{ "text": "欢迎回到On Insight\n如果我们要选一个节点\n作为全球AI产业\n从学术梦幻期彻底跨入\n现实暴击期的分水里\n我敢打赌\n绝不会是哪家大厂又在PPT里\n发布了多大规模参数的模型\n基准测试又突破了多高的分数\n真正的变局往往发生在那些看起来极度安静\n甚至带着几分诡异的高层人事变动力\n就在最近谷歌内部发生了一场\n足以重塑整个硅谷底层格局的震荡\n在谷歌效力了整整27年\n被全球程序员尊为代码之神的\n谷歌首席科学家Jeff Dean正式宣布离职\n如果这只是一位老将退影的话\n倒也不至于让硅谷各大投资机构和安全部门彻夜未免\n真正认识背后发紧的是\nJFDN离职的时候带走了一个堪称\n硅谷全明星的豪华降维打击阵容\n这其中包括与JFDN合作20多年的底层系统传奇搭档\n桑杰·格马瓦特\n还有Google D.Mine的研究副总裁\n詹姆勒模型最核心的技术操盘手奥里奥尔·维尼亚尔斯\n以及AutoML领域先驱和Google Brain的联合创始人黎光莲\n这四个人加在一起\n几乎就是半个谷歌AI与底层分布式系统的技术灵魂\n他们组建了一家名为Discovery Loop的新公司\n按理说面对如此核心的技术骨干离职\n谷歌应该爆发一场血腥的公关战\n对吧\n但诡异的事情来了\n在继续后面的内容之前\n我想先请你帮我一个小忙\n如果你经常观看我的视频\n但还没有订阅我的频道\n欢迎点击订阅\n这样你就不会错过\n后续更多的科学资讯和深度解析\n那如果你觉得这期视频对你有所帮助\n也请帮忙点赞和分享\n让更多人看到这些内容\n那如果你希望给予我更多的支持\n也欢迎加入我的会员\n每月一杯咖啡的钱\n就是对我持续创作最大的支持\n好的我们继续\n这个诡异的事情就是\n谷歌不仅没有阻拦\n反而第一时间成为了Discovery Loop的初始投资方\n并且双手奉上了Google Cloud的超级算力集群\n双方直接签下了长期云服务绑定协议\n不仅如此\n就在同一天\nGoogle DeepMind内部也完成了一场\n近乎手术刀级别的高层重组\nDeepMind的灵魂人物Demith Hossabies\n正式卸任DeepMind CEO的日常管理\n身任Alphabet的首席科学家\n去操盘更为宏大的ATI理论中局\n而接替他全盘操盘Google DeepMind的新帅\n是原DevMind的CTO\nDQN与WaveNet的发明人\n科雷·卡伍克·乔鲁\n他越过了所有中间管理层\n直接向谷歌CEO桑德尔PX1汇报\n听到这里\n你脑海里一定盘旋着一个极其巨大的悬念\n就是谷歌为什么会心甘情愿的\n让自己的顶尖大脑集体外逃\n甚至还要掏钱掏算力\n帮他们去创业呢\n而那位既懂DQN强化学习\n又懂Vivnet原生语音的工程狂人\n柯雷接棒之后\n到底要把Gemlet带向何方呢\n今天我们就有一期节目\n彻底拆解这场看似是高层离职潮\n实则是谷歌AI技术大分流的\n硅谷终极棋局\n要解开这个谜团\n我们首先得搞清楚\nJeff Dean带着这帮神仙阵容杀出去\n到底要去干一件什么样的事情\n在Discovery Loop的官宣里\nJeff Dean直接描绘了公司的愿景\n那就是构建自动化的科学方法闭环\n简单来说\n过去几百年\n人类探索未知的方式\n本质上都是串行实验\n就是科学家提出假设\n亲手写代码\n跑一次实验\n等几天看结果\n分析数据\n然后再手动修改参数\n跑下一次\n在这套流程里\n最瓶颈的变量不是算力\n而是人类脑力的思考速度\n与手动操作的物理实验\n而Discovery Loop要构建一套\n完全自主运行的AIA层架构\n做到自动提出科研假设\n自动编写实验代码\n自动部署分布式计算\n自动分析实验指标\n最后自动递归迭代\n你注意看Discovery Loop的\n手机应用场景选择\n他们没有第一天就跑去见物理实验室\n而是选择吃自己的狗粮\n把机器学习研发本身作为第一个实验战场\n也就是说他们要用AI去自动设计\n自动搜索\n自动验证下一代超越Transformer的模型架构\n很多朋友可能会问\n用AI自动写代码做研究\n之前日本Sakana AI的项目不是也做过吗\n这正是我们接下来要拆解的关键\nSatan的AI的项目本质上是一个应用层的agent流程封装\n它是拿现成的GPT4O或者Cloud外包一层prompt\n去模拟科学家写代码跑实验写论文\n它的上限完全被底下调用的通用大模型给封死了\n而Discovery Loop完全是另一个维度的存在\nJeff Dean和桑杰·格瓦马特是写出MapReduce Bigtable的超级系统架构师\n李光莲是AutoML的鼻祖\n奥里尔尔维尼亚尔斯\n是顶尖的神经网络搜索与多模态专家\n他们要做的是从底层分布式计算\n硬件算力调度\n神经架构搜索到算法闭环的系统级重构\n他们不是在用AI写论文\n他们是在用超级计算机集群\n让AI去做数以万计的硬件级与架构级实验\n直接在算法底层寻找下一个Transformer\n那么回到我们开头的悬念\n就是谷歌为什么不把这个超级团队留在内部\n而是支持他们独立成立一家公共利益公司呢\n因为谷歌算透了一笔账\n像AI自主研发下一代AI这种前沿探索\n风险极高烧钱极大失败率也极高\n如果放在谷歌Digmy内部\n研究员们每天要面对华尔街的财报审查和季度KPI\n就会被无限牵扯精力去搞产品适配\n通过把JFDN团队外部化\n谷歌既让这一帮顶尖天才\n摆脱了大公司病和商业汇报的枷锁\n又能通过初始投资方和独家GoCloud算力的身份\n牢牢锁定未来潜在的技术回流与云服务收益\n这一招叫做前沿高风险战略的外部结偶\n如果说Jeff Dean的离职是把远期高风险探索拆分了出去\n那么接下来发生的事情才是对Jamlight产品生态影响最深远\n也是最暴力的一环\n接掌Google D.My的全局率颖的科雷卡伍克·乔武\n是一个与Demis Hossabies风格截然不同的工程技术派\nDemis是国际象棋神童和神经科学家\n它充满着理想主义\n而柯雷是把深度强化学习\n真正推向AlphaGo破局\n把WaveNet徐徐语音\n真正变成极异手机里语音引擎的超级工程操盘手\n在柯雷掌舵之后\nGoogle G-Mind释放出了一个极其明确的信号\n就是全面放弃学术包袱\n转向极致工程演进与大规模商业化暴击\n在他的主持下\nGemlet的底层架构\nGamlight的底层架构正在发生三重极其硬核的眼镜\n首先最根本的变化是\n彻底打破对预训练参数崇拜的单位幻想\n全面转向后训练与推理期计算的深度融合\n在过去两年行业盲目堆叠预训练参数\n但预训练越往后编辑效应递减越严重\n算力成本高到令人发指\n柯雷结合他在DQN强化学习领域的积累\n给Gemla引入了过程奖励模型PRM\n与蒙特卡洛书搜索MCTS\n打个比方传统的语言模型\n就像在考场上没有任何草稿纸\n靠直觉备答案的考生\n遇到简单的题那是没问题的\n但遇到复杂逻辑推导或者长流程编程\n中间要是错一步后面就全乱了\n而PRM与MCTS的融合\n相当于给Gemla配了一张无线扩展的虚拟草稿纸\n和一位微观监考官\n在推理的每一步\nPRM都会对它的逻辑合理性进行毫秒级评分\n同时MCTS数搜索会在后台预判\n未来几十步的走向\n如果发现某个分支走向死胡同\n模型会在推理阶段自动撤回\n并重新搜索最右键\n这意味着\n在不用无限膨胀预讯链参数的前提下\nGEMLINE处理高难度编程\n复杂数学推导和A乘的规划的成功率\n称指数级提升\n紧接着第二种转变在于\n系统级推理效率的极限拉链\n那就是构建Gemini Spark端云协同分层架构\n这里稍微提一下我个人对Spark的使用体验\n虽然它还是Beta版本\n但先说结论\n我觉得它超级好用\n我会单独出一期视频\n聊一下Gemini Spark和Hermes Agent的配合使用\n作为Google订阅的长期体验者\n我明显感觉最近Gemini的使用限额提高了\n他们虽然没有发布像Fable 5这类的模型\n但是无论是Gemini 3.6 Flash\n还是Gemlet 3.1 Pro都变得更聪明了\n好我们继续这个第二重转变\n在商业高并发场景中\n用户问一句今天天气怎么样\n如果都要调用一次千亿参数的单体大模型\n显存和电力开销能把大厂烧穿\n克雷推动的Gemlet Spark架构\n把整套AI系统做成了像特种部队一样的分层调度\n前端部署及轻量及低延迟的Gemlet Spark代理模型\n负责高频意图识别\n上下门过滤和轻量任务\n只有当前线判定\n涉及复杂长链条逻辑的时候\n才会无缝调动云端期间大模型\n这种分层架构\n直接把单位Token成本打了下来\n这才是支撑数亿并发的商业硬实力\n而第三种变化\n则是柯雷将他的看家本领\nVIVNET原生音频与流逝生成\n深度融合到了詹姆賴的全模态底座中\n以往的多模态AI\n往往是文本模型外挂语音识别\n与语音合成模块的拼接车\n延迟高语调生硬\n而柯雷主导下的詹姆賴\n实现了真正的原生全模态流逝交互\n不仅能实现毫秒级\n实时音视频双向对打\n还能在听懂说话的同时\n自主执行代码\n甚至在手机屏幕上动态生成\nUI界面与交互空间\n这已经不是在聊天了\n这是直接把自然语言变成了实时软件生成的界面\n那底层架构演进的再漂亮\n如果不能转化为商业流水\n在今天也是一张废纸\n科雷·卡伍克·乔鲁这次最让人惊悚的调整\n在于他不仅管科研\n还同时兼任了Gemline产品与开发者生态的最高指挥官\n在过去的大厂体制里\n科研实验室和商品产品部门是完全脱节的\n实验室发了论文\n产品部门要花大半年去沟通打包\n而柯雷一把抓之后\n直接祭出了商业落地的三大杀招\n在终端入口上\n他们开始加速清洗传统的语音助手\n全面用GM来替换掉Android和VLOS上\n那个指挥社闹钟的assistant\n赋予他操作系统级的最高上下文读取权限\n与跨应用API调用能力\n直接把它打造成操作系统的中央控制大脑\n比如在ValOS手表或者车载系统里\n你只需要说帮我安排今晚与张总的商务晚餐\n并准备一份他公司最新财报的背景简报\nGamlight就会自动调取邮件和地图订餐厅\n同时在云端提取财报生成简报推送给你\n与此同时\n他们还大幅压缩了科研成果\n转化到商业API的时间\n在谷歌云Vertex AI和开发者API上\n前沿团队在Achern代码生成\n自动模型评估和推理加速上的突破\n能以周级别的速度直接部署上线\n让科研成果在第一秒\n直接转化为商业账单\n形成技术产品与算力的超级费轮\n更关键的是\n谷歌正在悄无声息地改变自己的盈利模式\n当用户不再翻页点击搜索广告链接\n而是让Gamlet给出最终答案时\n谷歌的解法是\n直接占据agent决策路径的入口\n当Gamlet在每日简报或者个人助理里\n帮你规划行程推荐服务时\n它在生成方案的同时\n精准接入商业API入口\n谷歌由此从一个信息检索广告商\n变成了全局智能决策的开门人\n抽取服务分账和佣金\n拆决员Jeff Dean的离职创业\n与柯磊掌舵下的Deep Mind重组\n我们终于能看清这场硅谷AI大戏的全貌了\n硅谷的AI战争正是告别了依靠概念和论文引用量\n获取高估值的学术竞赛阶段\n全面跨入了拼推理成本\n拼系统稳定性\n拼终端控制力\n以及拼商业造学能力的工程与商业深水去\n对于我们所有在做AI研发架构设计\n或者技术创业的朋友来说\n谷歌的这一系列辩证\n留下了三条极其宝贵的工程启发\n首先不要盲目迷信\n纯域训练参数的单位堆叠\n多去研究推理器计算\n通过过程奖励模型打分与蒙特卡洛数搜索\n让模型在推理阶段学会自我纠错\n这是低成本实现逻辑跨越的唯一解法\n其次要学会拥抱端云分层架构\n平气无论任务大小\n一力调用千亿大模型的粗暴思维\n学习Jamler Spark的设计\n用轻量级模型做好前置意图路由\n才能在高并发任务中守住成本底线\n最后我们需要重新思考软件入口的定义\n未来的软件竞争不再是争夺\n手机屏幕上的APP图标位置\n而是争夺能否被系统及agent\nAPI决策链所调用\n技术不相信幻觉\n商业不相信眼泪\n这里是All Insight\n我们下期见", "language": null, "duration": 929.100063, "segments": [ { "start": 0.0, "end": 11.48, "text": "欢迎回到On Insight", "chunk": 0, "language": "zh", "avg_logprob": -0.07785013343105797, "no_speech_prob": 1.5937284478240166e-11, "compression_ratio": 1.0909090909090908 }, { "start": 11.48, "end": 14.12, "text": "如果我们要选一个节点", "chunk": 0, "language": "zh", "avg_logprob": -0.07785013343105797, "no_speech_prob": 1.5937284478240166e-11, "compression_ratio": 1.0909090909090908 }, { "start": 14.12, "end": 15.92, "text": "作为全球AI产业", "chunk": 0, "language": "zh", "avg_logprob": -0.07785013343105797, "no_speech_prob": 1.5937284478240166e-11, "compression_ratio": 1.0909090909090908 }, { "start": 15.92, "end": 18.72, "text": "从学术梦幻期彻底跨入", "chunk": 0, "language": "zh", "avg_logprob": -0.07785013343105797, "no_speech_prob": 1.5937284478240166e-11, "compression_ratio": 1.0909090909090908 }, { "start": 18.72, "end": 20.52, "text": "现实暴击期的分水里", "chunk": 0, "language": "zh", "avg_logprob": -0.07785013343105797, "no_speech_prob": 1.5937284478240166e-11, "compression_ratio": 1.0909090909090908 }, { "start": 20.52, "end": 21.44, "text": "我敢打赌", "chunk": 0, "language": "zh", "avg_logprob": -0.07785013343105797, "no_speech_prob": 1.5937284478240166e-11, "compression_ratio": 1.0909090909090908 }, { "start": 21.44, "end": 24.18, "text": "绝不会是哪家大厂又在PPT里", "chunk": 0, "language": "zh", "avg_logprob": -0.07785013343105797, "no_speech_prob": 1.5937284478240166e-11, "compression_ratio": 1.0909090909090908 }, { "start": 24.18, "end": 26.48, "text": "发布了多大规模参数的模型", "chunk": 0, "language": "zh", "avg_logprob": -0.07785013343105797, "no_speech_prob": 1.5937284478240166e-11, "compression_ratio": 1.0909090909090908 }, { "start": 26.48, "end": 28.9, "text": "基准测试又突破了多高的分数", "chunk": 0, "language": "zh", "avg_logprob": -0.07785013343105797, "no_speech_prob": 1.5937284478240166e-11, "compression_ratio": 1.0909090909090908 }, { "start": 28.9, "end": 33.68, "text": "真正的变局往往发生在那些看起来极度安静", "chunk": 0, "language": "zh", "avg_logprob": -0.07811544467876484, "no_speech_prob": 1.3255970973680142e-11, "compression_ratio": 1.1986062717770034 }, { "start": 33.68, "end": 37.24, "text": "甚至带着几分诡异的高层人事变动力", "chunk": 0, "language": "zh", "avg_logprob": -0.07811544467876484, "no_speech_prob": 1.3255970973680142e-11, "compression_ratio": 1.1986062717770034 }, { "start": 37.24, "end": 39.78, "text": "就在最近谷歌内部发生了一场", "chunk": 0, "language": "zh", "avg_logprob": -0.07811544467876484, "no_speech_prob": 1.3255970973680142e-11, "compression_ratio": 1.1986062717770034 }, { "start": 39.78, "end": 43.4, "text": "足以重塑整个硅谷底层格局的震荡", "chunk": 0, "language": "zh", "avg_logprob": -0.07811544467876484, "no_speech_prob": 1.3255970973680142e-11, "compression_ratio": 1.1986062717770034 }, { "start": 43.4, "end": 46.04, "text": "在谷歌效力了整整27年", "chunk": 0, "language": "zh", "avg_logprob": -0.07811544467876484, "no_speech_prob": 1.3255970973680142e-11, "compression_ratio": 1.1986062717770034 }, { "start": 46.04, "end": 49.26, "text": "被全球程序员尊为代码之神的", "chunk": 0, "language": "zh", "avg_logprob": -0.07811544467876484, "no_speech_prob": 1.3255970973680142e-11, "compression_ratio": 1.1986062717770034 }, { "start": 49.26, "end": 52.44, "text": "谷歌首席科学家Jeff Dean正式宣布离职", "chunk": 0, "language": "zh", "avg_logprob": -0.07811544467876484, "no_speech_prob": 1.3255970973680142e-11, "compression_ratio": 1.1986062717770034 }, { "start": 52.44, "end": 55.24, "text": "如果这只是一位老将退影的话", "chunk": 0, "language": "zh", "avg_logprob": -0.07811544467876484, "no_speech_prob": 1.3255970973680142e-11, "compression_ratio": 1.1986062717770034 }, { "start": 55.24, "end": 59.96, "text": "倒也不至于让硅谷各大投资机构和安全部门彻夜未免", "chunk": 0, "language": "zh", "avg_logprob": -0.1514168990285773, "no_speech_prob": 1.5211231579881357e-11, "compression_ratio": 1.1446153846153846 }, { "start": 59.96, "end": 62.32, "text": "真正认识背后发紧的是", "chunk": 0, "language": "zh", "avg_logprob": -0.1514168990285773, "no_speech_prob": 1.5211231579881357e-11, "compression_ratio": 1.1446153846153846 }, { "start": 62.32, "end": 65.52, "text": "JFDN离职的时候带走了一个堪称", "chunk": 0, "language": "zh", "avg_logprob": -0.1514168990285773, "no_speech_prob": 1.5211231579881357e-11, "compression_ratio": 1.1446153846153846 }, { "start": 65.52, "end": 69.18, "text": "硅谷全明星的豪华降维打击阵容", "chunk": 0, "language": "zh", "avg_logprob": -0.1514168990285773, "no_speech_prob": 1.5211231579881357e-11, "compression_ratio": 1.1446153846153846 }, { "start": 69.18, "end": 74.66, "text": "这其中包括与JFDN合作20多年的底层系统传奇搭档", "chunk": 0, "language": "zh", "avg_logprob": -0.1514168990285773, "no_speech_prob": 1.5211231579881357e-11, "compression_ratio": 1.1446153846153846 }, { "start": 74.66, "end": 76.38, "text": "桑杰·格马瓦特", "chunk": 0, "language": "zh", "avg_logprob": -0.1514168990285773, "no_speech_prob": 1.5211231579881357e-11, "compression_ratio": 1.1446153846153846 }, { "start": 76.38, "end": 79.16, "text": "还有Google D.Mine的研究副总裁", "chunk": 0, "language": "zh", "avg_logprob": -0.1514168990285773, "no_speech_prob": 1.5211231579881357e-11, "compression_ratio": 1.1446153846153846 }, { "start": 79.16, "end": 83.94, "text": "詹姆勒模型最核心的技术操盘手奥里奥尔·维尼亚尔斯", "chunk": 0, "language": "zh", "avg_logprob": -0.1514168990285773, "no_speech_prob": 1.5211231579881357e-11, "compression_ratio": 1.1446153846153846 }, { "start": 83.94, "end": 89.0, "text": "以及AutoML领域先驱和Google Brain的联合创始人黎光莲", "chunk": 0, "language": "zh", "avg_logprob": -0.08055820021518441, "no_speech_prob": 1.4224471427126684e-11, "compression_ratio": 1.1636904761904763 }, { "start": 89.0, "end": 91.06, "text": "这四个人加在一起", "chunk": 0, "language": "zh", "avg_logprob": -0.08055820021518441, "no_speech_prob": 1.4224471427126684e-11, "compression_ratio": 1.1636904761904763 }, { "start": 91.06, "end": 96.04, "text": "几乎就是半个谷歌AI与底层分布式系统的技术灵魂", "chunk": 0, "language": "zh", "avg_logprob": -0.08055820021518441, "no_speech_prob": 1.4224471427126684e-11, "compression_ratio": 1.1636904761904763 }, { "start": 96.04, "end": 99.6, "text": "他们组建了一家名为Discovery Loop的新公司", "chunk": 0, "language": "zh", "avg_logprob": -0.08055820021518441, "no_speech_prob": 1.4224471427126684e-11, "compression_ratio": 1.1636904761904763 }, { "start": 99.6, "end": 103.62, "text": "按理说面对如此核心的技术骨干离职", "chunk": 0, "language": "zh", "avg_logprob": -0.08055820021518441, "no_speech_prob": 1.4224471427126684e-11, "compression_ratio": 1.1636904761904763 }, { "start": 103.62, "end": 106.7, "text": "谷歌应该爆发一场血腥的公关战", "chunk": 0, "language": "zh", "avg_logprob": -0.08055820021518441, "no_speech_prob": 1.4224471427126684e-11, "compression_ratio": 1.1636904761904763 }, { "start": 106.7, "end": 107.3, "text": "对吧", "chunk": 0, "language": "zh", "avg_logprob": -0.08055820021518441, "no_speech_prob": 1.4224471427126684e-11, "compression_ratio": 1.1636904761904763 }, { "start": 107.3, "end": 109.0, "text": "但诡异的事情来了", "chunk": 0, "language": "zh", "avg_logprob": -0.08055820021518441, "no_speech_prob": 1.4224471427126684e-11, "compression_ratio": 1.1636904761904763 }, { "start": 109.0, "end": 111.18, "text": "在继续后面的内容之前", "chunk": 0, "language": "zh", "avg_logprob": -0.08055820021518441, "no_speech_prob": 1.4224471427126684e-11, "compression_ratio": 1.1636904761904763 }, { "start": 111.18, "end": 113.24, "text": "我想先请你帮我一个小忙", "chunk": 0, "language": "zh", "avg_logprob": -0.08055820021518441, "no_speech_prob": 1.4224471427126684e-11, "compression_ratio": 1.1636904761904763 }, { "start": 113.24, "end": 115.1, "text": "如果你经常观看我的视频", "chunk": 0, "language": "zh", "avg_logprob": -0.08728083307871562, "no_speech_prob": 1.5094826430472885e-11, "compression_ratio": 1.4158415841584158 }, { "start": 115.1, "end": 116.5, "text": "但还没有订阅我的频道", "chunk": 0, "language": "zh", "avg_logprob": -0.08728083307871562, "no_speech_prob": 1.5094826430472885e-11, "compression_ratio": 1.4158415841584158 }, { "start": 116.5, "end": 118.04, "text": "欢迎点击订阅", "chunk": 0, "language": "zh", "avg_logprob": -0.08728083307871562, "no_speech_prob": 1.5094826430472885e-11, "compression_ratio": 1.4158415841584158 }, { "start": 118.04, "end": 119.76, "text": "这样你就不会错过", "chunk": 0, "language": "zh", "avg_logprob": -0.08728083307871562, "no_speech_prob": 1.5094826430472885e-11, "compression_ratio": 1.4158415841584158 }, { "start": 119.76, "end": 122.46, "text": "后续更多的科学资讯和深度解析", "chunk": 0, "language": "zh", "avg_logprob": -0.08728083307871562, "no_speech_prob": 1.5094826430472885e-11, "compression_ratio": 1.4158415841584158 }, { "start": 122.46, "end": 125.8, "text": "那如果你觉得这期视频对你有所帮助", "chunk": 0, "language": "zh", "avg_logprob": -0.08728083307871562, "no_speech_prob": 1.5094826430472885e-11, "compression_ratio": 1.4158415841584158 }, { "start": 125.8, "end": 127.58, "text": "也请帮忙点赞和分享", "chunk": 0, "language": "zh", "avg_logprob": -0.08728083307871562, "no_speech_prob": 1.5094826430472885e-11, "compression_ratio": 1.4158415841584158 }, { "start": 127.58, "end": 129.36, "text": "让更多人看到这些内容", "chunk": 0, "language": "zh", "avg_logprob": -0.08728083307871562, "no_speech_prob": 1.5094826430472885e-11, "compression_ratio": 1.4158415841584158 }, { "start": 129.36, "end": 132.28, "text": "那如果你希望给予我更多的支持", "chunk": 0, "language": "zh", "avg_logprob": -0.08728083307871562, "no_speech_prob": 1.5094826430472885e-11, "compression_ratio": 1.4158415841584158 }, { "start": 132.28, "end": 133.88, "text": "也欢迎加入我的会员", "chunk": 0, "language": "zh", "avg_logprob": -0.08728083307871562, "no_speech_prob": 1.5094826430472885e-11, "compression_ratio": 1.4158415841584158 }, { "start": 133.88, "end": 135.74, "text": "每月一杯咖啡的钱", "chunk": 0, "language": "zh", "avg_logprob": -0.08728083307871562, "no_speech_prob": 1.5094826430472885e-11, "compression_ratio": 1.4158415841584158 }, { "start": 135.74, "end": 138.44, "text": "就是对我持续创作最大的支持", "chunk": 0, "language": "zh", "avg_logprob": -0.08728083307871562, "no_speech_prob": 1.5094826430472885e-11, "compression_ratio": 1.4158415841584158 }, { "start": 138.44, "end": 139.9, "text": "好的我们继续", "chunk": 0, "language": "zh", "avg_logprob": -0.08728083307871562, "no_speech_prob": 1.5094826430472885e-11, "compression_ratio": 1.4158415841584158 }, { "start": 139.9, "end": 141.84, "text": "这个诡异的事情就是", "chunk": 0, "language": "zh", "avg_logprob": -0.08728083307871562, "no_speech_prob": 1.5094826430472885e-11, "compression_ratio": 1.4158415841584158 }, { "start": 141.84, "end": 143.76, "text": "谷歌不仅没有阻拦", "chunk": 0, "language": "zh", "avg_logprob": -0.12579099004140157, "no_speech_prob": 1.8713076227472314e-11, "compression_ratio": 1.1514195583596214 }, { "start": 143.76, "end": 148.0, "text": "反而第一时间成为了Discovery Loop的初始投资方", "chunk": 0, "language": "zh", "avg_logprob": -0.12579099004140157, "no_speech_prob": 1.8713076227472314e-11, "compression_ratio": 1.1514195583596214 }, { "start": 148.0, "end": 152.24, "text": "并且双手奉上了Google Cloud的超级算力集群", "chunk": 0, "language": "zh", "avg_logprob": -0.12579099004140157, "no_speech_prob": 1.8713076227472314e-11, "compression_ratio": 1.1514195583596214 }, { "start": 152.24, "end": 156.54, "text": "双方直接签下了长期云服务绑定协议", "chunk": 0, "language": "zh", "avg_logprob": -0.12579099004140157, "no_speech_prob": 1.8713076227472314e-11, "compression_ratio": 1.1514195583596214 }, { "start": 156.54, "end": 157.12, "text": "不仅如此", "chunk": 0, "language": "zh", "avg_logprob": -0.12579099004140157, "no_speech_prob": 1.8713076227472314e-11, "compression_ratio": 1.1514195583596214 }, { "start": 157.12, "end": 158.34, "text": "就在同一天", "chunk": 0, "language": "zh", "avg_logprob": -0.12579099004140157, "no_speech_prob": 1.8713076227472314e-11, "compression_ratio": 1.1514195583596214 }, { "start": 158.34, "end": 160.68, "text": "Google DeepMind内部也完成了一场", "chunk": 0, "language": "zh", "avg_logprob": -0.12579099004140157, "no_speech_prob": 1.8713076227472314e-11, "compression_ratio": 1.1514195583596214 }, { "start": 160.68, "end": 163.22, "text": "近乎手术刀级别的高层重组", "chunk": 0, "language": "zh", "avg_logprob": -0.12579099004140157, "no_speech_prob": 1.8713076227472314e-11, "compression_ratio": 1.1514195583596214 }, { "start": 163.22, "end": 165.96, "text": "DeepMind的灵魂人物Demith Hossabies", "chunk": 0, "language": "zh", "avg_logprob": -0.12579099004140157, "no_speech_prob": 1.8713076227472314e-11, "compression_ratio": 1.1514195583596214 }, { "start": 165.96, "end": 168.94, "text": "正式卸任DeepMind CEO的日常管理", "chunk": 0, "language": "zh", "avg_logprob": -0.12579099004140157, "no_speech_prob": 1.8713076227472314e-11, "compression_ratio": 1.1514195583596214 }, { "start": 168.94, "end": 171.46, "text": "身任Alphabet的首席科学家", "chunk": 0, "language": "zh", "avg_logprob": -0.1509473954124012, "no_speech_prob": 1.7337898478020364e-11, "compression_ratio": 1.124223602484472 }, { "start": 171.46, "end": 174.52, "text": "去操盘更为宏大的ATI理论中局", "chunk": 0, "language": "zh", "avg_logprob": -0.1509473954124012, "no_speech_prob": 1.7337898478020364e-11, "compression_ratio": 1.124223602484472 }, { "start": 174.52, "end": 177.88, "text": "而接替他全盘操盘Google DeepMind的新帅", "chunk": 0, "language": "zh", "avg_logprob": -0.1509473954124012, "no_speech_prob": 1.7337898478020364e-11, "compression_ratio": 1.124223602484472 }, { "start": 177.88, "end": 180.14, "text": "是原DevMind的CTO", "chunk": 0, "language": "zh", "avg_logprob": -0.1509473954124012, "no_speech_prob": 1.7337898478020364e-11, "compression_ratio": 1.124223602484472 }, { "start": 180.14, "end": 182.52, "text": "DQN与WaveNet的发明人", "chunk": 0, "language": "zh", "avg_logprob": -0.1509473954124012, "no_speech_prob": 1.7337898478020364e-11, "compression_ratio": 1.124223602484472 }, { "start": 182.52, "end": 184.0, "text": "科雷·卡伍克·乔鲁", "chunk": 0, "language": "zh", "avg_logprob": -0.1509473954124012, "no_speech_prob": 1.7337898478020364e-11, "compression_ratio": 1.124223602484472 }, { "start": 184.0, "end": 186.56, "text": "他越过了所有中间管理层", "chunk": 0, "language": "zh", "avg_logprob": -0.1509473954124012, "no_speech_prob": 1.7337898478020364e-11, "compression_ratio": 1.124223602484472 }, { "start": 186.56, "end": 189.44, "text": "直接向谷歌CEO桑德尔PX1汇报", "chunk": 0, "language": "zh", "avg_logprob": -0.1509473954124012, "no_speech_prob": 1.7337898478020364e-11, "compression_ratio": 1.124223602484472 }, { "start": 189.44, "end": 190.66, "text": "听到这里", "chunk": 0, "language": "zh", "avg_logprob": -0.1509473954124012, "no_speech_prob": 1.7337898478020364e-11, "compression_ratio": 1.124223602484472 }, { "start": 190.66, "end": 194.42, "text": "你脑海里一定盘旋着一个极其巨大的悬念", "chunk": 0, "language": "zh", "avg_logprob": -0.1509473954124012, "no_speech_prob": 1.7337898478020364e-11, "compression_ratio": 1.124223602484472 }, { "start": 194.42, "end": 196.94, "text": "就是谷歌为什么会心甘情愿的", "chunk": 0, "language": "zh", "avg_logprob": -0.1509473954124012, "no_speech_prob": 1.7337898478020364e-11, "compression_ratio": 1.124223602484472 }, { "start": 196.94, "end": 199.22, "text": "让自己的顶尖大脑集体外逃", "chunk": 0, "language": "zh", "avg_logprob": -0.1243240714731796, "no_speech_prob": 1.7431726201388997e-11, "compression_ratio": 1.147239263803681 }, { "start": 199.22, "end": 201.34, "text": "甚至还要掏钱掏算力", "chunk": 0, "language": "zh", "avg_logprob": -0.1243240714731796, "no_speech_prob": 1.7431726201388997e-11, "compression_ratio": 1.147239263803681 }, { "start": 201.34, "end": 202.74, "text": "帮他们去创业呢", "chunk": 0, "language": "zh", "avg_logprob": -0.1243240714731796, "no_speech_prob": 1.7431726201388997e-11, "compression_ratio": 1.147239263803681 }, { "start": 202.74, "end": 205.74, "text": "而那位既懂DQN强化学习", "chunk": 0, "language": "zh", "avg_logprob": -0.1243240714731796, "no_speech_prob": 1.7431726201388997e-11, "compression_ratio": 1.147239263803681 }, { "start": 205.74, "end": 208.88, "text": "又懂Vivnet原生语音的工程狂人", "chunk": 0, "language": "zh", "avg_logprob": -0.1243240714731796, "no_speech_prob": 1.7431726201388997e-11, "compression_ratio": 1.147239263803681 }, { "start": 208.88, "end": 210.26, "text": "柯雷接棒之后", "chunk": 0, "language": "zh", "avg_logprob": -0.1243240714731796, "no_speech_prob": 1.7431726201388997e-11, "compression_ratio": 1.147239263803681 }, { "start": 210.26, "end": 212.84, "text": "到底要把Gemlet带向何方呢", "chunk": 0, "language": "zh", "avg_logprob": -0.1243240714731796, "no_speech_prob": 1.7431726201388997e-11, "compression_ratio": 1.147239263803681 }, { "start": 212.84, "end": 215.32, "text": "今天我们就有一期节目", "chunk": 0, "language": "zh", "avg_logprob": -0.1243240714731796, "no_speech_prob": 1.7431726201388997e-11, "compression_ratio": 1.147239263803681 }, { "start": 215.32, "end": 218.46, "text": "彻底拆解这场看似是高层离职潮", "chunk": 0, "language": "zh", "avg_logprob": -0.1243240714731796, "no_speech_prob": 1.7431726201388997e-11, "compression_ratio": 1.147239263803681 }, { "start": 218.46, "end": 221.72, "text": "实则是谷歌AI技术大分流的", "chunk": 0, "language": "zh", "avg_logprob": -0.1243240714731796, "no_speech_prob": 1.7431726201388997e-11, "compression_ratio": 1.147239263803681 }, { "start": 221.72, "end": 223.12, "text": "硅谷终极棋局", "chunk": 0, "language": "zh", "avg_logprob": -0.1243240714731796, "no_speech_prob": 1.7431726201388997e-11, "compression_ratio": 1.147239263803681 }, { "start": 223.12, "end": 224.82, "text": "要解开这个谜团", "chunk": 0, "language": "zh", "avg_logprob": -0.1243240714731796, "no_speech_prob": 1.7431726201388997e-11, "compression_ratio": 1.147239263803681 }, { "start": 224.82, "end": 226.24, "text": "我们首先得搞清楚", "chunk": 0, "language": "zh", "avg_logprob": -0.1243240714731796, "no_speech_prob": 1.7431726201388997e-11, "compression_ratio": 1.147239263803681 }, { "start": 226.24, "end": 229.36, "text": "Jeff Dean带着这帮神仙阵容杀出去", "chunk": 0, "language": "zh", "avg_logprob": -0.09454160325982597, "no_speech_prob": 1.7845730101995194e-11, "compression_ratio": 1.1787878787878787 }, { "start": 229.36, "end": 231.7, "text": "到底要去干一件什么样的事情", "chunk": 0, "language": "zh", "avg_logprob": -0.09454160325982597, "no_speech_prob": 1.7845730101995194e-11, "compression_ratio": 1.1787878787878787 }, { "start": 231.7, "end": 234.2, "text": "在Discovery Loop的官宣里", "chunk": 0, "language": "zh", "avg_logprob": -0.09454160325982597, "no_speech_prob": 1.7845730101995194e-11, "compression_ratio": 1.1787878787878787 }, { "start": 234.2, "end": 236.96, "text": "Jeff Dean直接描绘了公司的愿景", "chunk": 0, "language": "zh", "avg_logprob": -0.09454160325982597, "no_speech_prob": 1.7845730101995194e-11, "compression_ratio": 1.1787878787878787 }, { "start": 236.96, "end": 240.92, "text": "那就是构建自动化的科学方法闭环", "chunk": 0, "language": "zh", "avg_logprob": -0.09454160325982597, "no_speech_prob": 1.7845730101995194e-11, "compression_ratio": 1.1787878787878787 }, { "start": 240.92, "end": 241.9, "text": "简单来说", "chunk": 0, "language": "zh", "avg_logprob": -0.09454160325982597, "no_speech_prob": 1.7845730101995194e-11, "compression_ratio": 1.1787878787878787 }, { "start": 241.9, "end": 243.38, "text": "过去几百年", "chunk": 0, "language": "zh", "avg_logprob": -0.09454160325982597, "no_speech_prob": 1.7845730101995194e-11, "compression_ratio": 1.1787878787878787 }, { "start": 243.38, "end": 245.5, "text": "人类探索未知的方式", "chunk": 0, "language": "zh", "avg_logprob": -0.09454160325982597, "no_speech_prob": 1.7845730101995194e-11, "compression_ratio": 1.1787878787878787 }, { "start": 245.5, "end": 247.3, "text": "本质上都是串行实验", "chunk": 0, "language": "zh", "avg_logprob": -0.09454160325982597, "no_speech_prob": 1.7845730101995194e-11, "compression_ratio": 1.1787878787878787 }, { "start": 247.3, "end": 249.6, "text": "就是科学家提出假设", "chunk": 0, "language": "zh", "avg_logprob": -0.09454160325982597, "no_speech_prob": 1.7845730101995194e-11, "compression_ratio": 1.1787878787878787 }, { "start": 249.6, "end": 250.88, "text": "亲手写代码", "chunk": 0, "language": "zh", "avg_logprob": -0.09454160325982597, "no_speech_prob": 1.7845730101995194e-11, "compression_ratio": 1.1787878787878787 }, { "start": 250.88, "end": 251.76, "text": "跑一次实验", "chunk": 0, "language": "zh", "avg_logprob": -0.09454160325982597, "no_speech_prob": 1.7845730101995194e-11, "compression_ratio": 1.1787878787878787 }, { "start": 251.76, "end": 253.14, "text": "等几天看结果", "chunk": 0, "language": "zh", "avg_logprob": -0.09454160325982597, "no_speech_prob": 1.7845730101995194e-11, "compression_ratio": 1.1787878787878787 }, { "start": 253.14, "end": 253.9, "text": "分析数据", "chunk": 0, "language": "zh", "avg_logprob": -0.09454160325982597, "no_speech_prob": 1.7845730101995194e-11, "compression_ratio": 1.1787878787878787 }, { "start": 253.9, "end": 255.88, "text": "然后再手动修改参数", "chunk": 0, "language": "zh", "avg_logprob": -0.09454160325982597, "no_speech_prob": 1.7845730101995194e-11, "compression_ratio": 1.1787878787878787 }, { "start": 255.88, "end": 256.6, "text": "跑下一次", "chunk": 0, "language": "zh", "avg_logprob": -0.08476443317055044, "no_speech_prob": 1.4740271503388414e-11, "compression_ratio": 1.2333333333333334 }, { "start": 256.6, "end": 258.3, "text": "在这套流程里", "chunk": 0, "language": "zh", "avg_logprob": -0.08476443317055044, "no_speech_prob": 1.4740271503388414e-11, "compression_ratio": 1.2333333333333334 }, { "start": 258.3, "end": 260.72, "text": "最瓶颈的变量不是算力", "chunk": 0, "language": "zh", "avg_logprob": -0.08476443317055044, "no_speech_prob": 1.4740271503388414e-11, "compression_ratio": 1.2333333333333334 }, { "start": 260.72, "end": 263.3, "text": "而是人类脑力的思考速度", "chunk": 0, "language": "zh", "avg_logprob": -0.08476443317055044, "no_speech_prob": 1.4740271503388414e-11, "compression_ratio": 1.2333333333333334 }, { "start": 263.3, "end": 265.6, "text": "与手动操作的物理实验", "chunk": 0, "language": "zh", "avg_logprob": -0.08476443317055044, "no_speech_prob": 1.4740271503388414e-11, "compression_ratio": 1.2333333333333334 }, { "start": 265.6, "end": 268.48, "text": "而Discovery Loop要构建一套", "chunk": 0, "language": "zh", "avg_logprob": -0.08476443317055044, "no_speech_prob": 1.4740271503388414e-11, "compression_ratio": 1.2333333333333334 }, { "start": 268.48, "end": 271.28, "text": "完全自主运行的AIA层架构", "chunk": 0, "language": "zh", "avg_logprob": -0.08476443317055044, "no_speech_prob": 1.4740271503388414e-11, "compression_ratio": 1.2333333333333334 }, { "start": 271.28, "end": 273.72, "text": "做到自动提出科研假设", "chunk": 0, "language": "zh", "avg_logprob": -0.08476443317055044, "no_speech_prob": 1.4740271503388414e-11, "compression_ratio": 1.2333333333333334 }, { "start": 273.72, "end": 275.62, "text": "自动编写实验代码", "chunk": 0, "language": "zh", "avg_logprob": -0.08476443317055044, "no_speech_prob": 1.4740271503388414e-11, "compression_ratio": 1.2333333333333334 }, { "start": 275.62, "end": 277.4, "text": "自动部署分布式计算", "chunk": 0, "language": "zh", "avg_logprob": -0.08476443317055044, "no_speech_prob": 1.4740271503388414e-11, "compression_ratio": 1.2333333333333334 }, { "start": 277.4, "end": 279.22, "text": "自动分析实验指标", "chunk": 0, "language": "zh", "avg_logprob": -0.08476443317055044, "no_speech_prob": 1.4740271503388414e-11, "compression_ratio": 1.2333333333333334 }, { "start": 279.22, "end": 281.16, "text": "最后自动递归迭代", "chunk": 0, "language": "zh", "avg_logprob": -0.08476443317055044, "no_speech_prob": 1.4740271503388414e-11, "compression_ratio": 1.2333333333333334 }, { "start": 281.16, "end": 283.38, "text": "你注意看Discovery Loop的", "chunk": 0, "language": "zh", "avg_logprob": -0.08476443317055044, "no_speech_prob": 1.4740271503388414e-11, "compression_ratio": 1.2333333333333334 }, { "start": 283.38, "end": 285.14, "text": "手机应用场景选择", "chunk": 0, "language": "zh", "avg_logprob": -0.08476443317055044, "no_speech_prob": 1.4740271503388414e-11, "compression_ratio": 1.2333333333333334 }, { "start": 285.14, "end": 288.04, "text": "他们没有第一天就跑去见物理实验室", "chunk": 0, "language": "zh", "avg_logprob": -0.0741603359486322, "no_speech_prob": 1.3214746137635291e-11, "compression_ratio": 1.2056962025316456 }, { "start": 288.04, "end": 290.22, "text": "而是选择吃自己的狗粮", "chunk": 0, "language": "zh", "avg_logprob": -0.0741603359486322, "no_speech_prob": 1.3214746137635291e-11, "compression_ratio": 1.2056962025316456 }, { "start": 290.22, "end": 294.32, "text": "把机器学习研发本身作为第一个实验战场", "chunk": 0, "language": "zh", "avg_logprob": -0.0741603359486322, "no_speech_prob": 1.3214746137635291e-11, "compression_ratio": 1.2056962025316456 }, { "start": 294.32, "end": 297.58, "text": "也就是说他们要用AI去自动设计", "chunk": 0, "language": "zh", "avg_logprob": -0.0741603359486322, "no_speech_prob": 1.3214746137635291e-11, "compression_ratio": 1.2056962025316456 }, { "start": 297.58, "end": 298.56, "text": "自动搜索", "chunk": 0, "language": "zh", "avg_logprob": -0.0741603359486322, "no_speech_prob": 1.3214746137635291e-11, "compression_ratio": 1.2056962025316456 }, { "start": 298.56, "end": 302.26, "text": "自动验证下一代超越Transformer的模型架构", "chunk": 0, "language": "zh", "avg_logprob": -0.0741603359486322, "no_speech_prob": 1.3214746137635291e-11, "compression_ratio": 1.2056962025316456 }, { "start": 302.26, "end": 304.14, "text": "很多朋友可能会问", "chunk": 0, "language": "zh", "avg_logprob": -0.0741603359486322, "no_speech_prob": 1.3214746137635291e-11, "compression_ratio": 1.2056962025316456 }, { "start": 304.14, "end": 306.84, "text": "用AI自动写代码做研究", "chunk": 0, "language": "zh", "avg_logprob": -0.0741603359486322, "no_speech_prob": 1.3214746137635291e-11, "compression_ratio": 1.2056962025316456 }, { "start": 306.84, "end": 310.14, "text": "之前日本Sakana AI的项目不是也做过吗", "chunk": 0, "language": "zh", "avg_logprob": -0.0741603359486322, "no_speech_prob": 1.3214746137635291e-11, "compression_ratio": 1.2056962025316456 }, { "start": 310.14, "end": 313.2, "text": "这正是我们接下来要拆解的关键", "chunk": 0, "language": "zh", "avg_logprob": -0.0741603359486322, "no_speech_prob": 1.3214746137635291e-11, "compression_ratio": 1.2056962025316456 }, { "start": 313.2, "end": 318.1, "text": "Satan的AI的项目本质上是一个应用层的agent流程封装", "chunk": 0, "language": "zh", "avg_logprob": -0.11626734052385602, "no_speech_prob": 1.5439439657316534e-11, "compression_ratio": 1.0978593272171253 }, { "start": 318.1, "end": 322.7, "text": "它是拿现成的GPT4O或者Cloud外包一层prompt", "chunk": 0, "language": "zh", "avg_logprob": -0.11626734052385602, "no_speech_prob": 1.5439439657316534e-11, "compression_ratio": 1.0978593272171253 }, { "start": 322.7, "end": 326.3, "text": "去模拟科学家写代码跑实验写论文", "chunk": 0, "language": "zh", "avg_logprob": -0.11626734052385602, "no_speech_prob": 1.5439439657316534e-11, "compression_ratio": 1.0978593272171253 }, { "start": 326.3, "end": 331.38, "text": "它的上限完全被底下调用的通用大模型给封死了", "chunk": 0, "language": "zh", "avg_logprob": -0.11626734052385602, "no_speech_prob": 1.5439439657316534e-11, "compression_ratio": 1.0978593272171253 }, { "start": 331.38, "end": 335.14, "text": "而Discovery Loop完全是另一个维度的存在", "chunk": 0, "language": "zh", "avg_logprob": -0.11626734052385602, "no_speech_prob": 1.5439439657316534e-11, "compression_ratio": 1.0978593272171253 }, { "start": 335.14, "end": 342.54, "text": "Jeff Dean和桑杰·格瓦马特是写出MapReduce Bigtable的超级系统架构师", "chunk": 0, "language": "zh", "avg_logprob": -0.11626734052385602, "no_speech_prob": 1.5439439657316534e-11, "compression_ratio": 1.0978593272171253 }, { "start": 342.54, "end": 345.3, "text": "李光莲是AutoML的鼻祖", "chunk": 0, "language": "zh", "avg_logprob": -0.0700369768364485, "no_speech_prob": 1.5092796804005992e-11, "compression_ratio": 1.277992277992278 }, { "start": 345.3, "end": 347.02, "text": "奥里尔尔维尼亚尔斯", "chunk": 0, "language": "zh", "avg_logprob": -0.0700369768364485, "no_speech_prob": 1.5092796804005992e-11, "compression_ratio": 1.277992277992278 }, { "start": 347.02, "end": 350.58, "text": "是顶尖的神经网络搜索与多模态专家", "chunk": 0, "language": "zh", "avg_logprob": -0.0700369768364485, "no_speech_prob": 1.5092796804005992e-11, "compression_ratio": 1.277992277992278 }, { "start": 350.58, "end": 353.96, "text": "他们要做的是从底层分布式计算", "chunk": 0, "language": "zh", "avg_logprob": -0.0700369768364485, "no_speech_prob": 1.5092796804005992e-11, "compression_ratio": 1.277992277992278 }, { "start": 353.96, "end": 355.52, "text": "硬件算力调度", "chunk": 0, "language": "zh", "avg_logprob": -0.0700369768364485, "no_speech_prob": 1.5092796804005992e-11, "compression_ratio": 1.277992277992278 }, { "start": 355.52, "end": 360.12, "text": "神经架构搜索到算法闭环的系统级重构", "chunk": 0, "language": "zh", "avg_logprob": -0.0700369768364485, "no_speech_prob": 1.5092796804005992e-11, "compression_ratio": 1.277992277992278 }, { "start": 360.12, "end": 362.18, "text": "他们不是在用AI写论文", "chunk": 0, "language": "zh", "avg_logprob": -0.0700369768364485, "no_speech_prob": 1.5092796804005992e-11, "compression_ratio": 1.277992277992278 }, { "start": 362.18, "end": 364.66, "text": "他们是在用超级计算机集群", "chunk": 0, "language": "zh", "avg_logprob": -0.0700369768364485, "no_speech_prob": 1.5092796804005992e-11, "compression_ratio": 1.277992277992278 }, { "start": 364.66, "end": 369.26, "text": "让AI去做数以万计的硬件级与架构级实验", "chunk": 0, "language": "zh", "avg_logprob": -0.0700369768364485, "no_speech_prob": 1.5092796804005992e-11, "compression_ratio": 1.277992277992278 }, { "start": 369.26, "end": 373.0, "text": "直接在算法底层寻找下一个Transformer", "chunk": 0, "language": "zh", "avg_logprob": -0.09177416439714103, "no_speech_prob": 1.179939132511798e-11, "compression_ratio": 1.1587837837837838 }, { "start": 373.0, "end": 375.26, "text": "那么回到我们开头的悬念", "chunk": 0, "language": "zh", "avg_logprob": -0.09177416439714103, "no_speech_prob": 1.179939132511798e-11, "compression_ratio": 1.1587837837837838 }, { "start": 375.26, "end": 379.56, "text": "就是谷歌为什么不把这个超级团队留在内部", "chunk": 0, "language": "zh", "avg_logprob": -0.09177416439714103, "no_speech_prob": 1.179939132511798e-11, "compression_ratio": 1.1587837837837838 }, { "start": 379.56, "end": 383.34, "text": "而是支持他们独立成立一家公共利益公司呢", "chunk": 0, "language": "zh", "avg_logprob": -0.09177416439714103, "no_speech_prob": 1.179939132511798e-11, "compression_ratio": 1.1587837837837838 }, { "start": 383.34, "end": 385.98, "text": "因为谷歌算透了一笔账", "chunk": 0, "language": "zh", "avg_logprob": -0.09177416439714103, "no_speech_prob": 1.179939132511798e-11, "compression_ratio": 1.1587837837837838 }, { "start": 385.98, "end": 389.72, "text": "像AI自主研发下一代AI这种前沿探索", "chunk": 0, "language": "zh", "avg_logprob": -0.09177416439714103, "no_speech_prob": 1.179939132511798e-11, "compression_ratio": 1.1587837837837838 }, { "start": 389.72, "end": 393.64, "text": "风险极高烧钱极大失败率也极高", "chunk": 0, "language": "zh", "avg_logprob": -0.09177416439714103, "no_speech_prob": 1.179939132511798e-11, "compression_ratio": 1.1587837837837838 }, { "start": 393.64, "end": 395.72, "text": "如果放在谷歌Digmy内部", "chunk": 0, "language": "zh", "avg_logprob": -0.09177416439714103, "no_speech_prob": 1.179939132511798e-11, "compression_ratio": 1.1587837837837838 }, { "start": 395.72, "end": 400.44, "text": "研究员们每天要面对华尔街的财报审查和季度KPI", "chunk": 0, "language": "zh", "avg_logprob": -0.07157229795688536, "no_speech_prob": 1.1848798851799014e-11, "compression_ratio": 1.1305841924398625 }, { "start": 400.44, "end": 403.7, "text": "就会被无限牵扯精力去搞产品适配", "chunk": 0, "language": "zh", "avg_logprob": -0.07157229795688536, "no_speech_prob": 1.1848798851799014e-11, "compression_ratio": 1.1305841924398625 }, { "start": 403.7, "end": 406.26, "text": "通过把JFDN团队外部化", "chunk": 0, "language": "zh", "avg_logprob": -0.07157229795688536, "no_speech_prob": 1.1848798851799014e-11, "compression_ratio": 1.1305841924398625 }, { "start": 406.26, "end": 408.74, "text": "谷歌既让这一帮顶尖天才", "chunk": 0, "language": "zh", "avg_logprob": -0.07157229795688536, "no_speech_prob": 1.1848798851799014e-11, "compression_ratio": 1.1305841924398625 }, { "start": 408.74, "end": 412.28, "text": "摆脱了大公司病和商业汇报的枷锁", "chunk": 0, "language": "zh", "avg_logprob": -0.07157229795688536, "no_speech_prob": 1.1848798851799014e-11, "compression_ratio": 1.1305841924398625 }, { "start": 412.28, "end": 417.14, "text": "又能通过初始投资方和独家GoCloud算力的身份", "chunk": 0, "language": "zh", "avg_logprob": -0.07157229795688536, "no_speech_prob": 1.1848798851799014e-11, "compression_ratio": 1.1305841924398625 }, { "start": 417.14, "end": 421.82, "text": "牢牢锁定未来潜在的技术回流与云服务收益", "chunk": 0, "language": "zh", "avg_logprob": -0.07157229795688536, "no_speech_prob": 1.1848798851799014e-11, "compression_ratio": 1.1305841924398625 }, { "start": 421.82, "end": 426.58, "text": "这一招叫做前沿高风险战略的外部结偶", "chunk": 0, "language": "zh", "avg_logprob": -0.14204300774468315, "no_speech_prob": 1.8405497612117294e-11, "compression_ratio": 1.1467065868263473 }, { "start": 426.58, "end": 431.44, "text": "如果说Jeff Dean的离职是把远期高风险探索拆分了出去", "chunk": 0, "language": "zh", "avg_logprob": -0.14204300774468315, "no_speech_prob": 1.8405497612117294e-11, "compression_ratio": 1.1467065868263473 }, { "start": 431.44, "end": 436.74, "text": "那么接下来发生的事情才是对Jamlight产品生态影响最深远", "chunk": 0, "language": "zh", "avg_logprob": -0.14204300774468315, "no_speech_prob": 1.8405497612117294e-11, "compression_ratio": 1.1467065868263473 }, { "start": 436.74, "end": 438.24, "text": "也是最暴力的一环", "chunk": 0, "language": "zh", "avg_logprob": -0.14204300774468315, "no_speech_prob": 1.8405497612117294e-11, "compression_ratio": 1.1467065868263473 }, { "start": 438.24, "end": 442.7, "text": "接掌Google D.My的全局率颖的科雷卡伍克·乔武", "chunk": 0, "language": "zh", "avg_logprob": -0.14204300774468315, "no_speech_prob": 1.8405497612117294e-11, "compression_ratio": 1.1467065868263473 }, { "start": 442.7, "end": 447.18, "text": "是一个与Demis Hossabies风格截然不同的工程技术派", "chunk": 0, "language": "zh", "avg_logprob": -0.14204300774468315, "no_speech_prob": 1.8405497612117294e-11, "compression_ratio": 1.1467065868263473 }, { "start": 447.18, "end": 451.48, "text": "Demis是国际象棋神童和神经科学家", "chunk": 0, "language": "zh", "avg_logprob": -0.14204300774468315, "no_speech_prob": 1.8405497612117294e-11, "compression_ratio": 1.1467065868263473 }, { "start": 451.48, "end": 453.4, "text": "它充满着理想主义", "chunk": 0, "language": "zh", "avg_logprob": -0.1270897402400859, "no_speech_prob": 1.5438495967745602e-11, "compression_ratio": 1.1414141414141414 }, { "start": 453.4, "end": 456.18, "text": "而柯雷是把深度强化学习", "chunk": 0, "language": "zh", "avg_logprob": -0.1270897402400859, "no_speech_prob": 1.5438495967745602e-11, "compression_ratio": 1.1414141414141414 }, { "start": 456.18, "end": 458.66, "text": "真正推向AlphaGo破局", "chunk": 0, "language": "zh", "avg_logprob": -0.1270897402400859, "no_speech_prob": 1.5438495967745602e-11, "compression_ratio": 1.1414141414141414 }, { "start": 458.66, "end": 460.62, "text": "把WaveNet徐徐语音", "chunk": 0, "language": "zh", "avg_logprob": -0.1270897402400859, "no_speech_prob": 1.5438495967745602e-11, "compression_ratio": 1.1414141414141414 }, { "start": 460.62, "end": 465.62, "text": "真正变成极异手机里语音引擎的超级工程操盘手", "chunk": 0, "language": "zh", "avg_logprob": -0.1270897402400859, "no_speech_prob": 1.5438495967745602e-11, "compression_ratio": 1.1414141414141414 }, { "start": 465.62, "end": 467.76, "text": "在柯雷掌舵之后", "chunk": 0, "language": "zh", "avg_logprob": -0.1270897402400859, "no_speech_prob": 1.5438495967745602e-11, "compression_ratio": 1.1414141414141414 }, { "start": 467.76, "end": 470.64, "text": "Google G-Mind释放出了一个极其明确的信号", "chunk": 0, "language": "zh", "avg_logprob": -0.1270897402400859, "no_speech_prob": 1.5438495967745602e-11, "compression_ratio": 1.1414141414141414 }, { "start": 470.64, "end": 473.06, "text": "就是全面放弃学术包袱", "chunk": 0, "language": "zh", "avg_logprob": -0.1270897402400859, "no_speech_prob": 1.5438495967745602e-11, "compression_ratio": 1.1414141414141414 }, { "start": 473.06, "end": 477.24, "text": "转向极致工程演进与大规模商业化暴击", "chunk": 0, "language": "zh", "avg_logprob": -0.1270897402400859, "no_speech_prob": 1.5438495967745602e-11, "compression_ratio": 1.1414141414141414 }, { "start": 477.24, "end": 478.68, "text": "在他的主持下", "chunk": 0, "language": "zh", "avg_logprob": -0.1270897402400859, "no_speech_prob": 1.5438495967745602e-11, "compression_ratio": 1.1414141414141414 }, { "start": 478.68, "end": 479.96, "text": "Gemlet的底层架构", "chunk": 0, "language": "zh", "avg_logprob": -0.28152309145246235, "no_speech_prob": 7.54585803564023e-12, "compression_ratio": 0.7 }, { "start": 480.0, "end": 482.84, "text": "Gamlight的底层架构正在发生三重极其硬核的眼镜", "chunk": 1, "language": "zh", "avg_logprob": -0.10029149824573148, "no_speech_prob": 2.211964812703915e-11, "compression_ratio": 1.2019230769230769 }, { "start": 482.84, "end": 484.76, "text": "首先最根本的变化是", "chunk": 1, "language": "zh", "avg_logprob": -0.10029149824573148, "no_speech_prob": 2.211964812703915e-11, "compression_ratio": 1.2019230769230769 }, { "start": 484.76, "end": 488.84, "text": "彻底打破对预训练参数崇拜的单位幻想", "chunk": 1, "language": "zh", "avg_logprob": -0.10029149824573148, "no_speech_prob": 2.211964812703915e-11, "compression_ratio": 1.2019230769230769 }, { "start": 488.84, "end": 493.72, "text": "全面转向后训练与推理期计算的深度融合", "chunk": 1, "language": "zh", "avg_logprob": -0.10029149824573148, "no_speech_prob": 2.211964812703915e-11, "compression_ratio": 1.2019230769230769 }, { "start": 493.72, "end": 497.8, "text": "在过去两年行业盲目堆叠预训练参数", "chunk": 1, "language": "zh", "avg_logprob": -0.0894854135353472, "no_speech_prob": 1.3060411259424587e-11, "compression_ratio": 1.1476510067114094 }, { "start": 497.8, "end": 501.72, "text": "但预训练越往后编辑效应递减越严重", "chunk": 1, "language": "zh", "avg_logprob": -0.0894854135353472, "no_speech_prob": 1.3060411259424587e-11, "compression_ratio": 1.1476510067114094 }, { "start": 501.72, "end": 504.1, "text": "算力成本高到令人发指", "chunk": 1, "language": "zh", "avg_logprob": -0.0894854135353472, "no_speech_prob": 1.3060411259424587e-11, "compression_ratio": 1.1476510067114094 }, { "start": 504.1, "end": 507.9, "text": "柯雷结合他在DQN强化学习领域的积累", "chunk": 1, "language": "zh", "avg_logprob": -0.0894854135353472, "no_speech_prob": 1.3060411259424587e-11, "compression_ratio": 1.1476510067114094 }, { "start": 507.9, "end": 511.28, "text": "给Gemla引入了过程奖励模型PRM", "chunk": 1, "language": "zh", "avg_logprob": -0.0894854135353472, "no_speech_prob": 1.3060411259424587e-11, "compression_ratio": 1.1476510067114094 }, { "start": 511.28, "end": 514.16, "text": "与蒙特卡洛书搜索MCTS", "chunk": 1, "language": "zh", "avg_logprob": -0.0894854135353472, "no_speech_prob": 1.3060411259424587e-11, "compression_ratio": 1.1476510067114094 }, { "start": 514.16, "end": 517.08, "text": "打个比方传统的语言模型", "chunk": 1, "language": "zh", "avg_logprob": -0.0894854135353472, "no_speech_prob": 1.3060411259424587e-11, "compression_ratio": 1.1476510067114094 }, { "start": 517.08, "end": 519.82, "text": "就像在考场上没有任何草稿纸", "chunk": 1, "language": "zh", "avg_logprob": -0.0894854135353472, "no_speech_prob": 1.3060411259424587e-11, "compression_ratio": 1.1476510067114094 }, { "start": 519.82, "end": 521.78, "text": "靠直觉备答案的考生", "chunk": 1, "language": "zh", "avg_logprob": -0.0894854135353472, "no_speech_prob": 1.3060411259424587e-11, "compression_ratio": 1.1476510067114094 }, { "start": 521.78, "end": 524.18, "text": "遇到简单的题那是没问题的", "chunk": 1, "language": "zh", "avg_logprob": -0.11376734477717702, "no_speech_prob": 1.2319105804903252e-11, "compression_ratio": 1.1724137931034482 }, { "start": 524.18, "end": 527.52, "text": "但遇到复杂逻辑推导或者长流程编程", "chunk": 1, "language": "zh", "avg_logprob": -0.11376734477717702, "no_speech_prob": 1.2319105804903252e-11, "compression_ratio": 1.1724137931034482 }, { "start": 527.52, "end": 530.52, "text": "中间要是错一步后面就全乱了", "chunk": 1, "language": "zh", "avg_logprob": -0.11376734477717702, "no_speech_prob": 1.2319105804903252e-11, "compression_ratio": 1.1724137931034482 }, { "start": 530.52, "end": 533.46, "text": "而PRM与MCTS的融合", "chunk": 1, "language": "zh", "avg_logprob": -0.11376734477717702, "no_speech_prob": 1.2319105804903252e-11, "compression_ratio": 1.1724137931034482 }, { "start": 533.46, "end": 538.0, "text": "相当于给Gemla配了一张无线扩展的虚拟草稿纸", "chunk": 1, "language": "zh", "avg_logprob": -0.11376734477717702, "no_speech_prob": 1.2319105804903252e-11, "compression_ratio": 1.1724137931034482 }, { "start": 538.0, "end": 539.92, "text": "和一位微观监考官", "chunk": 1, "language": "zh", "avg_logprob": -0.11376734477717702, "no_speech_prob": 1.2319105804903252e-11, "compression_ratio": 1.1724137931034482 }, { "start": 539.92, "end": 541.52, "text": "在推理的每一步", "chunk": 1, "language": "zh", "avg_logprob": -0.11376734477717702, "no_speech_prob": 1.2319105804903252e-11, "compression_ratio": 1.1724137931034482 }, { "start": 541.52, "end": 546.08, "text": "PRM都会对它的逻辑合理性进行毫秒级评分", "chunk": 1, "language": "zh", "avg_logprob": -0.11376734477717702, "no_speech_prob": 1.2319105804903252e-11, "compression_ratio": 1.1724137931034482 }, { "start": 546.08, "end": 550.0, "text": "同时MCTS数搜索会在后台预判", "chunk": 1, "language": "zh", "avg_logprob": -0.11376734477717702, "no_speech_prob": 1.2319105804903252e-11, "compression_ratio": 1.1724137931034482 }, { "start": 550.0, "end": 551.7, "text": "未来几十步的走向", "chunk": 1, "language": "zh", "avg_logprob": -0.1314085006713867, "no_speech_prob": 1.2610554094016901e-11, "compression_ratio": 1.2007042253521127 }, { "start": 551.7, "end": 554.72, "text": "如果发现某个分支走向死胡同", "chunk": 1, "language": "zh", "avg_logprob": -0.1314085006713867, "no_speech_prob": 1.2610554094016901e-11, "compression_ratio": 1.2007042253521127 }, { "start": 554.72, "end": 557.44, "text": "模型会在推理阶段自动撤回", "chunk": 1, "language": "zh", "avg_logprob": -0.1314085006713867, "no_speech_prob": 1.2610554094016901e-11, "compression_ratio": 1.2007042253521127 }, { "start": 557.44, "end": 559.14, "text": "并重新搜索最右键", "chunk": 1, "language": "zh", "avg_logprob": -0.1314085006713867, "no_speech_prob": 1.2610554094016901e-11, "compression_ratio": 1.2007042253521127 }, { "start": 559.14, "end": 560.24, "text": "这意味着", "chunk": 1, "language": "zh", "avg_logprob": -0.1314085006713867, "no_speech_prob": 1.2610554094016901e-11, "compression_ratio": 1.2007042253521127 }, { "start": 560.24, "end": 564.18, "text": "在不用无限膨胀预讯链参数的前提下", "chunk": 1, "language": "zh", "avg_logprob": -0.1314085006713867, "no_speech_prob": 1.2610554094016901e-11, "compression_ratio": 1.2007042253521127 }, { "start": 564.18, "end": 566.24, "text": "GEMLINE处理高难度编程", "chunk": 1, "language": "zh", "avg_logprob": -0.1314085006713867, "no_speech_prob": 1.2610554094016901e-11, "compression_ratio": 1.2007042253521127 }, { "start": 566.24, "end": 569.62, "text": "复杂数学推导和A乘的规划的成功率", "chunk": 1, "language": "zh", "avg_logprob": -0.1314085006713867, "no_speech_prob": 1.2610554094016901e-11, "compression_ratio": 1.2007042253521127 }, { "start": 569.62, "end": 571.42, "text": "称指数级提升", "chunk": 1, "language": "zh", "avg_logprob": -0.1314085006713867, "no_speech_prob": 1.2610554094016901e-11, "compression_ratio": 1.2007042253521127 }, { "start": 571.42, "end": 573.94, "text": "紧接着第二种转变在于", "chunk": 1, "language": "zh", "avg_logprob": -0.1314085006713867, "no_speech_prob": 1.2610554094016901e-11, "compression_ratio": 1.2007042253521127 }, { "start": 573.94, "end": 576.54, "text": "系统级推理效率的极限拉链", "chunk": 1, "language": "zh", "avg_logprob": -0.1314085006713867, "no_speech_prob": 1.2610554094016901e-11, "compression_ratio": 1.2007042253521127 }, { "start": 576.54, "end": 580.66, "text": "那就是构建Gemini Spark端云协同分层架构", "chunk": 1, "language": "zh", "avg_logprob": -0.08911573345010931, "no_speech_prob": 1.3305702026289445e-11, "compression_ratio": 1.2328358208955223 }, { "start": 580.66, "end": 584.26, "text": "这里稍微提一下我个人对Spark的使用体验", "chunk": 1, "language": "zh", "avg_logprob": -0.08911573345010931, "no_speech_prob": 1.3305702026289445e-11, "compression_ratio": 1.2328358208955223 }, { "start": 584.26, "end": 586.06, "text": "虽然它还是Beta版本", "chunk": 1, "language": "zh", "avg_logprob": -0.08911573345010931, "no_speech_prob": 1.3305702026289445e-11, "compression_ratio": 1.2328358208955223 }, { "start": 586.06, "end": 587.26, "text": "但先说结论", "chunk": 1, "language": "zh", "avg_logprob": -0.08911573345010931, "no_speech_prob": 1.3305702026289445e-11, "compression_ratio": 1.2328358208955223 }, { "start": 587.26, "end": 589.08, "text": "我觉得它超级好用", "chunk": 1, "language": "zh", "avg_logprob": -0.08911573345010931, "no_speech_prob": 1.3305702026289445e-11, "compression_ratio": 1.2328358208955223 }, { "start": 589.08, "end": 590.74, "text": "我会单独出一期视频", "chunk": 1, "language": "zh", "avg_logprob": -0.08911573345010931, "no_speech_prob": 1.3305702026289445e-11, "compression_ratio": 1.2328358208955223 }, { "start": 590.74, "end": 594.1, "text": "聊一下Gemini Spark和Hermes Agent的配合使用", "chunk": 1, "language": "zh", "avg_logprob": -0.08911573345010931, "no_speech_prob": 1.3305702026289445e-11, "compression_ratio": 1.2328358208955223 }, { "start": 594.1, "end": 596.78, "text": "作为Google订阅的长期体验者", "chunk": 1, "language": "zh", "avg_logprob": -0.08911573345010931, "no_speech_prob": 1.3305702026289445e-11, "compression_ratio": 1.2328358208955223 }, { "start": 596.78, "end": 600.52, "text": "我明显感觉最近Gemini的使用限额提高了", "chunk": 1, "language": "zh", "avg_logprob": -0.08911573345010931, "no_speech_prob": 1.3305702026289445e-11, "compression_ratio": 1.2328358208955223 }, { "start": 600.52, "end": 603.88, "text": "他们虽然没有发布像Fable 5这类的模型", "chunk": 1, "language": "zh", "avg_logprob": -0.08911573345010931, "no_speech_prob": 1.3305702026289445e-11, "compression_ratio": 1.2328358208955223 }, { "start": 603.88, "end": 606.4, "text": "但是无论是Gemini 3.6 Flash", "chunk": 1, "language": "zh", "avg_logprob": -0.08911573345010931, "no_speech_prob": 1.3305702026289445e-11, "compression_ratio": 1.2328358208955223 }, { "start": 606.4, "end": 609.44, "text": "还是Gemlet 3.1 Pro都变得更聪明了", "chunk": 1, "language": "zh", "avg_logprob": -0.0855890873500279, "no_speech_prob": 2.0221558802435702e-11, "compression_ratio": 1.1711711711711712 }, { "start": 609.44, "end": 611.74, "text": "好我们继续这个第二重转变", "chunk": 1, "language": "zh", "avg_logprob": -0.0855890873500279, "no_speech_prob": 2.0221558802435702e-11, "compression_ratio": 1.1711711711711712 }, { "start": 611.74, "end": 614.26, "text": "在商业高并发场景中", "chunk": 1, "language": "zh", "avg_logprob": -0.0855890873500279, "no_speech_prob": 2.0221558802435702e-11, "compression_ratio": 1.1711711711711712 }, { "start": 614.26, "end": 616.42, "text": "用户问一句今天天气怎么样", "chunk": 1, "language": "zh", "avg_logprob": -0.0855890873500279, "no_speech_prob": 2.0221558802435702e-11, "compression_ratio": 1.1711711711711712 }, { "start": 616.42, "end": 620.68, "text": "如果都要调用一次千亿参数的单体大模型", "chunk": 1, "language": "zh", "avg_logprob": -0.0855890873500279, "no_speech_prob": 2.0221558802435702e-11, "compression_ratio": 1.1711711711711712 }, { "start": 620.68, "end": 623.82, "text": "显存和电力开销能把大厂烧穿", "chunk": 1, "language": "zh", "avg_logprob": -0.0855890873500279, "no_speech_prob": 2.0221558802435702e-11, "compression_ratio": 1.1711711711711712 }, { "start": 623.82, "end": 626.22, "text": "克雷推动的Gemlet Spark架构", "chunk": 1, "language": "zh", "avg_logprob": -0.0855890873500279, "no_speech_prob": 2.0221558802435702e-11, "compression_ratio": 1.1711711711711712 }, { "start": 626.22, "end": 630.9, "text": "把整套AI系统做成了像特种部队一样的分层调度", "chunk": 1, "language": "zh", "avg_logprob": -0.0855890873500279, "no_speech_prob": 2.0221558802435702e-11, "compression_ratio": 1.1711711711711712 }, { "start": 630.9, "end": 635.84, "text": "前端部署及轻量及低延迟的Gemlet Spark代理模型", "chunk": 1, "language": "zh", "avg_logprob": -0.0855890873500279, "no_speech_prob": 2.0221558802435702e-11, "compression_ratio": 1.1711711711711712 }, { "start": 635.84, "end": 637.8, "text": "负责高频意图识别", "chunk": 1, "language": "zh", "avg_logprob": -0.11474797825613421, "no_speech_prob": 1.3550213034951852e-11, "compression_ratio": 1.1512027491408934 }, { "start": 637.8, "end": 640.24, "text": "上下门过滤和轻量任务", "chunk": 1, "language": "zh", "avg_logprob": -0.11474797825613421, "no_speech_prob": 1.3550213034951852e-11, "compression_ratio": 1.1512027491408934 }, { "start": 640.24, "end": 641.82, "text": "只有当前线判定", "chunk": 1, "language": "zh", "avg_logprob": -0.11474797825613421, "no_speech_prob": 1.3550213034951852e-11, "compression_ratio": 1.1512027491408934 }, { "start": 641.82, "end": 644.18, "text": "涉及复杂长链条逻辑的时候", "chunk": 1, "language": "zh", "avg_logprob": -0.11474797825613421, "no_speech_prob": 1.3550213034951852e-11, "compression_ratio": 1.1512027491408934 }, { "start": 644.18, "end": 647.56, "text": "才会无缝调动云端期间大模型", "chunk": 1, "language": "zh", "avg_logprob": -0.11474797825613421, "no_speech_prob": 1.3550213034951852e-11, "compression_ratio": 1.1512027491408934 }, { "start": 647.56, "end": 648.96, "text": "这种分层架构", "chunk": 1, "language": "zh", "avg_logprob": -0.11474797825613421, "no_speech_prob": 1.3550213034951852e-11, "compression_ratio": 1.1512027491408934 }, { "start": 648.96, "end": 651.86, "text": "直接把单位Token成本打了下来", "chunk": 1, "language": "zh", "avg_logprob": -0.11474797825613421, "no_speech_prob": 1.3550213034951852e-11, "compression_ratio": 1.1512027491408934 }, { "start": 651.86, "end": 655.34, "text": "这才是支撑数亿并发的商业硬实力", "chunk": 1, "language": "zh", "avg_logprob": -0.11474797825613421, "no_speech_prob": 1.3550213034951852e-11, "compression_ratio": 1.1512027491408934 }, { "start": 655.34, "end": 656.9, "text": "而第三种变化", "chunk": 1, "language": "zh", "avg_logprob": -0.11474797825613421, "no_speech_prob": 1.3550213034951852e-11, "compression_ratio": 1.1512027491408934 }, { "start": 656.9, "end": 659.46, "text": "则是柯雷将他的看家本领", "chunk": 1, "language": "zh", "avg_logprob": -0.11474797825613421, "no_speech_prob": 1.3550213034951852e-11, "compression_ratio": 1.1512027491408934 }, { "start": 659.46, "end": 662.5, "text": "VIVNET原生音频与流逝生成", "chunk": 1, "language": "zh", "avg_logprob": -0.11474797825613421, "no_speech_prob": 1.3550213034951852e-11, "compression_ratio": 1.1512027491408934 }, { "start": 662.5, "end": 666.24, "text": "深度融合到了詹姆賴的全模态底座中", "chunk": 1, "language": "zh", "avg_logprob": -0.10299529824205624, "no_speech_prob": 1.4336070525144962e-11, "compression_ratio": 1.28719723183391 }, { "start": 666.24, "end": 668.02, "text": "以往的多模态AI", "chunk": 1, "language": "zh", "avg_logprob": -0.10299529824205624, "no_speech_prob": 1.4336070525144962e-11, "compression_ratio": 1.28719723183391 }, { "start": 668.02, "end": 671.42, "text": "往往是文本模型外挂语音识别", "chunk": 1, "language": "zh", "avg_logprob": -0.10299529824205624, "no_speech_prob": 1.4336070525144962e-11, "compression_ratio": 1.28719723183391 }, { "start": 671.42, "end": 673.96, "text": "与语音合成模块的拼接车", "chunk": 1, "language": "zh", "avg_logprob": -0.10299529824205624, "no_speech_prob": 1.4336070525144962e-11, "compression_ratio": 1.28719723183391 }, { "start": 673.96, "end": 676.12, "text": "延迟高语调生硬", "chunk": 1, "language": "zh", "avg_logprob": -0.10299529824205624, "no_speech_prob": 1.4336070525144962e-11, "compression_ratio": 1.28719723183391 }, { "start": 676.12, "end": 678.12, "text": "而柯雷主导下的詹姆賴", "chunk": 1, "language": "zh", "avg_logprob": -0.10299529824205624, "no_speech_prob": 1.4336070525144962e-11, "compression_ratio": 1.28719723183391 }, { "start": 678.12, "end": 681.44, "text": "实现了真正的原生全模态流逝交互", "chunk": 1, "language": "zh", "avg_logprob": -0.10299529824205624, "no_speech_prob": 1.4336070525144962e-11, "compression_ratio": 1.28719723183391 }, { "start": 681.44, "end": 683.28, "text": "不仅能实现毫秒级", "chunk": 1, "language": "zh", "avg_logprob": -0.10299529824205624, "no_speech_prob": 1.4336070525144962e-11, "compression_ratio": 1.28719723183391 }, { "start": 683.28, "end": 685.84, "text": "实时音视频双向对打", "chunk": 1, "language": "zh", "avg_logprob": -0.10299529824205624, "no_speech_prob": 1.4336070525144962e-11, "compression_ratio": 1.28719723183391 }, { "start": 685.84, "end": 687.64, "text": "还能在听懂说话的同时", "chunk": 1, "language": "zh", "avg_logprob": -0.10299529824205624, "no_speech_prob": 1.4336070525144962e-11, "compression_ratio": 1.28719723183391 }, { "start": 687.64, "end": 689.18, "text": "自主执行代码", "chunk": 1, "language": "zh", "avg_logprob": -0.10299529824205624, "no_speech_prob": 1.4336070525144962e-11, "compression_ratio": 1.28719723183391 }, { "start": 689.18, "end": 691.58, "text": "甚至在手机屏幕上动态生成", "chunk": 1, "language": "zh", "avg_logprob": -0.10299529824205624, "no_speech_prob": 1.4336070525144962e-11, "compression_ratio": 1.28719723183391 }, { "start": 691.58, "end": 693.6, "text": "UI界面与交互空间", "chunk": 1, "language": "zh", "avg_logprob": -0.09534252324753259, "no_speech_prob": 1.6312591902267748e-11, "compression_ratio": 1.2108626198083068 }, { "start": 693.6, "end": 695.6, "text": "这已经不是在聊天了", "chunk": 1, "language": "zh", "avg_logprob": -0.09534252324753259, "no_speech_prob": 1.6312591902267748e-11, "compression_ratio": 1.2108626198083068 }, { "start": 695.6, "end": 700.58, "text": "这是直接把自然语言变成了实时软件生成的界面", "chunk": 1, "language": "zh", "avg_logprob": -0.09534252324753259, "no_speech_prob": 1.6312591902267748e-11, "compression_ratio": 1.2108626198083068 }, { "start": 700.58, "end": 702.8, "text": "那底层架构演进的再漂亮", "chunk": 1, "language": "zh", "avg_logprob": -0.09534252324753259, "no_speech_prob": 1.6312591902267748e-11, "compression_ratio": 1.2108626198083068 }, { "start": 702.8, "end": 705.0, "text": "如果不能转化为商业流水", "chunk": 1, "language": "zh", "avg_logprob": -0.09534252324753259, "no_speech_prob": 1.6312591902267748e-11, "compression_ratio": 1.2108626198083068 }, { "start": 705.0, "end": 707.28, "text": "在今天也是一张废纸", "chunk": 1, "language": "zh", "avg_logprob": -0.09534252324753259, "no_speech_prob": 1.6312591902267748e-11, "compression_ratio": 1.2108626198083068 }, { "start": 707.28, "end": 710.9, "text": "科雷·卡伍克·乔鲁这次最让人惊悚的调整", "chunk": 1, "language": "zh", "avg_logprob": -0.09534252324753259, "no_speech_prob": 1.6312591902267748e-11, "compression_ratio": 1.2108626198083068 }, { "start": 710.9, "end": 712.68, "text": "在于他不仅管科研", "chunk": 1, "language": "zh", "avg_logprob": -0.09534252324753259, "no_speech_prob": 1.6312591902267748e-11, "compression_ratio": 1.2108626198083068 }, { "start": 712.68, "end": 717.6, "text": "还同时兼任了Gemline产品与开发者生态的最高指挥官", "chunk": 1, "language": "zh", "avg_logprob": -0.09534252324753259, "no_speech_prob": 1.6312591902267748e-11, "compression_ratio": 1.2108626198083068 }, { "start": 717.6, "end": 719.94, "text": "在过去的大厂体制里", "chunk": 1, "language": "zh", "avg_logprob": -0.09534252324753259, "no_speech_prob": 1.6312591902267748e-11, "compression_ratio": 1.2108626198083068 }, { "start": 719.94, "end": 724.36, "text": "科研实验室和商品产品部门是完全脱节的", "chunk": 1, "language": "zh", "avg_logprob": -0.09188194708390669, "no_speech_prob": 1.3213818927937382e-11, "compression_ratio": 1.1729559748427674 }, { "start": 724.36, "end": 725.58, "text": "实验室发了论文", "chunk": 1, "language": "zh", "avg_logprob": -0.09188194708390669, "no_speech_prob": 1.3213818927937382e-11, "compression_ratio": 1.1729559748427674 }, { "start": 725.58, "end": 728.78, "text": "产品部门要花大半年去沟通打包", "chunk": 1, "language": "zh", "avg_logprob": -0.09188194708390669, "no_speech_prob": 1.3213818927937382e-11, "compression_ratio": 1.1729559748427674 }, { "start": 728.78, "end": 730.74, "text": "而柯雷一把抓之后", "chunk": 1, "language": "zh", "avg_logprob": -0.09188194708390669, "no_speech_prob": 1.3213818927937382e-11, "compression_ratio": 1.1729559748427674 }, { "start": 730.74, "end": 733.82, "text": "直接祭出了商业落地的三大杀招", "chunk": 1, "language": "zh", "avg_logprob": -0.09188194708390669, "no_speech_prob": 1.3213818927937382e-11, "compression_ratio": 1.1729559748427674 }, { "start": 733.82, "end": 735.82, "text": "在终端入口上", "chunk": 1, "language": "zh", "avg_logprob": -0.09188194708390669, "no_speech_prob": 1.3213818927937382e-11, "compression_ratio": 1.1729559748427674 }, { "start": 735.82, "end": 739.0, "text": "他们开始加速清洗传统的语音助手", "chunk": 1, "language": "zh", "avg_logprob": -0.09188194708390669, "no_speech_prob": 1.3213818927937382e-11, "compression_ratio": 1.1729559748427674 }, { "start": 739.0, "end": 742.44, "text": "全面用GM来替换掉Android和VLOS上", "chunk": 1, "language": "zh", "avg_logprob": -0.09188194708390669, "no_speech_prob": 1.3213818927937382e-11, "compression_ratio": 1.1729559748427674 }, { "start": 742.44, "end": 744.56, "text": "那个指挥社闹钟的assistant", "chunk": 1, "language": "zh", "avg_logprob": -0.09188194708390669, "no_speech_prob": 1.3213818927937382e-11, "compression_ratio": 1.1729559748427674 }, { "start": 744.56, "end": 748.86, "text": "赋予他操作系统级的最高上下文读取权限", "chunk": 1, "language": "zh", "avg_logprob": -0.09188194708390669, "no_speech_prob": 1.3213818927937382e-11, "compression_ratio": 1.1729559748427674 }, { "start": 748.86, "end": 751.16, "text": "与跨应用API调用能力", "chunk": 1, "language": "zh", "avg_logprob": -0.09549988640679254, "no_speech_prob": 1.1937079796853212e-11, "compression_ratio": 1.1533101045296168 }, { "start": 751.16, "end": 755.04, "text": "直接把它打造成操作系统的中央控制大脑", "chunk": 1, "language": "zh", "avg_logprob": -0.09549988640679254, "no_speech_prob": 1.1937079796853212e-11, "compression_ratio": 1.1533101045296168 }, { "start": 755.04, "end": 758.52, "text": "比如在ValOS手表或者车载系统里", "chunk": 1, "language": "zh", "avg_logprob": -0.09549988640679254, "no_speech_prob": 1.1937079796853212e-11, "compression_ratio": 1.1533101045296168 }, { "start": 758.52, "end": 762.34, "text": "你只需要说帮我安排今晚与张总的商务晚餐", "chunk": 1, "language": "zh", "avg_logprob": -0.09549988640679254, "no_speech_prob": 1.1937079796853212e-11, "compression_ratio": 1.1533101045296168 }, { "start": 762.34, "end": 766.04, "text": "并准备一份他公司最新财报的背景简报", "chunk": 1, "language": "zh", "avg_logprob": -0.09549988640679254, "no_speech_prob": 1.1937079796853212e-11, "compression_ratio": 1.1533101045296168 }, { "start": 766.04, "end": 770.16, "text": "Gamlight就会自动调取邮件和地图订餐厅", "chunk": 1, "language": "zh", "avg_logprob": -0.09549988640679254, "no_speech_prob": 1.1937079796853212e-11, "compression_ratio": 1.1533101045296168 }, { "start": 770.16, "end": 774.44, "text": "同时在云端提取财报生成简报推送给你", "chunk": 1, "language": "zh", "avg_logprob": -0.09549988640679254, "no_speech_prob": 1.1937079796853212e-11, "compression_ratio": 1.1533101045296168 }, { "start": 774.44, "end": 775.58, "text": "与此同时", "chunk": 1, "language": "zh", "avg_logprob": -0.10155981132783086, "no_speech_prob": 1.2015840579471249e-11, "compression_ratio": 1.1815068493150684 }, { "start": 775.58, "end": 778.3, "text": "他们还大幅压缩了科研成果", "chunk": 1, "language": "zh", "avg_logprob": -0.10155981132783086, "no_speech_prob": 1.2015840579471249e-11, "compression_ratio": 1.1815068493150684 }, { "start": 778.3, "end": 780.44, "text": "转化到商业API的时间", "chunk": 1, "language": "zh", "avg_logprob": -0.10155981132783086, "no_speech_prob": 1.2015840579471249e-11, "compression_ratio": 1.1815068493150684 }, { "start": 780.44, "end": 784.44, "text": "在谷歌云Vertex AI和开发者API上", "chunk": 1, "language": "zh", "avg_logprob": -0.10155981132783086, "no_speech_prob": 1.2015840579471249e-11, "compression_ratio": 1.1815068493150684 }, { "start": 784.44, "end": 787.14, "text": "前沿团队在Achern代码生成", "chunk": 1, "language": "zh", "avg_logprob": -0.10155981132783086, "no_speech_prob": 1.2015840579471249e-11, "compression_ratio": 1.1815068493150684 }, { "start": 787.14, "end": 790.8, "text": "自动模型评估和推理加速上的突破", "chunk": 1, "language": "zh", "avg_logprob": -0.10155981132783086, "no_speech_prob": 1.2015840579471249e-11, "compression_ratio": 1.1815068493150684 }, { "start": 790.8, "end": 793.56, "text": "能以周级别的速度直接部署上线", "chunk": 1, "language": "zh", "avg_logprob": -0.10155981132783086, "no_speech_prob": 1.2015840579471249e-11, "compression_ratio": 1.1815068493150684 }, { "start": 793.56, "end": 795.72, "text": "让科研成果在第一秒", "chunk": 1, "language": "zh", "avg_logprob": -0.10155981132783086, "no_speech_prob": 1.2015840579471249e-11, "compression_ratio": 1.1815068493150684 }, { "start": 795.72, "end": 797.68, "text": "直接转化为商业账单", "chunk": 1, "language": "zh", "avg_logprob": -0.10155981132783086, "no_speech_prob": 1.2015840579471249e-11, "compression_ratio": 1.1815068493150684 }, { "start": 797.68, "end": 801.34, "text": "形成技术产品与算力的超级费轮", "chunk": 1, "language": "zh", "avg_logprob": -0.10155981132783086, "no_speech_prob": 1.2015840579471249e-11, "compression_ratio": 1.1815068493150684 }, { "start": 801.34, "end": 802.74, "text": "更关键的是", "chunk": 1, "language": "zh", "avg_logprob": -0.10155981132783086, "no_speech_prob": 1.2015840579471249e-11, "compression_ratio": 1.1815068493150684 }, { "start": 802.74, "end": 805.96, "text": "谷歌正在悄无声息地改变自己的盈利模式", "chunk": 1, "language": "zh", "avg_logprob": -0.0909455672077749, "no_speech_prob": 1.3475120325123768e-11, "compression_ratio": 1.212624584717608 }, { "start": 805.96, "end": 810.0, "text": "当用户不再翻页点击搜索广告链接", "chunk": 1, "language": "zh", "avg_logprob": -0.0909455672077749, "no_speech_prob": 1.3475120325123768e-11, "compression_ratio": 1.212624584717608 }, { "start": 810.0, "end": 812.74, "text": "而是让Gamlet给出最终答案时", "chunk": 1, "language": "zh", "avg_logprob": -0.0909455672077749, "no_speech_prob": 1.3475120325123768e-11, "compression_ratio": 1.212624584717608 }, { "start": 812.74, "end": 814.02, "text": "谷歌的解法是", "chunk": 1, "language": "zh", "avg_logprob": -0.0909455672077749, "no_speech_prob": 1.3475120325123768e-11, "compression_ratio": 1.212624584717608 }, { "start": 814.02, "end": 817.3, "text": "直接占据agent决策路径的入口", "chunk": 1, "language": "zh", "avg_logprob": -0.0909455672077749, "no_speech_prob": 1.3475120325123768e-11, "compression_ratio": 1.212624584717608 }, { "start": 817.3, "end": 820.64, "text": "当Gamlet在每日简报或者个人助理里", "chunk": 1, "language": "zh", "avg_logprob": -0.0909455672077749, "no_speech_prob": 1.3475120325123768e-11, "compression_ratio": 1.212624584717608 }, { "start": 820.64, "end": 823.26, "text": "帮你规划行程推荐服务时", "chunk": 1, "language": "zh", "avg_logprob": -0.0909455672077749, "no_speech_prob": 1.3475120325123768e-11, "compression_ratio": 1.212624584717608 }, { "start": 823.26, "end": 825.44, "text": "它在生成方案的同时", "chunk": 1, "language": "zh", "avg_logprob": -0.0909455672077749, "no_speech_prob": 1.3475120325123768e-11, "compression_ratio": 1.212624584717608 }, { "start": 825.44, "end": 827.96, "text": "精准接入商业API入口", "chunk": 1, "language": "zh", "avg_logprob": -0.0909455672077749, "no_speech_prob": 1.3475120325123768e-11, "compression_ratio": 1.212624584717608 }, { "start": 827.96, "end": 831.5, "text": "谷歌由此从一个信息检索广告商", "chunk": 1, "language": "zh", "avg_logprob": -0.0909455672077749, "no_speech_prob": 1.3475120325123768e-11, "compression_ratio": 1.212624584717608 }, { "start": 831.5, "end": 834.64, "text": "变成了全局智能决策的开门人", "chunk": 1, "language": "zh", "avg_logprob": -0.08348489576770414, "no_speech_prob": 1.3527291266302033e-11, "compression_ratio": 1.1428571428571428 }, { "start": 834.64, "end": 836.92, "text": "抽取服务分账和佣金", "chunk": 1, "language": "zh", "avg_logprob": -0.08348489576770414, "no_speech_prob": 1.3527291266302033e-11, "compression_ratio": 1.1428571428571428 }, { "start": 836.92, "end": 839.58, "text": "拆决员Jeff Dean的离职创业", "chunk": 1, "language": "zh", "avg_logprob": -0.08348489576770414, "no_speech_prob": 1.3527291266302033e-11, "compression_ratio": 1.1428571428571428 }, { "start": 839.58, "end": 842.14, "text": "与柯磊掌舵下的Deep Mind重组", "chunk": 1, "language": "zh", "avg_logprob": -0.08348489576770414, "no_speech_prob": 1.3527291266302033e-11, "compression_ratio": 1.1428571428571428 }, { "start": 842.14, "end": 846.44, "text": "我们终于能看清这场硅谷AI大戏的全貌了", "chunk": 1, "language": "zh", "avg_logprob": -0.08348489576770414, "no_speech_prob": 1.3527291266302033e-11, "compression_ratio": 1.1428571428571428 }, { "start": 846.44, "end": 851.14, "text": "硅谷的AI战争正是告别了依靠概念和论文引用量", "chunk": 1, "language": "zh", "avg_logprob": -0.08348489576770414, "no_speech_prob": 1.3527291266302033e-11, "compression_ratio": 1.1428571428571428 }, { "start": 851.14, "end": 853.78, "text": "获取高估值的学术竞赛阶段", "chunk": 1, "language": "zh", "avg_logprob": -0.08348489576770414, "no_speech_prob": 1.3527291266302033e-11, "compression_ratio": 1.1428571428571428 }, { "start": 853.78, "end": 856.7, "text": "全面跨入了拼推理成本", "chunk": 1, "language": "zh", "avg_logprob": -0.08348489576770414, "no_speech_prob": 1.3527291266302033e-11, "compression_ratio": 1.1428571428571428 }, { "start": 856.7, "end": 858.38, "text": "拼系统稳定性", "chunk": 1, "language": "zh", "avg_logprob": -0.08348489576770414, "no_speech_prob": 1.3527291266302033e-11, "compression_ratio": 1.1428571428571428 }, { "start": 858.38, "end": 859.88, "text": "拼终端控制力", "chunk": 1, "language": "zh", "avg_logprob": -0.08348489576770414, "no_speech_prob": 1.3527291266302033e-11, "compression_ratio": 1.1428571428571428 }, { "start": 859.88, "end": 863.92, "text": "以及拼商业造学能力的工程与商业深水去", "chunk": 1, "language": "zh", "avg_logprob": -0.07784231723342812, "no_speech_prob": 1.1754208717462689e-11, "compression_ratio": 1.2541254125412542 }, { "start": 863.92, "end": 867.8, "text": "对于我们所有在做AI研发架构设计", "chunk": 1, "language": "zh", "avg_logprob": -0.07784231723342812, "no_speech_prob": 1.1754208717462689e-11, "compression_ratio": 1.2541254125412542 }, { "start": 867.8, "end": 869.92, "text": "或者技术创业的朋友来说", "chunk": 1, "language": "zh", "avg_logprob": -0.07784231723342812, "no_speech_prob": 1.1754208717462689e-11, "compression_ratio": 1.2541254125412542 }, { "start": 869.92, "end": 871.72, "text": "谷歌的这一系列辩证", "chunk": 1, "language": "zh", "avg_logprob": -0.07784231723342812, "no_speech_prob": 1.1754208717462689e-11, "compression_ratio": 1.2541254125412542 }, { "start": 871.72, "end": 874.86, "text": "留下了三条极其宝贵的工程启发", "chunk": 1, "language": "zh", "avg_logprob": -0.07784231723342812, "no_speech_prob": 1.1754208717462689e-11, "compression_ratio": 1.2541254125412542 }, { "start": 874.86, "end": 876.94, "text": "首先不要盲目迷信", "chunk": 1, "language": "zh", "avg_logprob": -0.07784231723342812, "no_speech_prob": 1.1754208717462689e-11, "compression_ratio": 1.2541254125412542 }, { "start": 876.94, "end": 879.58, "text": "纯域训练参数的单位堆叠", "chunk": 1, "language": "zh", "avg_logprob": -0.07784231723342812, "no_speech_prob": 1.1754208717462689e-11, "compression_ratio": 1.2541254125412542 }, { "start": 879.58, "end": 881.94, "text": "多去研究推理器计算", "chunk": 1, "language": "zh", "avg_logprob": -0.07784231723342812, "no_speech_prob": 1.1754208717462689e-11, "compression_ratio": 1.2541254125412542 }, { "start": 881.94, "end": 886.12, "text": "通过过程奖励模型打分与蒙特卡洛数搜索", "chunk": 1, "language": "zh", "avg_logprob": -0.07784231723342812, "no_speech_prob": 1.1754208717462689e-11, "compression_ratio": 1.2541254125412542 }, { "start": 886.12, "end": 889.3, "text": "让模型在推理阶段学会自我纠错", "chunk": 1, "language": "zh", "avg_logprob": -0.07784231723342812, "no_speech_prob": 1.1754208717462689e-11, "compression_ratio": 1.2541254125412542 }, { "start": 889.3, "end": 893.12, "text": "这是低成本实现逻辑跨越的唯一解法", "chunk": 1, "language": "zh", "avg_logprob": -0.10054173495600133, "no_speech_prob": 1.1170740547850055e-11, "compression_ratio": 1.1981424148606812 }, { "start": 893.12, "end": 896.58, "text": "其次要学会拥抱端云分层架构", "chunk": 1, "language": "zh", "avg_logprob": -0.10054173495600133, "no_speech_prob": 1.1170740547850055e-11, "compression_ratio": 1.1981424148606812 }, { "start": 896.58, "end": 898.52, "text": "平气无论任务大小", "chunk": 1, "language": "zh", "avg_logprob": -0.10054173495600133, "no_speech_prob": 1.1170740547850055e-11, "compression_ratio": 1.1981424148606812 }, { "start": 898.52, "end": 901.62, "text": "一力调用千亿大模型的粗暴思维", "chunk": 1, "language": "zh", "avg_logprob": -0.10054173495600133, "no_speech_prob": 1.1170740547850055e-11, "compression_ratio": 1.1981424148606812 }, { "start": 901.62, "end": 903.4, "text": "学习Jamler Spark的设计", "chunk": 1, "language": "zh", "avg_logprob": -0.10054173495600133, "no_speech_prob": 1.1170740547850055e-11, "compression_ratio": 1.1981424148606812 }, { "start": 903.4, "end": 906.9, "text": "用轻量级模型做好前置意图路由", "chunk": 1, "language": "zh", "avg_logprob": -0.10054173495600133, "no_speech_prob": 1.1170740547850055e-11, "compression_ratio": 1.1981424148606812 }, { "start": 906.9, "end": 910.02, "text": "才能在高并发任务中守住成本底线", "chunk": 1, "language": "zh", "avg_logprob": -0.10054173495600133, "no_speech_prob": 1.1170740547850055e-11, "compression_ratio": 1.1981424148606812 }, { "start": 910.02, "end": 913.8, "text": "最后我们需要重新思考软件入口的定义", "chunk": 1, "language": "zh", "avg_logprob": -0.10054173495600133, "no_speech_prob": 1.1170740547850055e-11, "compression_ratio": 1.1981424148606812 }, { "start": 913.8, "end": 916.26, "text": "未来的软件竞争不再是争夺", "chunk": 1, "language": "zh", "avg_logprob": -0.10054173495600133, "no_speech_prob": 1.1170740547850055e-11, "compression_ratio": 1.1981424148606812 }, { "start": 916.26, "end": 918.42, "text": "手机屏幕上的APP图标位置", "chunk": 1, "language": "zh", "avg_logprob": -0.10054173495600133, "no_speech_prob": 1.1170740547850055e-11, "compression_ratio": 1.1981424148606812 }, { "start": 918.42, "end": 920.98, "text": "而是争夺能否被系统及agent", "chunk": 1, "language": "zh", "avg_logprob": -0.09754295782609419, "no_speech_prob": 9.368133005449586e-12, "compression_ratio": 0.9366197183098591 }, { "start": 920.98, "end": 923.16, "text": "API决策链所调用", "chunk": 1, "language": "zh", "avg_logprob": -0.09754295782609419, "no_speech_prob": 9.368133005449586e-12, "compression_ratio": 0.9366197183098591 }, { "start": 923.16, "end": 925.12, "text": "技术不相信幻觉", "chunk": 1, "language": "zh", "avg_logprob": -0.09754295782609419, "no_speech_prob": 9.368133005449586e-12, "compression_ratio": 0.9366197183098591 }, { "start": 925.12, "end": 926.72, "text": "商业不相信眼泪", "chunk": 1, "language": "zh", "avg_logprob": -0.09754295782609419, "no_speech_prob": 9.368133005449586e-12, "compression_ratio": 0.9366197183098591 }, { "start": 926.72, "end": 927.86, "text": "这里是All Insight", "chunk": 1, "language": "zh", "avg_logprob": -0.09754295782609419, "no_speech_prob": 9.368133005449586e-12, "compression_ratio": 0.9366197183098591 }, { "start": 927.86, "end": 928.76, "text": "我们下期见", "chunk": 1, "language": "zh", "avg_logprob": -0.09754295782609419, "no_speech_prob": 9.368133005449586e-12, "compression_ratio": 0.9366197183098591 } ], "segmented_transcription": { "status": "SEGMENTED_TRANSCRIPT_OK", "created_at": "2026-08-10T08:44:43", "manifest": "/Users/sagawa/AI_WORK/video_notes/url/20260810_084409__谷歌AI大地震_代码之神_带队离职_为什么Gemini反而变得更好用了/segmented_work/chunk_manifest.json", "audio": "/Users/sagawa/AI_WORK/video_notes/url/20260810_084409__谷歌AI大地震_代码之神_带队离职_为什么Gemini反而变得更好用了/audio_16k_mono.wav", "audio_sha256": "69f574de4422402e2af2e35d0d2cc18fbc185cc27794fddd081de3781065194d", "audio_duration": 929.100063, "model": "mlx-community/whisper-large-v3-turbo", "language_hint": null, "condition_on_previous_text": false, "chunk_count": 2, "resumed_chunks": 0, "merged_segment_count": 330, "final_segment_end": 928.76, "global_repeat_run": 1, "global_repeat_text": "欢迎回到On Insight", "tail_repeat_run": 1, "tail_repeat_text": "获取高估值的学术竞赛阶段", "hallucination_region_count": 0, "hallucination_regions": [], "qc_ok": true, "qc_errors": [] } }