech_prob": 1.4911225032498976e-11,
      "compression_ratio": 1.3123028391167193
    },
    {
      "start": 1078.152,
      "end": 1081.332,
      "text": "债务直接省掉了字符串生成这一步",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09138413911224694,
      "no_speech_prob": 1.4911225032498976e-11,
      "compression_ratio": 1.3123028391167193
    },
    {
      "start": 1081.332,
      "end": 1084.532,
      "text": "他不需要把决定翻译成一段语言",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09138413911224694,
      "no_speech_prob": 1.4911225032498976e-11,
      "compression_ratio": 1.3123028391167193
    },
    {
      "start": 1084.532,
      "end": 1087.532,
      "text": "再让软件把语言重新翻译成决定",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09138413911224694,
      "no_speech_prob": 1.4911225032498976e-11,
      "compression_ratio": 1.3123028391167193
    },
    {
      "start": 1087.532,
      "end": 1090.552,
      "text": "他的输出空间已经提前确定",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09138413911224694,
      "no_speech_prob": 1.4911225032498976e-11,
      "compression_ratio": 1.3123028391167193
    },
    {
      "start": 1090.552,
      "end": 1094.372,
      "text": "而且多个问题可以在同一次询问里并行计算",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09138413911224694,
      "no_speech_prob": 1.4911225032498976e-11,
      "compression_ratio": 1.3123028391167193
    },
    {
      "start": 1094.372,
      "end": 1096.872,
      "text": "TypeSafe公布的端到端延迟",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09138413911224694,
      "no_speech_prob": 1.4911225032498976e-11,
      "compression_ratio": 1.3123028391167193
    },
    {
      "start": 1096.872,
      "end": 1099.052,
      "text": "大约是70到500毫秒",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.12516392536044862,
      "no_speech_prob": 1.36730357938597e-11,
      "compression_ratio": 1.155223880597015
    },
    {
      "start": 1099.052,
      "end": 1101.952,
      "text": "输入价格是每百万Token4.2美分",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.12516392536044862,
      "no_speech_prob": 1.36730357938597e-11,
      "compression_ratio": 1.155223880597015
    },
    {
      "start": 1101.952,
      "end": 1104.032,
      "text": "输出则没有单独计费",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.12516392536044862,
      "no_speech_prob": 1.36730357938597e-11,
      "compression_ratio": 1.155223880597015
    },
    {
      "start": 1104.032,
      "end": 1106.672,
      "text": "所以从物理账上看",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.12516392536044862,
      "no_speech_prob": 1.36730357938597e-11,
      "compression_ratio": 1.155223880597015
    },
    {
      "start": 1106.672,
      "end": 1109.812,
      "text": "Jav比大元模型快很多便宜很多",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.12516392536044862,
      "no_speech_prob": 1.36730357938597e-11,
      "compression_ratio": 1.155223880597015
    },
    {
      "start": 1109.812,
      "end": 1111.372,
      "text": "是完全说得通的",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.12516392536044862,
      "no_speech_prob": 1.36730357938597e-11,
      "compression_ratio": 1.155223880597015
    },
    {
      "start": 1111.372,
      "end": 1113.952,
      "text": "但我们真正需要问的是另一件事",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.12516392536044862,
      "no_speech_prob": 1.36730357938597e-11,
      "compression_ratio": 1.155223880597015
    },
    {
      "start": 1113.952,
      "end": 1117.292,
      "text": "就是他们比较的是不是同一种工作呢",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.12516392536044862,
      "no_speech_prob": 1.36730357938597e-11,
      "compression_ratio": 1.155223880597015
    },
    {
      "start": 1117.292,
      "end": 1119.672,
      "text": "TypeSafe的Workflow Benchmark",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.12516392536044862,
      "no_speech_prob": 1.36730357938597e-11,
      "compression_ratio": 1.155223880597015
    },
    {
      "start": 1119.672,
      "end": 1122.032,
      "text": "不是让Jav和GPT Cloud",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.12516392536044862,
      "no_speech_prob": 1.36730357938597e-11,
      "compression_ratio": 1.155223880597015
    },
    {
      "start": 1122.032,
      "end": 1123.672,
      "text": "各自写一篇文章",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.12516392536044862,
      "no_speech_prob": 1.36730357938597e-11,
      "compression_ratio": 1.155223880597015
    },
    {
      "start": 1123.672,
      "end": 1126.592,
      "text": "也不是让他们解决一道开放式的数学题",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.12516392536044862,
      "no_speech_prob": 1.36730357938597e-11,
      "compression_ratio": 1.155223880597015
    },
    {
      "start": 1126.592,
      "end": 1131.372,
      "text": "而是专门选择软件自动化里面的System-1任务",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09508907160110022,
      "no_speech_prob": 1.2954306898016466e-11,
      "compression_ratio": 1.26578073089701
    },
    {
      "start": 1131.372,
      "end": 1134.472,
      "text": "对照的是同样接入这套工作流的",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09508907160110022,
      "no_speech_prob": 1.2954306898016466e-11,
      "compression_ratio": 1.26578073089701
    },
    {
      "start": 1134.472,
      "end": 1137.612,
      "text": "按要求返回结果化结果的通用LM",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09508907160110022,
      "no_speech_prob": 1.2954306898016466e-11,
      "compression_ratio": 1.26578073089701
    },
    {
      "start": 1137.612,
      "end": 1139.252,
      "text": "比如处理一张发票",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09508907160110022,
      "no_speech_prob": 1.2954306898016466e-11,
      "compression_ratio": 1.26578073089701
    },
    {
      "start": 1139.252,
      "end": 1141.832,
      "text": "系统真正需要的是这几个决定",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09508907160110022,
      "no_speech_prob": 1.2954306898016466e-11,
      "compression_ratio": 1.26578073089701
    },
    {
      "start": 1141.832,
      "end": 1146.112,
      "text": "付款、赞款、要求补文件、送人工审批",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09508907160110022,
      "no_speech_prob": 1.2954306898016466e-11,
      "compression_ratio": 1.26578073089701
    },
    {
      "start": 1146.112,
      "end": 1147.212,
      "text": "还是怀疑欺诈",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09508907160110022,
      "no_speech_prob": 1.2954306898016466e-11,
      "compression_ratio": 1.26578073089701
    },
    {
      "start": 1147.212,
      "end": 1150.672,
      "text": "又比如一个客服agent完成了一轮任务",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09508907160110022,
      "no_speech_prob": 1.2954306898016466e-11,
      "compression_ratio": 1.26578073089701
    },
    {
      "start": 1150.672,
      "end": 1155.172,
      "text": "系统要判断的是自动关闭、人工复合、优化复合",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.09508907160110022,
      "no_speech_prob": 1.2954306898016466e-11,
      "compression_ratio": 1.26578073089701
    },
    {
      "start": 1155.172,
      "end": 1157.612,
      "text": "还是直接通知值班人员",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08563758423609763,
      "no_speech_prob": 1.351566341484256e-11,
      "compression_ratio": 1.1337579617834395
    },
    {
      "start": 1157.612,
      "end": 1162.272,
      "text": "这些任务天然就非常适合被拆成很多窄判断",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08563758423609763,
      "no_speech_prob": 1.351566341484256e-11,
      "compression_ratio": 1.1337579617834395
    },
    {
      "start": 1162.272,
      "end": 1163.992,
      "text": "然后再用代码组合",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08563758423609763,
      "no_speech_prob": 1.351566341484256e-11,
      "compression_ratio": 1.1337579617834395
    },
    {
      "start": 1163.992,
      "end": 1166.312,
      "text": "而这正好是Jav的主场",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08563758423609763,
      "no_speech_prob": 1.351566341484256e-11,
      "compression_ratio": 1.1337579617834395
    },
    {
      "start": 1166.312,
      "end": 1170.132,
      "text": "TypeSave甚至在自己的发布文章里主动承认",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08563758423609763,
      "no_speech_prob": 1.351566341484256e-11,
      "compression_ratio": 1.1337579617834395
    },
    {
      "start": 1170.132,
      "end": 1172.412,
      "text": "他们展示的side-by-side demo",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08563758423609763,
      "no_speech_prob": 1.351566341484256e-11,
      "compression_ratio": 1.1337579617834395
    },
    {
      "start": 1172.412,
      "end": 1175.112,
      "text": "使用了比较短和比较密集的输入",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08563758423609763,
      "no_speech_prob": 1.351566341484256e-11,
      "compression_ratio": 1.1337579617834395
    },
    {
      "start": 1175.112,
      "end": 1177.192,
      "text": "这会让Jav看起来更有优势",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08563758423609763,
      "no_speech_prob": 1.351566341484256e-11,
      "compression_ratio": 1.1337579617834395
    },
    {
      "start": 1177.192,
      "end": 1180.592,
      "text": "193.6倍和444.6倍",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08563758423609763,
      "no_speech_prob": 1.351566341484256e-11,
      "compression_ratio": 1.1337579617834395
    },
    {
      "start": 1180.592,
      "end": 1183.172,
      "text": "来自他们设计的工作流程评估",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08563758423609763,
      "no_speech_prob": 1.351566341484256e-11,
      "compression_ratio": 1.1337579617834395
    },
    {
      "start": 1183.172,
      "end": 1186.472,
      "text": "他们认为这应该处在现实收益的高端",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07926952314076934,
      "no_speech_prob": 1.3635453009752663e-11,
      "compression_ratio": 1.294478527607362
    },
    {
      "start": 1186.472,
      "end": 1191.192,
      "text": "而且这些workflow是公司自己模型能力团队的人做的",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07926952314076934,
      "no_speech_prob": 1.3635453009752663e-11,
      "compression_ratio": 1.294478527607362
    },
    {
      "start": 1191.192,
      "end": 1193.012,
      "text": "所以他们也明确写了一句",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07926952314076934,
      "no_speech_prob": 1.3635453009752663e-11,
      "compression_ratio": 1.294478527607362
    },
    {
      "start": 1193.012,
      "end": 1194.632,
      "text": "可能存在偏见",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07926952314076934,
      "no_speech_prob": 1.3635453009752663e-11,
      "compression_ratio": 1.294478527607362
    },
    {
      "start": 1194.632,
      "end": 1196.432,
      "text": "这一点反而很重要",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07926952314076934,
      "no_speech_prob": 1.3635453009752663e-11,
      "compression_ratio": 1.294478527607362
    },
    {
      "start": 1196.432,
      "end": 1200.552,
      "text": "因为他告诉我们200倍不是一个模型智力倍率",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07926952314076934,
      "no_speech_prob": 1.3635453009752663e-11,
      "compression_ratio": 1.294478527607362
    },
    {
      "start": 1200.552,
      "end": 1201.992,
      "text": "他真正表达的是",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07926952314076934,
      "no_speech_prob": 1.3635453009752663e-11,
      "compression_ratio": 1.294478527607362
    },
    {
      "start": 1201.992,
      "end": 1204.692,
      "text": "帮任务本来就是大量窄决策",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07926952314076934,
      "no_speech_prob": 1.3635453009752663e-11,
      "compression_ratio": 1.294478527607362
    },
    {
      "start": 1204.692,
      "end": 1208.932,
      "text": "而你还让一个通用生成模型不断生成字符串的时候",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07926952314076934,
      "no_speech_prob": 1.3635453009752663e-11,
      "compression_ratio": 1.294478527607362
    },
    {
      "start": 1208.932,
      "end": 1212.592,
      "text": "你可能正在为很多根本不需要的能力付钱",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07926952314076934,
      "no_speech_prob": 1.3635453009752663e-11,
      "compression_ratio": 1.294478527607362
    },
    {
      "start": 1212.592,
      "end": 1214.952,
      "text": "TypeSafe的工作流程评估",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06355127503600301,
      "no_speech_prob": 1.1753225996613548e-11,
      "compression_ratio": 1.2236842105263157
    },
    {
      "start": 1214.952,
      "end": 1218.012,
      "text": "其实还暴露了另外一个很有意思的结果",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06355127503600301,
      "no_speech_prob": 1.1753225996613548e-11,
      "compression_ratio": 1.2236842105263157
    },
    {
      "start": 1218.012,
      "end": 1221.552,
      "text": "就是同一个模型把任务拆成workflow",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06355127503600301,
      "no_speech_prob": 1.1753225996613548e-11,
      "compression_ratio": 1.2236842105263157
    },
    {
      "start": 1221.552,
      "end": 1223.912,
      "text": "让代码负责确定性的规则",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06355127503600301,
      "no_speech_prob": 1.1753225996613548e-11,
      "compression_ratio": 1.2236842105263157
    },
    {
      "start": 1223.912,
      "end": 1227.612,
      "text": "让模型只负责那些真正需要判断的部分的话",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06355127503600301,
      "no_speech_prob": 1.1753225996613548e-11,
      "compression_ratio": 1.2236842105263157
    },
    {
      "start": 1227.612,
      "end": 1229.172,
      "text": "那么在这些测试里",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06355127503600301,
      "no_speech_prob": 1.1753225996613548e-11,
      "compression_ratio": 1.2236842105263157
    },
    {
      "start": 1229.172,
      "end": 1232.552,
      "text": "就比把整套规则塞进一个巨大pront里",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06355127503600301,
      "no_speech_prob": 1.1753225996613548e-11,
      "compression_ratio": 1.2236842105263157
    },
    {
      "start": 1232.552,
      "end": 1236.352,
      "text": "一次性解决更准确更快和更便宜",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06355127503600301,
      "no_speech_prob": 1.1753225996613548e-11,
      "compression_ratio": 1.2236842105263157
    },
    {
      "start": 1236.352,
      "end": 1239.592,
      "text": "所以这笔账不只是和模型本身有关",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.06355127503600301,
      "no_speech_prob": 1.1753225996613548e-11,
      "compression_ratio": 1.2236842105263157
    },
    {
      "start": 1239.592,
      "end": 1241.412,
      "text": "也和任务怎么拆",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08784027333639882,
      "no_speech_prob": 1.2847263183846103e-11,
      "compression_ratio": 1.2884012539184952
    },
    {
      "start": 1241.412,
      "end": 1243.312,
      "text": "软件怎么调用它有关",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08784027333639882,
      "no_speech_prob": 1.2847263183846103e-11,
      "compression_ratio": 1.2884012539184952
    },
    {
      "start": 1243.312,
      "end": 1246.492,
      "text": "那如果把这个思路放进一个完整的agent",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08784027333639882,
      "no_speech_prob": 1.2847263183846103e-11,
      "compression_ratio": 1.2884012539184952
    },
    {
      "start": 1246.492,
      "end": 1248.472,
      "text": "分工又会变成什么样呢",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08784027333639882,
      "no_speech_prob": 1.2847263183846103e-11,
      "compression_ratio": 1.2884012539184952
    },
    {
      "start": 1248.472,
      "end": 1249.712,
      "text": "讲到这里",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08784027333639882,
      "no_speech_prob": 1.2847263183846103e-11,
      "compression_ratio": 1.2884012539184952
    },
    {
      "start": 1249.712,
      "end": 1251.672,
      "text": "我觉得Jab最有意思的地方",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08784027333639882,
      "no_speech_prob": 1.2847263183846103e-11,
      "compression_ratio": 1.2884012539184952
    },
    {
      "start": 1251.672,
      "end": 1253.532,
      "text": "已经不是Jab本身了",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08784027333639882,
      "no_speech_prob": 1.2847263183846103e-11,
      "compression_ratio": 1.2884012539184952
    },
    {
      "start": 1253.532,
      "end": 1256.152,
      "text": "而是一个真正长期运行的agent",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08784027333639882,
      "no_speech_prob": 1.2847263183846103e-11,
      "compression_ratio": 1.2884012539184952
    },
    {
      "start": 1256.152,
      "end": 1259.732,
      "text": "到底应该把多少事情交给同一个大模型呢",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08784027333639882,
      "no_speech_prob": 1.2847263183846103e-11,
      "compression_ratio": 1.2884012539184952
    },
    {
      "start": 1259.732,
      "end": 1262.852,
      "text": "一个agent可能连续工作几十分钟",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08784027333639882,
      "no_speech_prob": 1.2847263183846103e-11,
      "compression_ratio": 1.2884012539184952
    },
    {
      "start": 1262.852,
      "end": 1264.112,
      "text": "甚至几个小时",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08784027333639882,
      "no_speech_prob": 1.2847263183846103e-11,
      "compression_ratio": 1.2884012539184952
    },
    {
      "start": 1264.112,
      "end": 1266.892,
      "text": "中间经历几十几百个决策点",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08784027333639882,
      "no_speech_prob": 1.2847263183846103e-11,
      "compression_ratio": 1.2884012539184952
    },
    {
      "start": 1266.892,
      "end": 1269.072,
      "text": "真正需要复杂推理的",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08784027333639882,
      "no_speech_prob": 1.2847263183846103e-11,
      "compression_ratio": 1.2884012539184952
    },
    {
      "start": 1269.072,
      "end": 1270.472,
      "text": "也许只有其中几步",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08501256647564116,
      "no_speech_prob": 1.5755062185429658e-11,
      "compression_ratio": 1.2458471760797343
    },
    {
      "start": 1270.472,
      "end": 1271.972,
      "text": "剩下大量的工作",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08501256647564116,
      "no_speech_prob": 1.5755062185429658e-11,
      "compression_ratio": 1.2458471760797343
    },
    {
      "start": 1271.972,
      "end": 1274.552,
      "text": "其实只是判断走哪条路",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08501256647564116,
      "no_speech_prob": 1.5755062185429658e-11,
      "compression_ratio": 1.2458471760797343
    },
    {
      "start": 1274.552,
      "end": 1275.612,
      "text": "调用哪个工具",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08501256647564116,
      "no_speech_prob": 1.5755062185429658e-11,
      "compression_ratio": 1.2458471760797343
    },
    {
      "start": 1275.612,
      "end": 1277.252,
      "text": "这个结果能不能接受",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08501256647564116,
      "no_speech_prob": 1.5755062185429658e-11,
      "compression_ratio": 1.2458471760797343
    },
    {
      "start": 1277.252,
      "end": 1278.952,
      "text": "需不需要再检查一次",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08501256647564116,
      "no_speech_prob": 1.5755062185429658e-11,
      "compression_ratio": 1.2458471760797343
    },
    {
      "start": 1278.952,
      "end": 1281.552,
      "text": "而Jav指向的是另一种架构",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08501256647564116,
      "no_speech_prob": 1.5755062185429658e-11,
      "compression_ratio": 1.2458471760797343
    },
    {
      "start": 1281.552,
      "end": 1283.372,
      "text": "真正开放复杂",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08501256647564116,
      "no_speech_prob": 1.5755062185429658e-11,
      "compression_ratio": 1.2458471760797343
    },
    {
      "start": 1283.372,
      "end": 1285.492,
      "text": "需要理解模糊目标",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08501256647564116,
      "no_speech_prob": 1.5755062185429658e-11,
      "compression_ratio": 1.2458471760797343
    },
    {
      "start": 1285.492,
      "end": 1288.332,
      "text": "长程规划和深度推理的问题",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08501256647564116,
      "no_speech_prob": 1.5755062185429658e-11,
      "compression_ratio": 1.2458471760797343
    },
    {
      "start": 1288.332,
      "end": 1289.972,
      "text": "仍然交给前沿模型",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08501256647564116,
      "no_speech_prob": 1.5755062185429658e-11,
      "compression_ratio": 1.2458471760797343
    },
    {
      "start": 1289.972,
      "end": 1292.172,
      "text": "如果是确定性的规则",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08501256647564116,
      "no_speech_prob": 1.5755062185429658e-11,
      "compression_ratio": 1.2458471760797343
    },
    {
      "start": 1292.172,
      "end": 1293.712,
      "text": "比如金额超过多少",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08501256647564116,
      "no_speech_prob": 1.5755062185429658e-11,
      "compression_ratio": 1.2458471760797343
    },
    {
      "start": 1293.712,
      "end": 1295.192,
      "text": "必须人工审批的",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08501256647564116,
      "no_speech_prob": 1.5755062185429658e-11,
      "compression_ratio": 1.2458471760797343
    },
    {
      "start": 1295.192,
      "end": 1296.552,
      "text": "就直接交给代码",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08501256647564116,
      "no_speech_prob": 1.5755062185429658e-11,
      "compression_ratio": 1.2458471760797343
    },
    {
      "start": 1296.552,
      "end": 1300.472,
      "text": "如果是大量高频窄而不断变化的判断",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08677304626270464,
      "no_speech_prob": 1.3209397985158855e-11,
      "compression_ratio": 1.2668810289389068
    },
    {
      "start": 1300.472,
      "end": 1303.132,
      "text": "就可以交给像Jav这种决策模型",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08677304626270464,
      "no_speech_prob": 1.3209397985158855e-11,
      "compression_ratio": 1.2668810289389068
    },
    {
      "start": 1303.132,
      "end": 1306.392,
      "text": "结果是不是满足某个明确标准",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08677304626270464,
      "no_speech_prob": 1.3209397985158855e-11,
      "compression_ratio": 1.2668810289389068
    },
    {
      "start": 1306.392,
      "end": 1308.032,
      "text": "可以再交给Verify",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08677304626270464,
      "no_speech_prob": 1.3209397985158855e-11,
      "compression_ratio": 1.2668810289389068
    },
    {
      "start": 1308.032,
      "end": 1310.992,
      "text": "需要长期保存的信息进入Memory",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08677304626270464,
      "no_speech_prob": 1.3209397985158855e-11,
      "compression_ratio": 1.2668810289389068
    },
    {
      "start": 1310.992,
      "end": 1313.632,
      "text": "需要真正改变外部世界的动作",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08677304626270464,
      "no_speech_prob": 1.3209397985158855e-11,
      "compression_ratio": 1.2668810289389068
    },
    {
      "start": 1313.632,
      "end": 1314.672,
      "text": "就交给Tools",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08677304626270464,
      "no_speech_prob": 1.3209397985158855e-11,
      "compression_ratio": 1.2668810289389068
    },
    {
      "start": 1314.672,
      "end": 1317.132,
      "text": "最后真正重要的",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08677304626270464,
      "no_speech_prob": 1.3209397985158855e-11,
      "compression_ratio": 1.2668810289389068
    },
    {
      "start": 1317.132,
      "end": 1319.092,
      "text": "可能不再是其中某一个模型",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08677304626270464,
      "no_speech_prob": 1.3209397985158855e-11,
      "compression_ratio": 1.2668810289389068
    },
    {
      "start": 1319.092,
      "end": 1320.412,
      "text": "而是外面的Harness",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08677304626270464,
      "no_speech_prob": 1.3209397985158855e-11,
      "compression_ratio": 1.2668810289389068
    },
    {
      "start": 1320.412,
      "end": 1323.992,
      "text": "因为Harness决定谁在什么时候被调用",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08677304626270464,
      "no_speech_prob": 1.3209397985158855e-11,
      "compression_ratio": 1.2668810289389068
    },
    {
      "start": 1323.992,
      "end": 1325.992,
      "text": "什么结果可以直接通过",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.08677304626270464,
      "no_speech_prob": 1.3209397985158855e-11,
      "compression_ratio": 1.2668810289389068
    },
    {
      "start": 1325.992,
      "end": 1327.792,
      "text": "什么结果必须升级",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07763803836911223,
      "no_speech_prob": 1.258090680245072e-11,
      "compression_ratio": 1.0917874396135265
    },
    {
      "start": 1327.792,
      "end": 1330.112,
      "text": "什么任务值得花几十倍的成本",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07763803836911223,
      "no_speech_prob": 1.258090680245072e-11,
      "compression_ratio": 1.0917874396135265
    },
    {
      "start": 1330.112,
      "end": 1331.792,
      "text": "去唤醒最强的模型",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07763803836911223,
      "no_speech_prob": 1.258090680245072e-11,
      "compression_ratio": 1.0917874396135265
    },
    {
      "start": 1331.792,
      "end": 1335.172,
      "text": "当软件真正需要的只是一个决定",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07763803836911223,
      "no_speech_prob": 1.258090680245072e-11,
      "compression_ratio": 1.0917874396135265
    },
    {
      "start": 1335.172,
      "end": 1338.352,
      "text": "我们为什么一定要让AI先开口说话呢",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07763803836911223,
      "no_speech_prob": 1.258090680245072e-11,
      "compression_ratio": 1.0917874396135265
    },
    {
      "start": 1338.352,
      "end": 1339.672,
      "text": "这里是All Insight",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07763803836911223,
      "no_speech_prob": 1.258090680245072e-11,
      "compression_ratio": 1.0917874396135265
    },
    {
      "start": 1339.672,
      "end": 1340.332,
      "text": "感谢收看",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07763803836911223,
      "no_speech_prob": 1.258090680245072e-11,
      "compression_ratio": 1.0917874396135265
    },
    {
      "start": 1340.332,
      "end": 1341.192,
      "text": "我们下期再见",
      "chunk": 2,
      "language": "zh",
      "avg_logprob": -0.07763803836911223,
      "no_speech_prob": 1.258090680245072e-11,
      "compression_ratio": 1.0917874396135265
    }
  ],
  "segmented_transcription": {
    "status": "SEGMENTED_TRANSCRIPT_OK",
    "created_at": "2026-09-18T09:45:06",
    "manifest": "/Users/sagawa/AI_WORK/video_notes/url/20260918_094420__比_LLM_快_200_倍_输出成本为_0_参与创造_ChatGPT_的人_把_AI_的_说话_能力删掉了/segmented_work/chunk_manifest.json",
    "audio": "/Users/sagawa/AI_WORK/video_notes/url/20260918_094420__比_LLM_快_200_倍_输出成本为_0_参与创造_ChatGPT_的人_把_AI_的_说话_能力删掉了/audio_16k_mono.wav",
    "audio_sha256": "186102fe19cd186b6b0c7b592d9a9666320fcf131992572b112a78077a5ffb8e",
    "audio_duration": 1341.486438,
    "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": 3,
    "resumed_chunks": 0,
    "merged_segment_count": 516,
    "cleaned_segment_count": 516,
    "removed_hallucination_segment_count": 0,
    "final_segment_end": 1341.192,
    "global_repeat_run": 1,
    "global_repeat_text": "9月15日迪奥哥·阿尔梅达在X上发了一条帖子。第一句话就很抓人。他说,参与创造XGP之后,我一直在问自己,为什么这些已经超过人类水平的聊天模型,没有带来AGI呢?这个问题从他嘴里说出来,确实有点特别。阿尔梅达是Intract GPT的作者之一,也参与过GPT4的RLHF和指令遵循的相关工作。",
    "tail_repeat_run": 1,
    "tail_repeat_text": "需不需要再检查一次",
    "hallucination_region_count": 0,
    "hallucination_regions": [],
    "qc_ok": true,
    "qc_errors": []
  }
}