{
  "text": "Chad GPT主创人员之一\n表示他让他感到心碎\nDiogo Ameta曾是OpenAI研究员\n参与构建了聊天机器人\n并共同发明了RLHF\n即基于人类反馈的强化学习\n这很可能是造就当前AI时代\n最关键的技术\n但他却感到失望\n用他的话来说\n我们空有惊人潜力\n却无法将其转化为实际用途\n他的观点是\n整个行业一直在针对人类语言进行优化\n模型在这方面已出色了四年之久\n但这并不能带来自动化\n毕竟计算机的语言与人类不同\n他在发布文中\n反问\n模型在对话方面已超越人类多年\n那自动化究竟在哪\n他曾以为对话模型能通向API调用\n但尽管外界喧嚣不断\n显然还缺了关键的一环\n于是两年前他离开了OpenAI\n创办了TypeSafe AI\n并出任联合创始人兼CEO\n他们融资4000万美元\n前新研发两年\n并于9月15日周二发布了Jeff\n这是他们首个系统一模型\n一类全新的前沿模型\n能直接为软件提供快速的结构化决策\n最奇特的一点是它基于Transformer架构\n但它并非大语言模型\n它不输出文本而是返回概率\n继他们所说的效准决策\nAlmeta将其称为机器原生人工智能\n因为如果人工智能要改变工作流程\n人类不能是智能的唯一受众\nTypeSafe将其称为前沿智能函数\n输入非结构化状态输出类型化的概率决策\n输出结果可以是JSON\n也可以只是一个简单的字符串\n比如我的卡被重复扣费了\n随后你可以使用名为Choice Score的原语来提问\n以及Nil他们会返回类型化的选择评分\n或是带有概率的不耳直答案全部并行评估\n以工单分类为例\n假设你需要决定有哪个部门\n处理客户咨询\n你会得到账务0.08\n技术0.85\n销售0.07\n智信度0.82\n这对人类而言毫无意义\n但对代码来说却非常完美\n因为可以直接据此进行逻辑分支\n这就是名字的由来\n类型安全\n是指程序捕获异常数据的方式\n即防止程序遇到预期之外的情况\n比如试图用整数\n去除以字符串而使用Jeff\n你可以预先定义输出格式\n它绝不会输出结构以外的内容\n省去了解析与验证的麻烦\n而大语言模型输出的只是一个字符串\n可能是对话回复\n可能是代码\n可能是幻觉\n也可能是拒绝回答\n又或是带有风险的结构化数据\n稍不留神就会失控\n你当然能用代码封装大语言模型\n但这正是Jeff的优势所在\nTypeSafe声称它不会产生幻觉\n在处理系统一任务时\n其表现足以媲美大语言模型\n且速度快了两个数量级\n效率也更高在其底层架构下\n采用了一种全新的并行采样器架构\n以及一种名为强化学习的训练方法\n即基于效准决策\n简称RLCD\nThe Register将RLCD归类为一种架构\n但它本质上是一种训练方法\n也是他们对RLHF的回应\n像GPD 5.6 Terra这类模型\n每次只能生成一个Token\n每个Token都依赖前一个Jeff就能一次性生成所有结果\n通过单次硬件感知查询实现\n它基于结构化程序状态而非序列消息构建\n他们认为放弃字符串赋予了模型超能力\n而且速度快得惊人这就是TypeSafe\n前沿大模型处理耗时\n3到329秒端到端\n人类用者还行\n但在代码自动化场景下就是瓶颈Jeff只需\n70到500毫秒\n速度快了40到200倍\n在TechCrunch报道的对比演示中\n截图显示Jeff响应仅需0.114秒\n而Tara则用了8.566秒\n更惊人的是价格\n大模型每百万输入\nToken收费20美分到10美元\n输出费用是输入的5倍\n而Jeff每百万Token仅需0.042美元\n也就是每10亿Token仅需42美元\n没错他们按10亿单位计费\n输出成本极低\n几乎可以忽略不计\nTara模型输入为每百万两美元\n输出则为每百万十二美元\n数据显示Jeff模型\n比Fibo 5.1便宜了238倍\n再来说说智信度\n这恐怕是最被低估的一点大模型\n往往过度自信\n且对自身确信度的评估前后不一\n如果一个模型能在95%的情况下完成任务\n但却无法告诉你\n它何时会出错\n你就无法实现自动化\nJeff的智信度经过了较准\n智信度越高准确率也就越高\n相似的输入会得到相似的答案\nIron DL的首席技术官\nArmond Risser表示\n这让开源模型得以发挥其效用\n他告诉TechCrunch\n这实际上是将\n幻觉问题交由用户自行把控\n知性度5成10如抛硬币直接忽略\n9成50则果断执行\n面对高调的声称Type Safe\n心知肚明\n正如文中那句\n非凡的结论需要非凡的证据\n下方还附上了有图有真相的证据配文\n还带了个图指加油的表情\n他们自称欢迎质疑\n因为他们自己就是怀疑论者\n值得称赞的是\n他们在速度优化上确实下足了功夫\n你可以亲自验证\n这些模型多运行在西海岸的笔记本上\n也就是该服务的所在地定价方面很透明\n但他们坦言无法证明其未获补贴\n并预计价格未来会进一步下降\n且不存在类型错误\n虽然一个反例就能将其推翻\n但这在数学上是不可能的\n他们说在Terra演示中\n他们坦言查询过程经过了高度简化\n使用了可读的问题件外加一段\n有利于Jeff的密集状态描述\n从而美化了其表现\n他们选用了默认推理的Terra\n认为其智力水平最接近Jeff\n唯一的意见分歧在于流失可能性的判断\n他们称之为真正的模棱两可有趣的是\n正是类似的演示说服了他们全力投入\n更有利的证据是他们的工作流评估\n旨在衡量AI在代码环境下的表现\n没有标准答案标签\n也不对评估框架进行微调\n迎面导致过拟和每个模型都运行相同的工作流\n并根据GT6 Extra的平均表现评分\n以及Fable 5.1他们说Jeff稳居帕雷托前沿\n在近两个数量级的范围内\n且由大语言模型处理所有逻辑\n比起遵循工作流思维链的表现明显差得多\n这些调用远比演示案例复杂\n更接近实际生产环境\n他们将结果发布在评估网站上\n他们认为可靠的工作流是有许多小型的独立\n由概率驱动的问题组成最终导向一个分支\n这就是\n他们主页上那些数据的来源也就是100\n93.6倍的提速\n以及444.6倍的成本缩减\nVirtue将其四舍侮入为100\n94和400\n即45倍\n这大概是实际应用中的上限\n他们声称这些工作流并非刻意挑选\n且不在训练数据分布之内\n但毕竟是他们自家团队构建的\n所以难免存在偏差Astra\n和Fable的参考案例\n可能更偏向OpenAI和Anthlopic\n所以他们认为自己可能还低估了性能\n以及DeepSeq这些大模型\n都是通过TypeSafe的开源系统运行的\n即针对Python的大模型封装库\n他们称这是获取大模型决策最准确的方式\n但处理幻觉时速度较慢成本也更高\n他们视类型安全为基本门槛\n对智能体来说工具调用产生幻觉是不可接受的\n但对于有延迟要求的系统\n这是致命的缺陷\n或是深埋在依赖链的底层\n即便是最强的模型也难免出错\n它们的大模型指标来自OpenRouter\n复杂查询会被自动路由到更强的模型\nJeff的零错误率不是测出来的\n而是由模式匹配保证的\nThe Register的反驳很重肯\nJeff并不生成自然语言\n所以这种比较并不公平\n概率模型仍可能出错\n它虽不会捏造虚假的法律引文\n但它也不会被那样使用\n现在来看点有趣的\n他们做了一个玩毁灭战士的Jeff AI\n以此掩饰其具备的实时智能\n据他们说\n工程师曾担心每秒十次查询\n以及每小时七美元的成本\n团队认为这比预期更低\n他读取的是文本\n而非图像格式的游戏状态\n暂时还不是\n他们也承认非AI机器人玩起来会更强\n但他们想要的是能适应多种游戏状态表达的AI\n且能听从指令用他们的话说着酷毙了\n演示和黑客松活动即将开启还有危机竞速\n仅靠点击链接在成百上千个页面间穿梭\n每一步都有上千种选择\n充分展现了其极高的处理效率\n面对海量选项不产生幻觉的优势会不断叠加\nJeff支持的基数上限为200\n一旦超过这个数值\n他会独立评估各个链接再进行选择\n因此偶尔会出现减速\n毕竟大模型没有推理过程整体提速空间有限\n但Astra设为最低档以保证流畅度的情况除外\n而且Jeff往往能以更少的步骤完成任务\n这是我最喜欢的一点\n第二和第三项挑战的起始页面都随机抽到了小黄鸭\n直到团队指出来作者才发现这一点\n开发者们蜂拥而至导致需求激增\n以至于TypeSafe的API一度无法为用户提供服务\n据TechCrunch报道\n开发者将其视为一种更经济\n更稳健的智能集成方案\nAdversal公司的工程师Brenet Sharma表示\n他们此前一直在使用指令安全分类器\n基于OpenAI的GPD-CO mini模型\n但换成Jeff后\n处理速度提升了5到18倍\n且准确度更高\nBrio AI的CTO Nikhil Madhoka\n在处理商务邮件时对比了Gemini\nGemini的准确率略胜一筹\n但成本却高出10到20倍\n让他感到震惊的是\n况且Jeff是唯一能提供真实概率反馈的模型\n他认为这使其成为自动化工作流的理想之选\nAI Gateway的数据表现简直惊人\n发布24小时内\nJeff吸引的付费团队数量是此前任何模型的两倍多\n这是Gateway历史上普及最快的模型\n12小时内就超越了所有竞品\n第18小时就覆盖了1%的付费团队\n第24小时是接近13%\n是GPT 5.6系列的两倍\n更是Fable 5.1的六倍多\n而其他近期发布的模型均未超过7%\n首日数据就足以证明\n专用模型投入生产环境的速度有多快\n但真正的考验在于持久性谈到具体用途\nPerser提到智能体可以自主选择下一个工具\n或在子智能体的工作流中决定是继续重视还是询问用户\n亦或在行动前先评估风险并决定是否终止\n并将无法确定的情况转交人工处理\nTypeSafe将其称为智能异服语句或模糊规则\n因为手写逻辑往往过于脆弱\n他们还建议使用Map\n将PB级数据精炼为特征以支持实时应用\n毕竟100毫秒的延迟完全在用户体验的可控范围内\n从推理过程到越狱尝试全程进行验证\nAmita希望人们能用它来监控智能体\n由于智能体互间成本高昂\nRuniker建议采用实时模型路由\nThe Register补充到\n聊天机器人人由LLM架构主导\n编程助手代码智能体\n以及数学证明或演示等可验证任务\n而这正如TypeSafe戏称的那样\n效果时好时坏\n至于内部究竟如何\nAmita手口如瓶\n外界普遍怀疑其底层使用的是\n开源权重的大模型\n尽管Jeff撑起FAQ文档内容并不短\nJeff模型完全采用合成数据进行训练\nOmeta告诉TechCrunch\n资产数据是他人生中最成功的豪赌之一\nif\n比发布产品更成功\n也优于RLHF\nTypeSafe自诩为一家核心的数据研究实验室\n公司半数员工几乎主导了\n统计学可控合成数据这一细分领域\n这已成他毕生乐趣\n即便你请求他们也不会用你的数据训练请见谅\n想了解细节\n得先被他们录用\n这就是他们所奉行的苦涩教训\n优化任务远比堆砌数据和算力重要\n或是RLHF这类针对文本优化的算法\n人类评估员偏好虽然适合聊天\n却不适用于自动化任务\n且RLHF仅适用于程序可教验的任务\n导致其现实判断波动极大\n智能表现不够稳健\n系统1这一名称源自卡尼曼的思考快与曼\n尽管在传统认知中系统1往往意味着一出错\n但他们认为这些模型其实比其他方案更可靠\n他们还完全放弃了公开基准测试\n仅提供一次性评估\n并推动用户自行构建评估体系\n因为系统一任务更易于评估\n甚至还有篇题为反基准优化的文章\nJeff这个名字源自威廉·Tauphel\n这位19世纪的经济学家提出了一个悖论\n当蒸汽机效率提升\n煤炭需求反而不降反升\n他们压住的是\n智能成本每降低一个数量级\n就会解锁多出数个数量级的应用场景\n阿尔梅达预见智能软件将无处不再呈现出涌现性\n且更像早期互联网那样分布化\n而非如今的超级应用\n但The Register指出\n这建立在一个假设之上\n其Token的需求会像能源一样广泛\n这尚未有定论PO研究中心\n发现许多能用上AI的人\n根本找不到用处\nCNBC调查显示\n另一些人则出于道德顾虑排斥AI\n目前Jeff依然是特例\n但Reneker预计竞争终会出现\n他感叹我们本应早点看透这一点\n毕竟大语言模型太便宜了\n再加上补贴人们往往懒得去创新\n但Type Safe正在开发更多新模态版本\n并开始从候补名单中邀请开发者\n询问他们想实现什么样的科幻构想\n当Type Crunch问起Type Safe是否属于前沿实验室时\nAlmeta直言前沿实验室的主打产品\n无非是恐惧和炒作\n它追求的是真正智能而非常无限财富的豪赌\n也不是某种宗教\n更不是在数据中心里创造神\n总之这是Jeff\n欢迎在评论区留言分享\n你会如何使用它\n感谢观看\n我们下期见\nThank you.",
  "language": "multilingual",
  "duration": 880.570375,
  "segments": [
    {
      "start": 0.0,
      "end": 2.0,
      "text": "Chad GPT主创人员之一",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 2.0,
      "end": 5.0,
      "text": "表示他让他感到心碎",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 5.0,
      "end": 8.0,
      "text": "Diogo Ameta曾是OpenAI研究员",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 9.495,
      "end": 11.005,
      "text": "参与构建了聊天机器人",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 11.005,
      "end": 13.125,
      "text": "并共同发明了RLHF",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 13.125,
      "end": 15.405,
      "text": "即基于人类反馈的强化学习",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 15.405,
      "end": 18.145,
      "text": "这很可能是造就当前AI时代",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 18.145,
      "end": 19.485,
      "text": "最关键的技术",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 19.485,
      "end": 20.865,
      "text": "但他却感到失望",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 20.865,
      "end": 21.705,
      "text": "用他的话来说",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 21.705,
      "end": 23.345,
      "text": "我们空有惊人潜力",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 23.345,
      "end": 25.725,
      "text": "却无法将其转化为实际用途",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 25.725,
      "end": 26.745,
      "text": "他的观点是",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 26.745,
      "end": 29.725,
      "text": "整个行业一直在针对人类语言进行优化",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 29.725,
      "end": 32.405,
      "text": "模型在这方面已出色了四年之久",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 32.405,
      "end": 34.425,
      "text": "但这并不能带来自动化",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 34.425,
      "end": 36.625,
      "text": "毕竟计算机的语言与人类不同",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 36.625,
      "end": 37.665,
      "text": "他在发布文中",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 37.665,
      "end": 38.465,
      "text": "反问",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 38.465,
      "end": 41.825,
      "text": "模型在对话方面已超越人类多年",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 41.825,
      "end": 43.905,
      "text": "那自动化究竟在哪",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 43.905,
      "end": 46.885,
      "text": "他曾以为对话模型能通向API调用",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 46.885,
      "end": 48.585,
      "text": "但尽管外界喧嚣不断",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 48.585,
      "end": 50.045,
      "text": "显然还缺了关键的一环",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 50.045,
      "end": 52.505,
      "text": "于是两年前他离开了OpenAI",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 52.505,
      "end": 54.285,
      "text": "创办了TypeSafe AI",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 54.285,
      "end": 57.725,
      "text": "并出任联合创始人兼CEO",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 57.725,
      "end": 59.505,
      "text": "他们融资4000万美元",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 59.505,
      "end": 60.925,
      "text": "前新研发两年",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 60.925,
      "end": 64.005,
      "text": "并于9月15日周二发布了Jeff",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 64.005,
      "end": 66.125,
      "text": "这是他们首个系统一模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 66.125,
      "end": 67.785,
      "text": "一类全新的前沿模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 67.785,
      "end": 70.765,
      "text": "能直接为软件提供快速的结构化决策",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 1,
      "language_run": 0
    },
    {
      "start": 72.647,
      "end": 74.977,
      "text": "最奇特的一点是它基于Transformer架构",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 2,
      "language_run": 0
    },
    {
      "start": 74.977,
      "end": 76.837,
      "text": "但它并非大语言模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 2,
      "language_run": 0
    },
    {
      "start": 76.837,
      "end": 78.997,
      "text": "它不输出文本而是返回概率",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 2,
      "language_run": 0
    },
    {
      "start": 78.997,
      "end": 81.197,
      "text": "继他们所说的效准决策",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 2,
      "language_run": 0
    },
    {
      "start": 81.197,
      "end": 84.457,
      "text": "Almeta将其称为机器原生人工智能",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 2,
      "language_run": 0
    },
    {
      "start": 86.795,
      "end": 89.485,
      "text": "因为如果人工智能要改变工作流程",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 89.485,
      "end": 91.885,
      "text": "人类不能是智能的唯一受众",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 91.885,
      "end": 95.265,
      "text": "TypeSafe将其称为前沿智能函数",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 95.265,
      "end": 98.445,
      "text": "输入非结构化状态输出类型化的概率决策",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 98.445,
      "end": 101.405,
      "text": "输出结果可以是JSON",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 101.405,
      "end": 103.405,
      "text": "也可以只是一个简单的字符串",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 103.405,
      "end": 105.545,
      "text": "比如我的卡被重复扣费了",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 105.545,
      "end": 108.805,
      "text": "随后你可以使用名为Choice Score的原语来提问",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 108.805,
      "end": 111.985,
      "text": "以及Nil他们会返回类型化的选择评分",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 111.985,
      "end": 115.225,
      "text": "或是带有概率的不耳直答案全部并行评估",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 115.225,
      "end": 116.785,
      "text": "以工单分类为例",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 116.785,
      "end": 119.025,
      "text": "假设你需要决定有哪个部门",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 119.025,
      "end": 120.505,
      "text": "处理客户咨询",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 120.505,
      "end": 123.065,
      "text": "你会得到账务0.08",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 123.065,
      "end": 124.365,
      "text": "技术0.85",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 124.365,
      "end": 127.705,
      "text": "销售0.07",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 127.705,
      "end": 128.825,
      "text": "智信度0.82",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 128.825,
      "end": 130.785,
      "text": "这对人类而言毫无意义",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 130.785,
      "end": 132.885,
      "text": "但对代码来说却非常完美",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 132.885,
      "end": 135.185,
      "text": "因为可以直接据此进行逻辑分支",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 135.185,
      "end": 136.565,
      "text": "这就是名字的由来",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 136.565,
      "end": 137.185,
      "text": "类型安全",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 137.185,
      "end": 139.405,
      "text": "是指程序捕获异常数据的方式",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 139.405,
      "end": 142.025,
      "text": "即防止程序遇到预期之外的情况",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 142.025,
      "end": 143.325,
      "text": "比如试图用整数",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 143.325,
      "end": 145.485,
      "text": "去除以字符串而使用Jeff",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 145.485,
      "end": 147.525,
      "text": "你可以预先定义输出格式",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 147.525,
      "end": 149.885,
      "text": "它绝不会输出结构以外的内容",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 149.885,
      "end": 151.905,
      "text": "省去了解析与验证的麻烦",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 151.905,
      "end": 154.305,
      "text": "而大语言模型输出的只是一个字符串",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 154.305,
      "end": 155.485,
      "text": "可能是对话回复",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 155.485,
      "end": 156.165,
      "text": "可能是代码",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 156.165,
      "end": 157.225,
      "text": "可能是幻觉",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 157.225,
      "end": 158.305,
      "text": "也可能是拒绝回答",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 158.305,
      "end": 160.745,
      "text": "又或是带有风险的结构化数据",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 160.745,
      "end": 162.085,
      "text": "稍不留神就会失控",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 162.085,
      "end": 163.985,
      "text": "你当然能用代码封装大语言模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 163.985,
      "end": 166.225,
      "text": "但这正是Jeff的优势所在",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 166.225,
      "end": 168.645,
      "text": "TypeSafe声称它不会产生幻觉",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 168.645,
      "end": 170.885,
      "text": "在处理系统一任务时",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 170.885,
      "end": 173.065,
      "text": "其表现足以媲美大语言模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 173.065,
      "end": 174.885,
      "text": "且速度快了两个数量级",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 174.885,
      "end": 177.245,
      "text": "效率也更高在其底层架构下",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 177.245,
      "end": 180.145,
      "text": "采用了一种全新的并行采样器架构",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 180.145,
      "end": 183.145,
      "text": "以及一种名为强化学习的训练方法",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 183.145,
      "end": 184.925,
      "text": "即基于效准决策",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 184.925,
      "end": 186.265,
      "text": "简称RLCD",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 186.265,
      "end": 190.065,
      "text": "The Register将RLCD归类为一种架构",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 190.065,
      "end": 192.185,
      "text": "但它本质上是一种训练方法",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 192.185,
      "end": 194.225,
      "text": "也是他们对RLHF的回应",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 194.225,
      "end": 196.605,
      "text": "像GPD 5.6 Terra这类模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 196.605,
      "end": 198.185,
      "text": "每次只能生成一个Token",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 198.185,
      "end": 203.025,
      "text": "每个Token都依赖前一个Jeff就能一次性生成所有结果",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 203.025,
      "end": 205.405,
      "text": "通过单次硬件感知查询实现",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 205.405,
      "end": 208.665,
      "text": "它基于结构化程序状态而非序列消息构建",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 208.665,
      "end": 212.185,
      "text": "他们认为放弃字符串赋予了模型超能力",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 212.185,
      "end": 214.545,
      "text": "而且速度快得惊人这就是TypeSafe",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 214.545,
      "end": 216.425,
      "text": "前沿大模型处理耗时",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 216.425,
      "end": 218.665,
      "text": "3到329秒端到端",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 218.665,
      "end": 220.345,
      "text": "人类用者还行",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 220.345,
      "end": 223.585,
      "text": "但在代码自动化场景下就是瓶颈Jeff只需",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 223.585,
      "end": 225.105,
      "text": "70到500毫秒",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 225.105,
      "end": 227.185,
      "text": "速度快了40到200倍",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 227.185,
      "end": 230.405,
      "text": "在TechCrunch报道的对比演示中",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 230.405,
      "end": 235.325,
      "text": "截图显示Jeff响应仅需0.114秒",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 235.325,
      "end": 238.545,
      "text": "而Tara则用了8.566秒",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 238.545,
      "end": 239.865,
      "text": "更惊人的是价格",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 239.865,
      "end": 241.425,
      "text": "大模型每百万输入",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 241.425,
      "end": 243.685,
      "text": "Token收费20美分到10美元",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 243.685,
      "end": 245.785,
      "text": "输出费用是输入的5倍",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 245.785,
      "end": 250.605,
      "text": "而Jeff每百万Token仅需0.042美元",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 250.605,
      "end": 253.305,
      "text": "也就是每10亿Token仅需42美元",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 253.305,
      "end": 254.745,
      "text": "没错他们按10亿单位计费",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 254.745,
      "end": 256.005,
      "text": "输出成本极低",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 256.005,
      "end": 256.965,
      "text": "几乎可以忽略不计",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 256.965,
      "end": 259.085,
      "text": "Tara模型输入为每百万两美元",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 259.085,
      "end": 261.665,
      "text": "输出则为每百万十二美元",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 261.665,
      "end": 263.325,
      "text": "数据显示Jeff模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 263.325,
      "end": 265.865,
      "text": "比Fibo 5.1便宜了238倍",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 265.865,
      "end": 267.425,
      "text": "再来说说智信度",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 267.425,
      "end": 270.905,
      "text": "这恐怕是最被低估的一点大模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 270.905,
      "end": 271.925,
      "text": "往往过度自信",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 271.925,
      "end": 274.445,
      "text": "且对自身确信度的评估前后不一",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 274.445,
      "end": 277.545,
      "text": "如果一个模型能在95%的情况下完成任务",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 277.545,
      "end": 278.905,
      "text": "但却无法告诉你",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 278.905,
      "end": 279.985,
      "text": "它何时会出错",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 279.985,
      "end": 281.225,
      "text": "你就无法实现自动化",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 281.225,
      "end": 283.405,
      "text": "Jeff的智信度经过了较准",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 283.405,
      "end": 286.545,
      "text": "智信度越高准确率也就越高",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 286.545,
      "end": 288.785,
      "text": "相似的输入会得到相似的答案",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 288.785,
      "end": 290.885,
      "text": "Iron DL的首席技术官",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 290.885,
      "end": 292.165,
      "text": "Armond Risser表示",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 292.165,
      "end": 294.305,
      "text": "这让开源模型得以发挥其效用",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 294.305,
      "end": 296.725,
      "text": "他告诉TechCrunch",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 296.725,
      "end": 298.065,
      "text": "这实际上是将",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 298.065,
      "end": 300.605,
      "text": "幻觉问题交由用户自行把控",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 300.605,
      "end": 303.125,
      "text": "知性度5成10如抛硬币直接忽略",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 303.125,
      "end": 305.025,
      "text": "9成50则果断执行",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 305.025,
      "end": 307.005,
      "text": "面对高调的声称Type Safe",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 307.005,
      "end": 307.805,
      "text": "心知肚明",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 307.805,
      "end": 309.505,
      "text": "正如文中那句",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 309.505,
      "end": 311.525,
      "text": "非凡的结论需要非凡的证据",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 311.525,
      "end": 314.525,
      "text": "下方还附上了有图有真相的证据配文",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 314.525,
      "end": 316.025,
      "text": "还带了个图指加油的表情",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 316.025,
      "end": 317.685,
      "text": "他们自称欢迎质疑",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 317.685,
      "end": 319.205,
      "text": "因为他们自己就是怀疑论者",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 319.205,
      "end": 320.465,
      "text": "值得称赞的是",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 320.465,
      "end": 322.825,
      "text": "他们在速度优化上确实下足了功夫",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 322.825,
      "end": 324.205,
      "text": "你可以亲自验证",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 324.205,
      "end": 327.205,
      "text": "这些模型多运行在西海岸的笔记本上",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 327.205,
      "end": 330.025,
      "text": "也就是该服务的所在地定价方面很透明",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 330.025,
      "end": 332.905,
      "text": "但他们坦言无法证明其未获补贴",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 332.905,
      "end": 335.445,
      "text": "并预计价格未来会进一步下降",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 335.445,
      "end": 337.205,
      "text": "且不存在类型错误",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 337.205,
      "end": 339.065,
      "text": "虽然一个反例就能将其推翻",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 339.065,
      "end": 341.105,
      "text": "但这在数学上是不可能的",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 341.105,
      "end": 342.945,
      "text": "他们说在Terra演示中",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 342.945,
      "end": 345.425,
      "text": "他们坦言查询过程经过了高度简化",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 345.425,
      "end": 347.585,
      "text": "使用了可读的问题件外加一段",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 347.585,
      "end": 349.585,
      "text": "有利于Jeff的密集状态描述",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 349.585,
      "end": 350.905,
      "text": "从而美化了其表现",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 350.905,
      "end": 352.905,
      "text": "他们选用了默认推理的Terra",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 352.905,
      "end": 355.265,
      "text": "认为其智力水平最接近Jeff",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 355.265,
      "end": 358.345,
      "text": "唯一的意见分歧在于流失可能性的判断",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 358.345,
      "end": 361.285,
      "text": "他们称之为真正的模棱两可有趣的是",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 361.285,
      "end": 364.465,
      "text": "正是类似的演示说服了他们全力投入",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 364.465,
      "end": 368.065,
      "text": "更有利的证据是他们的工作流评估",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 368.065,
      "end": 370.425,
      "text": "旨在衡量AI在代码环境下的表现",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 370.425,
      "end": 371.805,
      "text": "没有标准答案标签",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 371.805,
      "end": 373.465,
      "text": "也不对评估框架进行微调",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 373.465,
      "end": 377.265,
      "text": "迎面导致过拟和每个模型都运行相同的工作流",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 377.265,
      "end": 380.725,
      "text": "并根据GT6 Extra的平均表现评分",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 380.725,
      "end": 384.385,
      "text": "以及Fable 5.1他们说Jeff稳居帕雷托前沿",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 384.385,
      "end": 387.545,
      "text": "在近两个数量级的范围内",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 387.545,
      "end": 390.105,
      "text": "且由大语言模型处理所有逻辑",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 390.105,
      "end": 393.865,
      "text": "比起遵循工作流思维链的表现明显差得多",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 393.865,
      "end": 396.325,
      "text": "这些调用远比演示案例复杂",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 396.325,
      "end": 397.845,
      "text": "更接近实际生产环境",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 397.845,
      "end": 400.865,
      "text": "他们将结果发布在评估网站上",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 400.865,
      "end": 404.825,
      "text": "他们认为可靠的工作流是有许多小型的独立",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 404.825,
      "end": 408.245,
      "text": "由概率驱动的问题组成最终导向一个分支",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 408.245,
      "end": 408.705,
      "text": "这就是",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 408.705,
      "end": 411.345,
      "text": "他们主页上那些数据的来源也就是100",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 411.345,
      "end": 413.285,
      "text": "93.6倍的提速",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 413.285,
      "end": 416.245,
      "text": "以及444.6倍的成本缩减",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 416.245,
      "end": 418.785,
      "text": "Virtue将其四舍侮入为100",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 418.785,
      "end": 420.265,
      "text": "94和400",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 420.265,
      "end": 421.465,
      "text": "即45倍",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 421.465,
      "end": 423.265,
      "text": "这大概是实际应用中的上限",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 423.265,
      "end": 426.025,
      "text": "他们声称这些工作流并非刻意挑选",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 426.025,
      "end": 428.205,
      "text": "且不在训练数据分布之内",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 428.205,
      "end": 430.605,
      "text": "但毕竟是他们自家团队构建的",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 430.605,
      "end": 432.765,
      "text": "所以难免存在偏差Astra",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 432.765,
      "end": 434.285,
      "text": "和Fable的参考案例",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 434.285,
      "end": 436.745,
      "text": "可能更偏向OpenAI和Anthlopic",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 436.745,
      "end": 439.405,
      "text": "所以他们认为自己可能还低估了性能",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 439.405,
      "end": 441.225,
      "text": "以及DeepSeq这些大模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 441.225,
      "end": 443.665,
      "text": "都是通过TypeSafe的开源系统运行的",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 443.665,
      "end": 446.285,
      "text": "即针对Python的大模型封装库",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 446.285,
      "end": 449.365,
      "text": "他们称这是获取大模型决策最准确的方式",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 449.365,
      "end": 452.065,
      "text": "但处理幻觉时速度较慢成本也更高",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 452.065,
      "end": 454.385,
      "text": "他们视类型安全为基本门槛",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 454.385,
      "end": 457.825,
      "text": "对智能体来说工具调用产生幻觉是不可接受的",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 457.825,
      "end": 459.785,
      "text": "但对于有延迟要求的系统",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 459.785,
      "end": 460.825,
      "text": "这是致命的缺陷",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 460.825,
      "end": 462.905,
      "text": "或是深埋在依赖链的底层",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 462.905,
      "end": 465.385,
      "text": "即便是最强的模型也难免出错",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 465.385,
      "end": 468.825,
      "text": "它们的大模型指标来自OpenRouter",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 468.825,
      "end": 471.785,
      "text": "复杂查询会被自动路由到更强的模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 471.785,
      "end": 473.825,
      "text": "Jeff的零错误率不是测出来的",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 473.825,
      "end": 475.485,
      "text": "而是由模式匹配保证的",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 475.485,
      "end": 477.185,
      "text": "The Register的反驳很重肯",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 477.185,
      "end": 478.885,
      "text": "Jeff并不生成自然语言",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 3,
      "language_run": 1
    },
    {
      "start": 481.741,
      "end": 483.891,
      "text": "所以这种比较并不公平",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 483.891,
      "end": 485.931,
      "text": "概率模型仍可能出错",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 485.931,
      "end": 488.271,
      "text": "它虽不会捏造虚假的法律引文",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 488.271,
      "end": 490.211,
      "text": "但它也不会被那样使用",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 490.211,
      "end": 491.511,
      "text": "现在来看点有趣的",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 491.511,
      "end": 494.171,
      "text": "他们做了一个玩毁灭战士的Jeff AI",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 494.171,
      "end": 496.511,
      "text": "以此掩饰其具备的实时智能",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 496.511,
      "end": 498.011,
      "text": "据他们说",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 498.011,
      "end": 500.391,
      "text": "工程师曾担心每秒十次查询",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 500.391,
      "end": 502.571,
      "text": "以及每小时七美元的成本",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 502.571,
      "end": 504.511,
      "text": "团队认为这比预期更低",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 504.511,
      "end": 506.511,
      "text": "他读取的是文本",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 506.511,
      "end": 508.211,
      "text": "而非图像格式的游戏状态",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 508.211,
      "end": 509.171,
      "text": "暂时还不是",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 509.171,
      "end": 512.011,
      "text": "他们也承认非AI机器人玩起来会更强",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 512.011,
      "end": 515.511,
      "text": "但他们想要的是能适应多种游戏状态表达的AI",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 515.511,
      "end": 519.071,
      "text": "且能听从指令用他们的话说着酷毙了",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 519.071,
      "end": 522.731,
      "text": "演示和黑客松活动即将开启还有危机竞速",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 1,
      "language_run": 2
    },
    {
      "start": 524.314,
      "end": 527.564,
      "text": "仅靠点击链接在成百上千个页面间穿梭",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 2,
      "language_run": 2
    },
    {
      "start": 527.564,
      "end": 529.604,
      "text": "每一步都有上千种选择",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 2,
      "language_run": 2
    },
    {
      "start": 529.604,
      "end": 531.864,
      "text": "充分展现了其极高的处理效率",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 2,
      "language_run": 2
    },
    {
      "start": 533.428,
      "end": 536.678,
      "text": "面对海量选项不产生幻觉的优势会不断叠加",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 3,
      "language_run": 2
    },
    {
      "start": 536.678,
      "end": 539.138,
      "text": "Jeff支持的基数上限为200",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 3,
      "language_run": 2
    },
    {
      "start": 539.138,
      "end": 540.598,
      "text": "一旦超过这个数值",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 3,
      "language_run": 2
    },
    {
      "start": 540.598,
      "end": 543.258,
      "text": "他会独立评估各个链接再进行选择",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 3,
      "language_run": 2
    },
    {
      "start": 543.258,
      "end": 545.198,
      "text": "因此偶尔会出现减速",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 3,
      "language_run": 2
    },
    {
      "start": 545.198,
      "end": 548.898,
      "text": "毕竟大模型没有推理过程整体提速空间有限",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 3,
      "language_run": 2
    },
    {
      "start": 548.898,
      "end": 552.658,
      "text": "但Astra设为最低档以保证流畅度的情况除外",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 3,
      "language_run": 2
    },
    {
      "start": 552.658,
      "end": 555.778,
      "text": "而且Jeff往往能以更少的步骤完成任务",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 3,
      "language_run": 2
    },
    {
      "start": 555.778,
      "end": 556.958,
      "text": "这是我最喜欢的一点",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 3,
      "language_run": 2
    },
    {
      "start": 556.958,
      "end": 560.558,
      "text": "第二和第三项挑战的起始页面都随机抽到了小黄鸭",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 3,
      "language_run": 2
    },
    {
      "start": 560.558,
      "end": 563.318,
      "text": "直到团队指出来作者才发现这一点",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 3,
      "language_run": 2
    },
    {
      "start": 564.711,
      "end": 566.941,
      "text": "开发者们蜂拥而至导致需求激增",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 566.941,
      "end": 570.501,
      "text": "以至于TypeSafe的API一度无法为用户提供服务",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 570.501,
      "end": 571.541,
      "text": "据TechCrunch报道",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 571.541,
      "end": 573.541,
      "text": "开发者将其视为一种更经济",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 573.541,
      "end": 575.921,
      "text": "更稳健的智能集成方案",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 575.921,
      "end": 579.321,
      "text": "Adversal公司的工程师Brenet Sharma表示",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 579.321,
      "end": 582.481,
      "text": "他们此前一直在使用指令安全分类器",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 582.481,
      "end": 585.021,
      "text": "基于OpenAI的GPD-CO mini模型",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 585.021,
      "end": 586.201,
      "text": "但换成Jeff后",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 586.201,
      "end": 588.561,
      "text": "处理速度提升了5到18倍",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 588.561,
      "end": 590.221,
      "text": "且准确度更高",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 590.221,
      "end": 593.001,
      "text": "Brio AI的CTO Nikhil Madhoka",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 593.001,
      "end": 595.621,
      "text": "在处理商务邮件时对比了Gemini",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 595.621,
      "end": 597.721,
      "text": "Gemini的准确率略胜一筹",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 597.721,
      "end": 600.521,
      "text": "但成本却高出10到20倍",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 600.521,
      "end": 601.941,
      "text": "让他感到震惊的是",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 601.941,
      "end": 606.041,
      "text": "况且Jeff是唯一能提供真实概率反馈的模型",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 606.041,
      "end": 609.721,
      "text": "他认为这使其成为自动化工作流的理想之选",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 609.721,
      "end": 612.201,
      "text": "AI Gateway的数据表现简直惊人",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 612.201,
      "end": 613.661,
      "text": "发布24小时内",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 613.661,
      "end": 617.741,
      "text": "Jeff吸引的付费团队数量是此前任何模型的两倍多",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 617.741,
      "end": 620.581,
      "text": "这是Gateway历史上普及最快的模型",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 620.581,
      "end": 622.801,
      "text": "12小时内就超越了所有竞品",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 622.801,
      "end": 625.841,
      "text": "第18小时就覆盖了1%的付费团队",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 625.841,
      "end": 628.501,
      "text": "第24小时是接近13%",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 628.501,
      "end": 631.541,
      "text": "是GPT 5.6系列的两倍",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 631.541,
      "end": 634.401,
      "text": "更是Fable 5.1的六倍多",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 634.401,
      "end": 638.261,
      "text": "而其他近期发布的模型均未超过7%",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 638.261,
      "end": 639.881,
      "text": "首日数据就足以证明",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 639.881,
      "end": 643.741,
      "text": "专用模型投入生产环境的速度有多快",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 643.741,
      "end": 646.881,
      "text": "但真正的考验在于持久性谈到具体用途",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 646.881,
      "end": 650.361,
      "text": "Perser提到智能体可以自主选择下一个工具",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 650.361,
      "end": 655.441,
      "text": "或在子智能体的工作流中决定是继续重视还是询问用户",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 655.441,
      "end": 658.841,
      "text": "亦或在行动前先评估风险并决定是否终止",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 658.841,
      "end": 661.541,
      "text": "并将无法确定的情况转交人工处理",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 661.541,
      "end": 664.681,
      "text": "TypeSafe将其称为智能异服语句或模糊规则",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 664.681,
      "end": 667.161,
      "text": "因为手写逻辑往往过于脆弱",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 667.161,
      "end": 668.401,
      "text": "他们还建议使用Map",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 668.401,
      "end": 671.921,
      "text": "将PB级数据精炼为特征以支持实时应用",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 671.921,
      "end": 675.721,
      "text": "毕竟100毫秒的延迟完全在用户体验的可控范围内",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 675.721,
      "end": 678.661,
      "text": "从推理过程到越狱尝试全程进行验证",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 678.661,
      "end": 681.541,
      "text": "Amita希望人们能用它来监控智能体",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 681.541,
      "end": 683.941,
      "text": "由于智能体互间成本高昂",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 683.941,
      "end": 686.861,
      "text": "Runiker建议采用实时模型路由",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 686.861,
      "end": 688.421,
      "text": "The Register补充到",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 688.421,
      "end": 691.261,
      "text": "聊天机器人人由LLM架构主导",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 691.261,
      "end": 693.321,
      "text": "编程助手代码智能体",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 693.321,
      "end": 696.341,
      "text": "以及数学证明或演示等可验证任务",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 696.341,
      "end": 699.061,
      "text": "而这正如TypeSafe戏称的那样",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 699.061,
      "end": 700.741,
      "text": "效果时好时坏",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 700.741,
      "end": 702.281,
      "text": "至于内部究竟如何",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 702.281,
      "end": 703.641,
      "text": "Amita手口如瓶",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 703.641,
      "end": 706.481,
      "text": "外界普遍怀疑其底层使用的是",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 706.481,
      "end": 707.821,
      "text": "开源权重的大模型",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 707.821,
      "end": 710.641,
      "text": "尽管Jeff撑起FAQ文档内容并不短",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 710.641,
      "end": 714.301,
      "text": "Jeff模型完全采用合成数据进行训练",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 714.301,
      "end": 716.261,
      "text": "Ometa告诉TechCrunch",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 716.261,
      "end": 719.321,
      "text": "资产数据是他人生中最成功的豪赌之一",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 4,
      "language_run": 2
    },
    {
      "start": 719.586,
      "end": 720.686,
      "text": "if",
      "language": "en",
      "chunk": 1,
      "local_language_run": 5,
      "language_run": 3
    },
    {
      "start": 720.686,
      "end": 722.256,
      "text": "比发布产品更成功",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 722.256,
      "end": 723.816,
      "text": "也优于RLHF",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 723.816,
      "end": 727.136,
      "text": "TypeSafe自诩为一家核心的数据研究实验室",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 727.136,
      "end": 729.476,
      "text": "公司半数员工几乎主导了",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 729.476,
      "end": 732.576,
      "text": "统计学可控合成数据这一细分领域",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 732.576,
      "end": 734.396,
      "text": "这已成他毕生乐趣",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 734.396,
      "end": 737.556,
      "text": "即便你请求他们也不会用你的数据训练请见谅",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 737.556,
      "end": 738.736,
      "text": "想了解细节",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 738.736,
      "end": 740.456,
      "text": "得先被他们录用",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 740.456,
      "end": 743.276,
      "text": "这就是他们所奉行的苦涩教训",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 743.276,
      "end": 746.516,
      "text": "优化任务远比堆砌数据和算力重要",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 746.516,
      "end": 749.936,
      "text": "或是RLHF这类针对文本优化的算法",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 749.936,
      "end": 752.156,
      "text": "人类评估员偏好虽然适合聊天",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 752.156,
      "end": 753.516,
      "text": "却不适用于自动化任务",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 753.516,
      "end": 757.916,
      "text": "且RLHF仅适用于程序可教验的任务",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 757.916,
      "end": 760.296,
      "text": "导致其现实判断波动极大",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 760.296,
      "end": 761.796,
      "text": "智能表现不够稳健",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 761.796,
      "end": 765.256,
      "text": "系统1这一名称源自卡尼曼的思考快与曼",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 765.256,
      "end": 768.796,
      "text": "尽管在传统认知中系统1往往意味着一出错",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 768.796,
      "end": 772.816,
      "text": "但他们认为这些模型其实比其他方案更可靠",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 772.816,
      "end": 775.396,
      "text": "他们还完全放弃了公开基准测试",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 775.396,
      "end": 776.856,
      "text": "仅提供一次性评估",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 776.856,
      "end": 779.476,
      "text": "并推动用户自行构建评估体系",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 779.476,
      "end": 782.096,
      "text": "因为系统一任务更易于评估",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 782.096,
      "end": 785.396,
      "text": "甚至还有篇题为反基准优化的文章",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 785.396,
      "end": 787.876,
      "text": "Jeff这个名字源自威廉·Tauphel",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 787.876,
      "end": 791.236,
      "text": "这位19世纪的经济学家提出了一个悖论",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 791.236,
      "end": 793.036,
      "text": "当蒸汽机效率提升",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 793.036,
      "end": 795.176,
      "text": "煤炭需求反而不降反升",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 795.176,
      "end": 796.176,
      "text": "他们压住的是",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 796.176,
      "end": 799.736,
      "text": "智能成本每降低一个数量级",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 799.736,
      "end": 802.736,
      "text": "就会解锁多出数个数量级的应用场景",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 802.736,
      "end": 806.576,
      "text": "阿尔梅达预见智能软件将无处不再呈现出涌现性",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 806.576,
      "end": 808.816,
      "text": "且更像早期互联网那样分布化",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 808.816,
      "end": 810.156,
      "text": "而非如今的超级应用",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 810.156,
      "end": 811.716,
      "text": "但The Register指出",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 811.716,
      "end": 813.176,
      "text": "这建立在一个假设之上",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 813.176,
      "end": 815.056,
      "text": "其Token的需求会像能源一样广泛",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 815.056,
      "end": 817.276,
      "text": "这尚未有定论PO研究中心",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 817.276,
      "end": 819.536,
      "text": "发现许多能用上AI的人",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 819.536,
      "end": 820.716,
      "text": "根本找不到用处",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 820.716,
      "end": 822.476,
      "text": "CNBC调查显示",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 822.476,
      "end": 824.916,
      "text": "另一些人则出于道德顾虑排斥AI",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 824.916,
      "end": 827.536,
      "text": "目前Jeff依然是特例",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 827.536,
      "end": 830.076,
      "text": "但Reneker预计竞争终会出现",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 830.076,
      "end": 832.376,
      "text": "他感叹我们本应早点看透这一点",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 832.376,
      "end": 834.316,
      "text": "毕竟大语言模型太便宜了",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 834.316,
      "end": 836.976,
      "text": "再加上补贴人们往往懒得去创新",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 836.976,
      "end": 840.156,
      "text": "但Type Safe正在开发更多新模态版本",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 840.156,
      "end": 843.216,
      "text": "并开始从候补名单中邀请开发者",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 843.216,
      "end": 845.396,
      "text": "询问他们想实现什么样的科幻构想",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 845.396,
      "end": 848.776,
      "text": "当Type Crunch问起Type Safe是否属于前沿实验室时",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 848.776,
      "end": 851.176,
      "text": "Almeta直言前沿实验室的主打产品",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 851.176,
      "end": 852.336,
      "text": "无非是恐惧和炒作",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 852.336,
      "end": 855.916,
      "text": "它追求的是真正智能而非常无限财富的豪赌",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 855.916,
      "end": 857.316,
      "text": "也不是某种宗教",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 857.316,
      "end": 859.136,
      "text": "更不是在数据中心里创造神",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 859.136,
      "end": 860.516,
      "text": "总之这是Jeff",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 860.516,
      "end": 862.296,
      "text": "欢迎在评论区留言分享",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 862.296,
      "end": 863.456,
      "text": "你会如何使用它",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 863.456,
      "end": 864.096,
      "text": "感谢观看",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 864.096,
      "end": 865.296,
      "text": "我们下期见",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 6,
      "language_run": 4
    },
    {
      "start": 866.436,
      "end": 876.096,
      "text": "Thank you.",
      "language": "en",
      "chunk": 1,
      "local_language_run": 7,
      "language_run": 5
    }
  ],
  "segmented_multilingual_transcription": {
    "status": "STABLE_SEGMENTED_MULTILINGUAL_REVIEW",
    "created_at": "2026-09-28T15:23:52",
    "manifest": "/Users/sagawa/AI_WORK/video_notes/url/20260928_150626__JEV彻底引爆全网_这才是AI的全新形态_p01_中文配音/segmented_work/chunk_manifest.json",
    "audio": "/Users/sagawa/AI_WORK/video_notes/url/20260928_150626__JEV彻底引爆全网_这才是AI的全新形态_p01_中文配音/audio_16k_mono.wav",
    "audio_sha256": "45fd20c20bcfc692e155f5a5b3dfaba1cb4726c21881f5797ff0f80e90fa46ed",
    "audio_duration": 880.570375,
    "model": "/Users/sagawa/AI_WORK/cache/huggingface/hub/models--mlx-community--whisper-large-v3-turbo/snapshots/a4aaeec0636e6fef84abdcbe3544cb2bf7e9f6fb",
    "language_profile": "/Users/sagawa/AI_WORK/video_notes/url/20260928_150626__JEV彻底引爆全网_这才是AI的全新形态_p01_中文配音/multilingual_work/multilingual_language_profile.json",
    "dominant_languages": [
      "zh",
      "en"
    ],
    "suppressed_languages": [],
    "chunk_count": 2,
    "completed_chunks": 2,
    "resumed_chunks": 0,
    "merged_segment_count": 365,
    "detected_languages": [
      "en",
      "zh"
    ],
    "transcript_languages": [
      "en",
      "zh"
    ],
    "missing_languages": [],
    "unexpected_languages": [],
    "language_sequence": [
      "zh",
      "en",
      "zh",
      "en"
    ],
    "speech_region_seconds": 864.9200000000001,
    "speech_covered_seconds": 861.58,
    "speech_coverage_ratio": 0.9961383711788373,
    "tail_repeat_run": 1,
    "tail_repeat_text": "Chad GPT主创人员之一",
    "chunk_needs_review": [
      {
        "chunk": 1,
        "runs": [
          5
        ],
        "exit_code": 43
      }
    ],
    "qc_ok": false,
    "qc_errors": [
      "one or more chunks require review"
    ],
    "complete_review_artifact": true
  }
}