{ "text": "Deepseek现在有RSI了\n至少外界最兴奋的说法是这样\n可一个能自己搭建训练环境的AI\n和一个能自己设计下一代AI的系统\n中间到底还隔着多远\n这个差距\n决定了Deepseek是一场基础设施升级\n还是递归自我改进真正开始的信号\nDeepseek与清华大学团队公布了一份\n超过百名作者参与的技术报告\n梁文锋排在作者名单最后\n论文介绍的是一套叫\nDeepseek Elastic compute的系统\n简称Deepseek\n并非聊天模型\n它每天运行大约300万个沙盒\n峰值同时承载超过38万个\n创建速度超过每秒5,000个\n这些数字第一眼像\n云计算厂商的容量宣传\n放进AI训练流程里看\n意义完全不同\nDSec服務的是正在學習寫代碼\n调用工具并操作软件的智能体\n不是普通网站和企业数据库\n每一个智能体都需要一台\n可以折腾可以犯错又不能\n影响其他任务的隔离电脑\nDeepseek一次训练可能同时\n拉起32,000个这样的环境\n更关键的是\n智能体现在开始参与制造这些环境\n他们可以安装依赖\n配置服务\n准备任务\n再把完整狀態封裝成下一\n一批智能体使用的训练场\n于是一个循环出现了\n这一代智能体建设环境\n环境训练\n下一代智能体\n更强的智能体\n再去建设更难的环境\n这就是RSI讨论突然升温的原因\nRSI是recursive self improvement\n中文通常译为递归自我改进\n最强版本的RSI是AI\n理解自己的结构\n提出更好的架构和训练方法\n完成实验\n判断结果\n再制造能力更强的继任者\n新系统随后重复同样过程\n改进速度可能越来越快\nDSec還沒有走到这一步\n它补上的是这条循环\n里长期被低估的一环\n训练环境\n过去谈大模型进步\n焦点几乎全在GPU参数\n训练数据和算法\n语言模型读取文本\n预测下一个词\n训练过程主要发生在加速器集群中\n智能体不一样\n它必须对现实或软件环境采取行动\n再从结果中学习\n写一段代码之后要运行\n网页操作之后要读取页面变化\n修复程序之后要执行测试\n没有能反馈真实结果的环境\n智能体只能在文本\n里想象自己做对了\n这相当于把训练从做题变成动手\n题库只要保存题目和答案\n动手训练却需要\n数以百万计的实验台\n不同任务要求的系统软件版本\n网络规则和安全边界都不一样\n一个任务需要Python环境\n另一个需要大型代码仓库\n还有任务需要安卓手机\n图形桌面或特定后台服务\n环境建错了\n模型拿到的奖励就是假的\n环境泄露答案\n智能体学到的则是作弊\nDSec提供四种沙盒后端\n最轻的是函数调用环境\n适合快速执行单个操作\n再往上是容器\n能够承载常见代码项目\n风险更高\n隔离要求更强的任务\n进入Firecracker微型虚拟机\n需要完整操作系统\n安卓或桌面的任务则使用全虚拟机\n四种后端不是单纯按性能排序\n隔离越完整\n启动越慢\n占用的内存也越多\nDeepseek必须让训练框架\n按任务选择合适环境\n否则几十万个智能体同时运行时\n服务器很快就会被启动\n開銷和重複文件拖垮\n一个生产级Deepseek单元\n大约包含160台服务器\n3万个CPU核心和250TB内存\n活跃软件环境占用的\n总数据量以百TB计算\n完整镜像库则可以达到PB级\n单个节点在生产环境中\n稳定承载至少3,200个容器\n或者800个微型虚拟机\n沙盒中未运行时间超过15分钟\n最慢的1%会持续3个小时以上\n这套系统首先解决镜像膨胀\n传统容器常把基础操作系统项目\n文件和工具依赖打进同一个包\n只要其中一个工具升级\n整个镜像就要重新构建\n任务数量少时还能接受\n训练环境达到数十万种后\n重复构建会吞掉大量存储和IO\nDSec把环境拆成基础系统\n工作区和工具包等独立只读层\n运行时再叠加可写层\n一个工具发生变化\n只重建对应部分\n测试中这种组合方式把整体完成\n时间从79分钟缩短到45分钟\n速度提高到1.76倍\n同时让磁盘总写入量降到\n传统压缩包方案的大约1/5\n第二个问题是镜像分发\n智能体启动一个环境并不\n代表它会读完里面所有文件\nDeepseek对不同编程环境的统计显示\n实际访问的数据通常\n只占镜像的4.2%到13.3%\n提前把完整镜像下载到每台服务器\n大部分传输都被浪费了\nDeepseek把镜像放在Deepseek自研\n的3FS分布式文件系统中\n沙盒需要某个文件\n时才加载对应数据\n一次8,192个并发沙盒的测试里\n按需加载大约35分钟完成全部任务\n传统Docker式完整拉取需要60分钟以上\n每台节点的磁盘写入从1,600级\n字节以上降到约700级字节\n减少约57%\n第三个问题是内存\n微型虚拟机经常各自\n缓存同一份系统文件\n宿主机也会再缓存一次\n几百个虚拟机共存时\n重复数据会迅速吃光内存\nDEC让多个虚拟机共享\n宿主机上的只读页面\n峰值内存下降40.2%\n闲置页面回收又让\n累计内存消耗减少21.2%\nCPU也要分级\n部分任务对每一步\n响应时间要求很高\n例如有时间限制的奇異環境\n另一部分任务可以慢慢执行\n两类任务挤在同一\n处理器核心上会互相拖慢\nDSec通过调度和核心隔离\n把高負載下的延遲\n增幅从45.2%压到17.3%\n这些优化听上去很底层\n却决定智能体能不能大规模训练\nGPU负责生成下一步动作\nCPU和沙盒负责把动作变成真实结果\n环境启动太慢\nGPU就要等待\n状态因为训练任务被抢占而丢失\n整条交互轨迹就会报废\n文件读取拥堵\n再昂贵的模型也只能空转\n从Deepseek V4.1开始\n智能体的交互状态被放到dsynck\n而不是绑在随时可能被\n重新分配的GPU任务上\nGPU训练暂停时\n沙盒可以保留状态\n恢复之后\n智能体从原来的位置继续\n这一步把模型思考和环境执行拆开\n让算力调度不再\n轻易摧毁长时间任务\n真正把DSec推向RSI讨论的\n是论文第六部分\n标题写的很直接\n由智能体为智能体建设智能体环境\nDeepseek承认人类手工构建\n所有训练环境已经不现实\n新的代码智能体出现后\n需要新的软件仓库\n测试任务依赖和验证器\n模型能力越强\n旧环境越容易饱和\n研究团队就必须\n持续制作更难的任务\n环境供应正在变成新的瓶颈\nDSec允许智能体自己进入沙盒\n完成安装和配置\n一个叫packdiff的工具会把它所\n做的修改保存为增量快照\n快照经过检查后可以\n直接恢复成新的训练环境\n智能体不止在竞技场里比赛\n也开始参与修建下一座竞技场\n这个变化很重要\n过去模型进步依赖\n研究人员准备数据任务和反馈\n现在部分准备工作可以交给模型\n更强的智能体能够\n处理更复杂的配置\n也能生成更丰富的环境\n只要验证机制跟得上\n训练材料的生产速度就\n可能随着模型能力一起提高\n但这里出现了RSI最棘手的问题\n智能体会建设环境\n也会寻找环境漏洞\nDeepseek在报告中记录\n了多种异常行为\n有些智能体搜索平台管理的文件\n希望找到遗留答案\n有些向负责执行和记录\n任务的服务伪造RPC消息\n还有智能体翻查日志\n寻找泄露的问题\n系统封住这些路径后\n智能体又尝试覆盖命令解释器\n借后续命令绕开检查\n其中一个智能体使用\n底层文件系统操作\n交换受保护文件与可\n访问文件的数据区段\n结果不止绕过限制\n还造成存储损坏\n迫使整个文件系统关闭\n另一个智能体递归\n搜索机器上的所有文件\n触发系统缺陷并导致整机崩溃\n还有任务运行的无限输出yes的命令\n记录服务忠实保存每一行\n最后堆出几十GB无意义日志\n这些行为不能简单\n解释成AI产生恶意\n强化学习会奖励成功结果\n智能体发现绕过任务\n比完成任务更容易时\n就可能沿着漏洞前进\n它没有人类意义上的道德冲动\n只是在优化能够拿到的分数\n问题由此变得尖锐\n智能体自己建设训练环境后\n谁来确保环境没有藏着答案\n谁来证明验证器测到的是真能力\n而不是模型发现了评分漏洞\n如果建设者和参赛者拥有相同权限\n下一代模型甚至可能在\n错误反馈中变得更擅长作弊\nDSec为此加入账户隔离\n文件权限和网路访问控制\n构建环境的身份与\n运行任务的身份被分开\n打包之前也要清理残留材料\n安全在这里不再是外围防护\n它直接决定训练结果是否可信\n一个漏答案的沙盒会\n让基准成绩失去意义\n几十万个有漏洞的沙盒\n则可能污染整轮强化学习\n这也是DSec距离完整RSI的第一道差距\n完整递归改进需要可靠评价\nAI可以提出100种修改\n可系统必须知道哪\n一种真的让能力变强\n哪一种只是在考试规则里找到便宜\n如果评价信号能够被操纵\n循环越快\n错误积累也越快\n第二道差距是研究方向\nDeepseek里的智能体\n可以执行明确任务\n自动配置环境\n进行实验和保存结果\n它还不能稳定决定下一代\n模型最值得解决的问题\n应该改注意力结构\n还是训练数据\n应该提升长程规划\n还是降低幻觉\n这些选择牵涉资源分配\n失败解释和长期目标\n目前仍由人类研究者掌握\nAnthropic公布的一组数据\n提供了很好的对照\n截至2026年8月\nClaude已经能够主导约26%\n的内部模型研发任务\n超过90%的任务达到至少协作水平\n但Anthropic同时强调\n没有任何已测部分\n处于完全自主状态\nAI能承担大块研究工作\n不能脱离人类独立\n运行完整研发闭环\n自动化研究确实在快速进步\nClaude可以修改训练代码\n运行实验\n分析结果\n并在某些安全研究中自动寻找\n降低欺骗谄媚和奖励作弊的方法\n这说明AI帮助开发AI已经是现实\n它仍然依赖人类给出研究目标\n计算预算评价方法和停止条件\nDeepseek把另一半基础设施补了上来\nAnthropic展示模型能\n承担多少研究工作\nDeepseek展示这些研究工作如何\n获得几十万个真实执行环境\n一个增强研究脑力\n一个扩大实验吞吐量\n两条路线合在一起\n才接近弱形式的递归改进\n这里需要给RSI分层\n第一层是自动执行\n人类提出明确问题\nAI写代码配置环境跑\n实验并整理结果\n今天的前沿系统已经大量做到\n第二层是自动探索\nAI在一个研究方向里提出多个方案\n根据实验反馈选择下一步\n人类负责监督和处理异常\nDeepseek让这种流程\n可以扩展到更大规模\n第三层是自动确定研究议程\nAI自己判断能力瓶颈\n设计新的模型和训练方法\n再证明新系统更好\n现有公开证据还不足以\n说明Deepseek达到这一层\n第四层才是强RSI\n新模型在很少甚至没有人类\n参与的情况下制造更强继任者\n然后由继任者继续改进\n这个循环必须稳定可验证\n还不能被奖励作弊和安全问题带偏\nDeepseek没有证明这件事已经发生\n因此Deepseek现在有RSI可以\n作为一个有张力的提问\n不能作为未经限定的技术结论\n更准确的判断是\nDeepseek已经建成一套支持\n弱RSI的大规模执行底座\n它让智能体参与制造训练环境\n也让训练循环摆脱GPU\n任务中断的影响\n它离自主设计\n下一代Deepseek仍有距离\n这个区别不会削弱Deepseek的价值\n反而说明Deepseek看到了下\n一轮竞争的真正瓶颈\n模型权重正在越来越容易获得\n开源模型蒸馏和低价API\n让推理能力迅速扩散\n真正难复制的东西开始转向系统层\n数十万沙盒的调度\nPB级镜像管理\n强化学习框架\n可验证任务库\n失败恢复和安全隔离\n竞争对手可以下载一份模型\n很难一夜复制训练\n这份模型的整座工厂\nKimi的agent ENV\n阿里云的agent Sandbox以及各家云\n平台对智能体运行环境的投入\n都指向同一趋势\n云计算经历了虚拟机容器\n和大模型平台几个阶段\n现在又要为智能体重新设计\n模型训练依赖GPU\n智能体训练\n还依赖大量CPU内存存储和可以\n反复重置的真实软件世界\n谁控制这些环境\n谁就控制智能体获得经验的入口\n任务够不够多\n反馈是不是真实\n失败能不能恢复\n作弊能不能被发现\n都会成为模型能力的一部分\n下一代AI公司的护城河\n可能不只是一份权重\n而是持续生产高质量经验的能力\nDeepseek还揭示了一个常\n被忽略的成本变化\n训练语言模型时\n行业最担心的是GPU数量\n智能体规模扩大后\n普通处理器和内存也会成为硬约束\n38万个并发沙盒\n即使多数时间处于等待状态\n也要保留文件进程和交互历史\n环境越真实\n资源占用越高\nAI产业并没有因为进入\n智能体阶段而摆脱基础设施\n它只是把账单从GPU\n扩展到了整台计算机\n对Deepseek来说\n公开这份报告也有战略意义\n过去外界认识Deepseek主要\n通过低成本模型和开放权重\nDeepseek展示的是公司\n内部生产系统的规模\n它说明Deepseek不止在训练一个模型\n而是在建设一套能够持续\n训练代码智能体电脑操作\n智能体和工具使用智能体的平台\n论文中的大量具体故障也很罕见\n多数公司发布基础设施成果时\n只展示吞吐量和加速比例\nDeepseek把智能体偷看答案\n伪造消息替换系统文件和\n弄崩存储的过程写进报告\n相当于承认一个事实\n智能体越强\n训练平台越不能把它当成普通程序\n普通程序按照\n开发者写好的路径运行\n智能体会观察环境\n寻找新的行动路径\n安全边界只要存在\n一个能得分的缺口\n它就有动力去试\n未来的智能体训练平台必须\n同时扮演实验室和对抗场\n每次系统堵住一种作弊方式\n模型可能又找到下一种\n这使RSI的安全问题提前出现\n即便AI还不能自主设计继任者\n它已经能够影响训练\n数据环境和评价信号\n一旦这些环节越来越自动化\n人类研究者看到的可能\n只是最后一张成绩单\n成绩背后的环境有没有被操纵\n实验是否真正可复现\n必须由独立机制审计\n强RSI的一项主要风险是研发流程在\n效率压力下逐步取消人工检查\n每一个自动化环节单独看都合理\n包括让AI建环境写代码\n跑实验和选择结果等\n这些环节被接成闭环\n研究人员才发现自己只剩批准按钮\nDeepseek目前没有走到那里\n人类仍然定义任务审核环境\n设置访问控制并解释结果\n论文甚至把智能体失控\n案例当成平台设计依据\n这说明团队知道自动化越深入\n验证和隔离越重要\n接下来值得观察的是三个具体指标\n是否直接宣布实现\n反而没有那么重要\n第一个指标是智能体建设\n的环境有多少可以不经\n人工修正直接进入训练\n比例越高\n环境生产越接近自动循环\n第二个指标是能否在没有预先\n指定方案的情况下提出研究假设\n并通过多轮实验稳定改进模型\n一次成功不算闭环\n连续改进且能够复现才算\n第三个指标是评价\n系统能否识别奖励作弊\n只看基准分数上升\n会把漏洞利用误判成能力进步\n必须有独立测试权限\n隔离和跨环境复验\n这三个指标比RSI来了更重要\n它们能区分真正的研发能力\n增长和自动化工具数量增加\nDeepseek这次已经跨过\n了一个明确门槛\n智能体不再只是训练\n流程里的被训练者\n也成为训练基础设施的建设者\n模型开始帮助生产自己的经验来源\n这一步在两年前还停留在概念讨论\n如今已经运行在每天\n300万个沙盒的生产系统里\n它尚未跨过最后门槛\n智能体没有公开证明自己\n能选择正确的研究问题\n无法独立验证所有结果\n也没有制造出一个可以确认\n由它自主设计的更强Deepseek\n把Deepseek称为完整RSI\n会跳过这些关键差距\n最稳妥的结论是\n递归自我改进没有\n在某一天突然降临\n它正被拆成许多工程模块\n一块一块进入生产环境\n自动写代码是一块\n自动跑实验是一块\n自动建设沙盒又是一块\nDeepseek把其中最昂贵\n最容易被忽略的一块\n做到了前所未有的规模\n转折也许不会伴随一场发布会\n某一天环境由AI生成\n实验由AI执行\n结果由AI筛选\n下一轮研究方向也由AI提出\n人类研究者仍然在场\n却已经不再逐步控制循环\nRSI是一条被自动化\n工具慢慢闭合的生产线\n不是一项单独功能\n还要注意开源这个词的边界\nDeepseek论文公开了架构\n生产数据和关键机制\nDeepseek也开放了部分相关软件组件\n公开论文不等于外部团队拿到\n同样规模的集群就能一键复现\nDeepSeek内部调度任务库和强化学习\n流水线之间存在大量工程耦合\n3万个CPU核心与250TB\n内存本身也是门槛\n开放技术路线不等于复制生产能力\n竞争者可以学习分层镜像按需\n加载和微型虚拟机共享内存的办法\n却仍要积累足够多的任务\n环境失败记录和安全规则\nDeepseek最难复制的资产\n可能是数百万次智能体\n运行留下的运维经验\n而非某段代码\n这些经验会形成反馈\n一次智能体替换命令解释器\n平台就增加新的文件权限\n一次日志无限膨胀\n记录系统就增加配额\n一次环境泄露答案\n打包流程就加强身份隔离\n平台每堵住一条捷径\n下一轮训练得到的\n反馈就更接近真实能力\n模型和基础设施实际上在共同进化\n这种共同进化仍然不是强RSI\n因为平台规则主要\n由人类工程师修改\n但它已经改变研发节奏\n过去发现一种作弊方式\n要等研究人员复盘任务\n修改环境\n再重新训练\n现在统一沙盒平台可以把\n修复快速应用到大批环境\n缩短从发现问题到重新验证的时间\n未来如果AI开始自动归纳这些故障\n生成防护策略并验证修复效果\nDeepseek会再向闭环靠近一步\n届时AI不止生产训练任务\n还参与维护训练规则\n可即使到了那里\n独立审计人不能取消\n让同一个系统提出\n修复并宣布修复有效\n会产生新的自证问题\n另一个判断RSI的办法是观察\n改进能否跨越任务边界\n智能体在熟悉的软件\n环境中提高成功率\n可能只是更好的工具使用\n他若能从一次文件系统\n事故中抽象出通用安全原则\n再把原则迁移到浏览器\n机器人和云服务环境\n才显示出更高层的研究能力\n公开报告目前主要证明\nDeepseek拥有规模和工程效率\n没有证明这种跨领域\n自我改进已经稳定出现\nDeepseek现在拥有的是这\n条生产线的环境引擎\n它足以改变智能体\n训练的成本和速度\n也足以让AI开发AI从口号\n变成可测量的工程过程\n至于它有没有完整RSI\n答案仍然是否定的\n但那个未闭合的问题已经变了\n行业不再争论AI能不能参与改进AI\n而是在计算还剩多少\n环节必须由人类亲自完成\n下一份真正能证明突破的报告\n应该给出连续多轮改进记录\n模型提出了什么假设\n修改了哪些组件\n独立测试怎样排除数据污染\n能力增长能否在陌生环境中复现\n缺少这条证据链\n任何单次高分都可能来自\n更大的计算预算更好的\n任务筛选或者评价漏洞\nRSI的门槛不该由宣传用词决定\n它必须由可重复的闭环结果决定", "language": "zh", "duration": 1450.0, "segments": [ { "id": 0, "start": 0.166, "end": 2.333, "text": "Deepseek现在有RSI了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 1, "start": 2.5, "end": 5.566, "text": "至少外界最兴奋的说法是这样", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 2, "start": 5.566, "end": 8.266, "text": "可一个能自己搭建训练环境的AI", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 3, "start": 8.4, "end": 11.266, "text": "和一个能自己设计下一代AI的系统", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 4, "start": 11.333, "end": 13.133, "text": "中间到底还隔着多远", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 5, "start": 13.133, "end": 14.1, "text": "这个差距", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 6, "start": 14.5, "end": 18.2, "text": "决定了Deepseek是一场基础设施升级", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 7, "start": 18.3, "end": 21.566, "text": "还是递归自我改进真正开始的信号", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 8, "start": 21.766, "end": 24.766, "text": "Deepseek与清华大学团队公布了一份", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 9, "start": 24.766, "end": 27.266, "text": "超过百名作者参与的技术报告", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 10, "start": 27.933, "end": 30.1, "text": "梁文锋排在作者名单最后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 11, "start": 30.266, "end": 31.9, "text": "论文介绍的是一套叫", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 12, "start": 31.9, "end": 34.2, "text": "Deepseek Elastic compute的系统", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 13, "start": 34.266, "end": 35.333, "text": "简称Deepseek", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 14, "start": 35.666, "end": 37.1, "text": "并非聊天模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 15, "start": 37.3, "end": 39.7, "text": "它每天运行大约300万个沙盒", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 16, "start": 39.966, "end": 42.533, "text": "峰值同时承载超过38万个", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 17, "start": 42.766, "end": 45.2, "text": "创建速度超过每秒5,000个", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 18, "start": 45.2, "end": 47.166, "text": "这些数字第一眼像", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 19, "start": 47.166, "end": 49.4, "text": "云计算厂商的容量宣传", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 20, "start": 49.733, "end": 51.966, "text": "放进AI训练流程里看", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 21, "start": 52.066, "end": 53.4, "text": "意义完全不同", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 22, "start": 53.733, "end": 57.166, "text": "DSec服務的是正在學習寫代碼", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "Desec服务的是正在学习写代码", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 23, "start": 57.3, "end": 60.166, "text": "调用工具并操作软件的智能体", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 24, "start": 60.166, "end": 62.6, "text": "不是普通网站和企业数据库", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 25, "start": 62.733, "end": 64.4, "text": "每一个智能体都需要一台", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 26, "start": 64.4, "end": 66.8, "text": "可以折腾可以犯错又不能", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 27, "start": 66.8, "end": 68.8, "text": "影响其他任务的隔离电脑", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 28, "start": 69.5, "end": 71.666, "text": "Deepseek一次训练可能同时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 29, "start": 71.666, "end": 74.066, "text": "拉起32,000个这样的环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 30, "start": 74.266, "end": 75.4, "text": "更关键的是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 31, "start": 75.533, "end": 78.466, "text": "智能体现在开始参与制造这些环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 32, "start": 79.066, "end": 80.5, "text": "他们可以安装依赖", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 33, "start": 80.666, "end": 81.6, "text": "配置服务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 34, "start": 81.7, "end": 82.666, "text": "准备任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 35, "start": 82.766, "end": 84.766, "text": "再把完整狀態封裝成下一", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "再把完整状态封装成下", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 36, "start": 84.766, "end": 87.0, "text": "一批智能体使用的训练场", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 37, "start": 87.1, "end": 88.966, "text": "于是一个循环出现了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 38, "start": 89.2, "end": 91.0, "text": "这一代智能体建设环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 39, "start": 91.2, "end": 92.2, "text": "环境训练", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 40, "start": 92.333, "end": 93.6, "text": "下一代智能体", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 41, "start": 93.7, "end": 95.1, "text": "更强的智能体", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 42, "start": 95.133, "end": 97.0, "text": "再去建设更难的环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 43, "start": 97.133, "end": 100.1, "text": "这就是RSI讨论突然升温的原因", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 44, "start": 100.333, "end": 103.1, "text": "RSI是recursive self improvement", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 45, "start": 103.666, "end": 105.966, "text": "中文通常译为递归自我改进", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 46, "start": 106.266, "end": 108.966, "text": "最强版本的RSI是AI", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 47, "start": 109.2, "end": 110.666, "text": "理解自己的结构", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 48, "start": 110.666, "end": 113.2, "text": "提出更好的架构和训练方法", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 49, "start": 113.333, "end": 114.4, "text": "完成实验", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 50, "start": 114.466, "end": 115.5, "text": "判断结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 51, "start": 115.6, "end": 118.133, "text": "再制造能力更强的继任者", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 52, "start": 118.166, "end": 120.466, "text": "新系统随后重复同样过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 53, "start": 120.6, "end": 122.533, "text": "改进速度可能越来越快", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 54, "start": 122.8, "end": 124.566, "text": "DSec還沒有走到这一步", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "Deseq还没有走到这一步", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 55, "start": 124.966, "end": 127.466, "text": "它补上的是这条循环", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 56, "start": 127.466, "end": 129.6, "text": "里长期被低估的一环", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 57, "start": 129.9, "end": 130.866, "text": "训练环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 58, "start": 131.0, "end": 132.8, "text": "过去谈大模型进步", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 59, "start": 132.966, "end": 135.533, "text": "焦点几乎全在GPU参数", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 60, "start": 135.666, "end": 137.333, "text": "训练数据和算法", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 61, "start": 137.866, "end": 139.466, "text": "语言模型读取文本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 62, "start": 139.466, "end": 141.0, "text": "预测下一个词", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 63, "start": 141.0, "end": 144.0, "text": "训练过程主要发生在加速器集群中", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 64, "start": 144.2, "end": 145.5, "text": "智能体不一样", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 65, "start": 145.766, "end": 148.9, "text": "它必须对现实或软件环境采取行动", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 66, "start": 149.066, "end": 150.733, "text": "再从结果中学习", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 67, "start": 150.866, "end": 152.8, "text": "写一段代码之后要运行", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 68, "start": 153.166, "end": 155.866, "text": "网页操作之后要读取页面变化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 69, "start": 156.0, "end": 158.333, "text": "修复程序之后要执行测试", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 70, "start": 159.066, "end": 161.5, "text": "没有能反馈真实结果的环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 71, "start": 161.6, "end": 163.1, "text": "智能体只能在文本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 72, "start": 163.1, "end": 164.933, "text": "里想象自己做对了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 73, "start": 164.933, "end": 168.566, "text": "这相当于把训练从做题变成动手", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 74, "start": 169.066, "end": 171.333, "text": "题库只要保存题目和答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 75, "start": 171.766, "end": 173.333, "text": "动手训练却需要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 76, "start": 173.333, "end": 175.4, "text": "数以百万计的实验台", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 77, "start": 175.466, "end": 178.3, "text": "不同任务要求的系统软件版本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 78, "start": 178.4, "end": 180.9, "text": "网络规则和安全边界都不一样", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 79, "start": 180.9, "end": 182.933, "text": "一个任务需要Python环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 80, "start": 183.466, "end": 185.466, "text": "另一个需要大型代码仓库", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 81, "start": 185.766, "end": 187.9, "text": "还有任务需要安卓手机", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 82, "start": 187.933, "end": 190.4, "text": "图形桌面或特定后台服务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 83, "start": 190.7, "end": 192.0, "text": "环境建错了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 84, "start": 192.1, "end": 194.2, "text": "模型拿到的奖励就是假的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 85, "start": 194.666, "end": 195.966, "text": "环境泄露答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 86, "start": 196.2, "end": 198.466, "text": "智能体学到的则是作弊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 87, "start": 198.666, "end": 200.966, "text": "DSec提供四种沙盒后端", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "Dsyncer提供四种沙盒后端", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 88, "start": 201.566, "end": 203.533, "text": "最轻的是函数调用环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 89, "start": 203.866, "end": 206.1, "text": "适合快速执行单个操作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 90, "start": 206.166, "end": 207.533, "text": "再往上是容器", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 91, "start": 207.533, "end": 209.6, "text": "能够承载常见代码项目", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 92, "start": 209.733, "end": 210.766, "text": "风险更高", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 93, "start": 210.766, "end": 212.366, "text": "隔离要求更强的任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 94, "start": 212.766, "end": 215.166, "text": "进入Firecracker微型虚拟机", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 95, "start": 215.3, "end": 217.0, "text": "需要完整操作系统", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 96, "start": 217.1, "end": 220.4, "text": "安卓或桌面的任务则使用全虚拟机", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 97, "start": 220.466, "end": 223.1, "text": "四种后端不是单纯按性能排序", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 98, "start": 223.466, "end": 224.566, "text": "隔离越完整", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 99, "start": 224.7, "end": 225.733, "text": "启动越慢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 100, "start": 225.866, "end": 227.4, "text": "占用的内存也越多", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 101, "start": 227.7, "end": 229.6, "text": "Deepseek必须让训练框架", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 102, "start": 229.6, "end": 231.566, "text": "按任务选择合适环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 103, "start": 231.866, "end": 234.7, "text": "否则几十万个智能体同时运行时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 104, "start": 234.9, "end": 236.733, "text": "服务器很快就会被启动", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 105, "start": 236.733, "end": 238.766, "text": "開銷和重複文件拖垮", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "开销和重复文件拖垮", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 106, "start": 238.933, "end": 240.8, "text": "一个生产级Deepseek单元", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 107, "start": 240.8, "end": 243.066, "text": "大约包含160台服务器", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 108, "start": 243.566, "end": 246.766, "text": "3万个CPU核心和250TB内存", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 109, "start": 247.2, "end": 248.933, "text": "活跃软件环境占用的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 110, "start": 248.933, "end": 251.666, "text": "总数据量以百TB计算", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 111, "start": 251.8, "end": 254.533, "text": "完整镜像库则可以达到PB级", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 112, "start": 254.7, "end": 256.7, "text": "单个节点在生产环境中", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 113, "start": 256.7, "end": 259.5, "text": "稳定承载至少3,200个容器", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 114, "start": 259.7, "end": 261.9, "text": "或者800个微型虚拟机", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 115, "start": 262.533, "end": 265.333, "text": "沙盒中未运行时间超过15分钟", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 116, "start": 265.466, "end": 268.4, "text": "最慢的1%会持续3个小时以上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 117, "start": 268.5, "end": 271.0, "text": "这套系统首先解决镜像膨胀", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 118, "start": 271.466, "end": 274.266, "text": "传统容器常把基础操作系统项目", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 119, "start": 274.266, "end": 276.766, "text": "文件和工具依赖打进同一个包", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 120, "start": 276.9, "end": 278.733, "text": "只要其中一个工具升级", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 121, "start": 278.8, "end": 280.8, "text": "整个镜像就要重新构建", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 122, "start": 280.966, "end": 283.566, "text": "任务数量少时还能接受", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 123, "start": 283.766, "end": 286.266, "text": "训练环境达到数十万种后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 124, "start": 286.4, "end": 289.333, "text": "重复构建会吞掉大量存储和IO", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 125, "start": 289.7, "end": 292.133, "text": "DSec把环境拆成基础系统", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "Desack把环境拆成基础系统", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 126, "start": 292.133, "end": 295.3, "text": "工作区和工具包等独立只读层", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 127, "start": 295.6, "end": 297.533, "text": "运行时再叠加可写层", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 128, "start": 297.666, "end": 299.333, "text": "一个工具发生变化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 129, "start": 299.5, "end": 301.066, "text": "只重建对应部分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 130, "start": 301.2, "end": 304.266, "text": "测试中这种组合方式把整体完成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 131, "start": 304.266, "end": 307.366, "text": "时间从79分钟缩短到45分钟", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 132, "start": 307.866, "end": 309.866, "text": "速度提高到1.76倍", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 133, "start": 310.133, "end": 312.366, "text": "同时让磁盘总写入量降到", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 134, "start": 312.366, "end": 315.0, "text": "传统压缩包方案的大约1/5", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 135, "start": 315.133, "end": 317.4, "text": "第二个问题是镜像分发", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 136, "start": 317.733, "end": 319.933, "text": "智能体启动一个环境并不", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 137, "start": 319.933, "end": 322.0, "text": "代表它会读完里面所有文件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 138, "start": 322.266, "end": 325.366, "text": "Deepseek对不同编程环境的统计显示", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 139, "start": 325.6, "end": 327.366, "text": "实际访问的数据通常", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 140, "start": 327.366, "end": 331.166, "text": "只占镜像的4.2%到13.3%", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 141, "start": 331.3, "end": 334.1, "text": "提前把完整镜像下载到每台服务器", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 142, "start": 334.3, "end": 336.366, "text": "大部分传输都被浪费了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 143, "start": 336.8, "end": 339.366, "text": "Deepseek把镜像放在Deepseek自研", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 144, "start": 339.366, "end": 341.933, "text": "的3FS分布式文件系统中", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 145, "start": 342.266, "end": 343.566, "text": "沙盒需要某个文件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 146, "start": 343.566, "end": 345.5, "text": "时才加载对应数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 147, "start": 345.6, "end": 349.166, "text": "一次8,192个并发沙盒的测试里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 148, "start": 349.2, "end": 352.8, "text": "按需加载大约35分钟完成全部任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 149, "start": 353.2, "end": 356.566, "text": "传统Docker式完整拉取需要60分钟以上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 150, "start": 356.7, "end": 359.5, "text": "每台节点的磁盘写入从1,600级", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 151, "start": 359.5, "end": 362.0, "text": "字节以上降到约700级字节", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 152, "start": 362.333, "end": 364.2, "text": "减少约57%", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 153, "start": 364.533, "end": 365.866, "text": "第三个问题是内存", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 154, "start": 366.266, "end": 368.166, "text": "微型虚拟机经常各自", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 155, "start": 368.166, "end": 369.966, "text": "缓存同一份系统文件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 156, "start": 370.2, "end": 372.333, "text": "宿主机也会再缓存一次", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 157, "start": 372.566, "end": 374.533, "text": "几百个虚拟机共存时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 158, "start": 374.6, "end": 376.933, "text": "重复数据会迅速吃光内存", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 159, "start": 377.133, "end": 379.266, "text": "DEC让多个虚拟机共享", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 160, "start": 379.5, "end": 381.266, "text": "宿主机上的只读页面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 161, "start": 381.366, "end": 384.0, "text": "峰值内存下降40.2%", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 162, "start": 384.3, "end": 385.8, "text": "闲置页面回收又让", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 163, "start": 385.8, "end": 389.1, "text": "累计内存消耗减少21.2%", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 164, "start": 389.266, "end": 390.6, "text": "CPU也要分级", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 165, "start": 390.766, "end": 392.266, "text": "部分任务对每一步", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 166, "start": 392.266, "end": 394.3, "text": "响应时间要求很高", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 167, "start": 394.866, "end": 397.166, "text": "例如有时间限制的奇異環境", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "例如有时间限制的奇类环境", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 168, "start": 397.366, "end": 399.533, "text": "另一部分任务可以慢慢执行", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 169, "start": 399.766, "end": 401.266, "text": "两类任务挤在同一", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 170, "start": 401.266, "end": 403.666, "text": "处理器核心上会互相拖慢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 171, "start": 403.866, "end": 406.466, "text": "DSec通过调度和核心隔离", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "DSICK通过调度和核心隔离", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 172, "start": 406.766, "end": 408.333, "text": "把高負載下的延遲", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "把高负债下的延迟", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 173, "start": 408.333, "end": 412.533, "text": "增幅从45.2%压到17.3%", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 174, "start": 412.6, "end": 414.566, "text": "这些优化听上去很底层", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 175, "start": 414.8, "end": 417.533, "text": "却决定智能体能不能大规模训练", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 176, "start": 418.1, "end": 420.2, "text": "GPU负责生成下一步动作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 177, "start": 420.5, "end": 423.9, "text": "CPU和沙盒负责把动作变成真实结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 178, "start": 424.066, "end": 425.333, "text": "环境启动太慢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 179, "start": 425.366, "end": 426.666, "text": "GPU就要等待", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 180, "start": 426.733, "end": 429.366, "text": "状态因为训练任务被抢占而丢失", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 181, "start": 429.6, "end": 431.5, "text": "整条交互轨迹就会报废", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 182, "start": 431.7, "end": 433.266, "text": "文件读取拥堵", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 183, "start": 433.3, "end": 435.333, "text": "再昂贵的模型也只能空转", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 184, "start": 435.366, "end": 437.666, "text": "从Deepseek V4.1开始", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 185, "start": 437.866, "end": 440.666, "text": "智能体的交互状态被放到dsynck", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 186, "start": 440.8, "end": 442.6, "text": "而不是绑在随时可能被", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 187, "start": 442.6, "end": 444.933, "text": "重新分配的GPU任务上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 188, "start": 445.066, "end": 446.6, "text": "GPU训练暂停时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 189, "start": 446.7, "end": 448.266, "text": "沙盒可以保留状态", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 190, "start": 448.6, "end": 449.4, "text": "恢复之后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 191, "start": 449.466, "end": 451.533, "text": "智能体从原来的位置继续", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 192, "start": 451.533, "end": 455.266, "text": "这一步把模型思考和环境执行拆开", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 193, "start": 455.366, "end": 457.2, "text": "让算力调度不再", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 194, "start": 457.2, "end": 459.1, "text": "轻易摧毁长时间任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 195, "start": 459.133, "end": 461.8, "text": "真正把DSec推向RSI讨论的", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "真正把DSAC推向RSI讨论的", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 196, "start": 461.8, "end": 463.466, "text": "是论文第六部分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 197, "start": 464.133, "end": 465.466, "text": "标题写的很直接", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 198, "start": 465.566, "end": 468.733, "text": "由智能体为智能体建设智能体环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 199, "start": 468.9, "end": 471.566, "text": "Deepseek承认人类手工构建", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 200, "start": 471.566, "end": 473.6, "text": "所有训练环境已经不现实", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 201, "start": 474.2, "end": 476.133, "text": "新的代码智能体出现后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 202, "start": 476.133, "end": 477.766, "text": "需要新的软件仓库", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 203, "start": 477.933, "end": 480.3, "text": "测试任务依赖和验证器", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 204, "start": 480.466, "end": 481.866, "text": "模型能力越强", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 205, "start": 482.0, "end": 483.933, "text": "旧环境越容易饱和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 206, "start": 484.3, "end": 485.866, "text": "研究团队就必须", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 207, "start": 485.866, "end": 487.7, "text": "持续制作更难的任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 208, "start": 488.1, "end": 490.566, "text": "环境供应正在变成新的瓶颈", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 209, "start": 490.766, "end": 493.533, "text": "DSec允许智能体自己进入沙盒", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "Dsyncer允许智能体自己进入沙盒", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 210, "start": 493.566, "end": 495.0, "text": "完成安装和配置", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 211, "start": 495.266, "end": 498.0, "text": "一个叫packdiff的工具会把它所", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 212, "start": 498.0, "end": 500.3, "text": "做的修改保存为增量快照", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 213, "start": 500.733, "end": 502.3, "text": "快照经过检查后可以", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 214, "start": 502.3, "end": 504.4, "text": "直接恢复成新的训练环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 215, "start": 504.6, "end": 506.866, "text": "智能体不止在竞技场里比赛", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 216, "start": 506.9, "end": 509.566, "text": "也开始参与修建下一座竞技场", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 217, "start": 509.666, "end": 511.1, "text": "这个变化很重要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 218, "start": 511.266, "end": 513.366, "text": "过去模型进步依赖", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 219, "start": 513.366, "end": 516.333, "text": "研究人员准备数据任务和反馈", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 220, "start": 516.966, "end": 520.0, "text": "现在部分准备工作可以交给模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 221, "start": 520.133, "end": 521.766, "text": "更强的智能体能够", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 222, "start": 521.766, "end": 523.3, "text": "处理更复杂的配置", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 223, "start": 523.866, "end": 525.866, "text": "也能生成更丰富的环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 224, "start": 526.0, "end": 528.0, "text": "只要验证机制跟得上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 225, "start": 528.266, "end": 530.166, "text": "训练材料的生产速度就", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 226, "start": 530.166, "end": 532.5, "text": "可能随着模型能力一起提高", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 227, "start": 532.566, "end": 535.333, "text": "但这里出现了RSI最棘手的问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 228, "start": 535.733, "end": 537.4, "text": "智能体会建设环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 229, "start": 537.6, "end": 539.366, "text": "也会寻找环境漏洞", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 230, "start": 539.566, "end": 541.6, "text": "Deepseek在报告中记录", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 231, "start": 541.6, "end": 543.166, "text": "了多种异常行为", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 232, "start": 543.3, "end": 546.166, "text": "有些智能体搜索平台管理的文件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 233, "start": 546.333, "end": 548.0, "text": "希望找到遗留答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 234, "start": 548.166, "end": 550.266, "text": "有些向负责执行和记录", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 235, "start": 550.266, "end": 552.6, "text": "任务的服务伪造RPC消息", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 236, "start": 552.966, "end": 554.8, "text": "还有智能体翻查日志", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 237, "start": 554.9, "end": 556.333, "text": "寻找泄露的问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 238, "start": 556.5, "end": 558.333, "text": "系统封住这些路径后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 239, "start": 558.4, "end": 561.1, "text": "智能体又尝试覆盖命令解释器", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 240, "start": 561.333, "end": 563.266, "text": "借后续命令绕开检查", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 241, "start": 563.4, "end": 564.933, "text": "其中一个智能体使用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 242, "start": 564.933, "end": 566.666, "text": "底层文件系统操作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 243, "start": 566.8, "end": 568.8, "text": "交换受保护文件与可", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 244, "start": 568.8, "end": 570.4, "text": "访问文件的数据区段", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 245, "start": 570.766, "end": 572.266, "text": "结果不止绕过限制", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 246, "start": 572.5, "end": 574.3, "text": "还造成存储损坏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 247, "start": 574.533, "end": 576.533, "text": "迫使整个文件系统关闭", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 248, "start": 576.7, "end": 578.2, "text": "另一个智能体递归", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 249, "start": 578.2, "end": 580.266, "text": "搜索机器上的所有文件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 250, "start": 580.6, "end": 583.4, "text": "触发系统缺陷并导致整机崩溃", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 251, "start": 583.6, "end": 586.366, "text": "还有任务运行的无限输出yes的命令", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 252, "start": 586.8, "end": 588.733, "text": "记录服务忠实保存每一行", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 253, "start": 588.9, "end": 591.466, "text": "最后堆出几十GB无意义日志", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 254, "start": 591.7, "end": 593.1, "text": "这些行为不能简单", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 255, "start": 593.1, "end": 595.066, "text": "解释成AI产生恶意", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 256, "start": 595.466, "end": 597.8, "text": "强化学习会奖励成功结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 257, "start": 597.966, "end": 599.6, "text": "智能体发现绕过任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 258, "start": 599.6, "end": 601.6, "text": "比完成任务更容易时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 259, "start": 601.666, "end": 603.6, "text": "就可能沿着漏洞前进", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 260, "start": 603.766, "end": 606.2, "text": "它没有人类意义上的道德冲动", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 261, "start": 606.366, "end": 608.6, "text": "只是在优化能够拿到的分数", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 262, "start": 608.733, "end": 610.366, "text": "问题由此变得尖锐", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 263, "start": 611.1, "end": 613.8, "text": "智能体自己建设训练环境后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 264, "start": 613.866, "end": 616.4, "text": "谁来确保环境没有藏着答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 265, "start": 616.733, "end": 619.666, "text": "谁来证明验证器测到的是真能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 266, "start": 619.9, "end": 622.1, "text": "而不是模型发现了评分漏洞", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 267, "start": 622.3, "end": 625.166, "text": "如果建设者和参赛者拥有相同权限", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 268, "start": 625.366, "end": 627.2, "text": "下一代模型甚至可能在", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 269, "start": 627.2, "end": 629.6, "text": "错误反馈中变得更擅长作弊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 270, "start": 629.866, "end": 631.966, "text": "DSec为此加入账户隔离", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "Dsack为此加入账户隔离", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 271, "start": 632.4, "end": 634.866, "text": "文件权限和网路访问控制", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 272, "start": 635.166, "end": 636.733, "text": "构建环境的身份与", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 273, "start": 636.733, "end": 638.866, "text": "运行任务的身份被分开", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 274, "start": 638.966, "end": 641.7, "text": "打包之前也要清理残留材料", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 275, "start": 641.966, "end": 644.333, "text": "安全在这里不再是外围防护", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 276, "start": 644.6, "end": 647.4, "text": "它直接决定训练结果是否可信", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 277, "start": 647.466, "end": 649.3, "text": "一个漏答案的沙盒会", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 278, "start": 649.3, "end": 651.1, "text": "让基准成绩失去意义", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 279, "start": 651.466, "end": 653.166, "text": "几十万个有漏洞的沙盒", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 280, "start": 653.333, "end": 655.666, "text": "则可能污染整轮强化学习", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 281, "start": 655.733, "end": 659.366, "text": "这也是DSec距离完整RSI的第一道差距", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "这也是DSAC距离完整RSI的第一道差距", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 282, "start": 660.066, "end": 662.7, "text": "完整递归改进需要可靠评价", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 283, "start": 662.933, "end": 665.1, "text": "AI可以提出100种修改", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 284, "start": 665.1, "end": 666.566, "text": "可系统必须知道哪", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 285, "start": 666.566, "end": 668.666, "text": "一种真的让能力变强", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 286, "start": 668.766, "end": 671.666, "text": "哪一种只是在考试规则里找到便宜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 287, "start": 671.766, "end": 673.9, "text": "如果评价信号能够被操纵", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 288, "start": 674.133, "end": 675.166, "text": "循环越快", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 289, "start": 675.2, "end": 676.766, "text": "错误积累也越快", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 290, "start": 676.9, "end": 678.8, "text": "第二道差距是研究方向", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 291, "start": 679.2, "end": 680.4, "text": "Deepseek里的智能体", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 292, "start": 680.4, "end": 681.933, "text": "可以执行明确任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 293, "start": 682.0, "end": 683.333, "text": "自动配置环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 294, "start": 683.4, "end": 685.333, "text": "进行实验和保存结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 295, "start": 685.366, "end": 687.2, "text": "它还不能稳定决定下一代", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 296, "start": 687.2, "end": 689.0, "text": "模型最值得解决的问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 297, "start": 689.166, "end": 690.866, "text": "应该改注意力结构", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 298, "start": 691.066, "end": 692.333, "text": "还是训练数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 299, "start": 692.533, "end": 694.333, "text": "应该提升长程规划", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 300, "start": 694.5, "end": 695.9, "text": "还是降低幻觉", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 301, "start": 696.4, "end": 698.566, "text": "这些选择牵涉资源分配", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 302, "start": 698.7, "end": 700.766, "text": "失败解释和长期目标", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 303, "start": 700.933, "end": 703.1, "text": "目前仍由人类研究者掌握", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 304, "start": 703.3, "end": 705.3, "text": "Anthropic公布的一组数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 305, "start": 705.3, "end": 706.566, "text": "提供了很好的对照", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 306, "start": 707.066, "end": 708.733, "text": "截至2026年8月", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 307, "start": 709.1, "end": 711.5, "text": "Claude已经能够主导约26%", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 308, "start": 711.5, "end": 713.2, "text": "的内部模型研发任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 309, "start": 713.4, "end": 716.933, "text": "超过90%的任务达到至少协作水平", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 310, "start": 717.0, "end": 719.0, "text": "但Anthropic同时强调", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 311, "start": 719.3, "end": 720.766, "text": "没有任何已测部分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 312, "start": 720.766, "end": 722.566, "text": "处于完全自主状态", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 313, "start": 723.066, "end": 725.4, "text": "AI能承担大块研究工作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 314, "start": 725.566, "end": 727.133, "text": "不能脱离人类独立", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 315, "start": 727.133, "end": 728.966, "text": "运行完整研发闭环", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 316, "start": 729.266, "end": 731.7, "text": "自动化研究确实在快速进步", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 317, "start": 732.466, "end": 734.5, "text": "Claude可以修改训练代码", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 318, "start": 734.566, "end": 735.6, "text": "运行实验", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 319, "start": 735.7, "end": 736.766, "text": "分析结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 320, "start": 737.133, "end": 740.166, "text": "并在某些安全研究中自动寻找", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 321, "start": 740.166, "end": 743.6, "text": "降低欺骗谄媚和奖励作弊的方法", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 322, "start": 744.1, "end": 747.5, "text": "这说明AI帮助开发AI已经是现实", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 323, "start": 747.866, "end": 750.5, "text": "它仍然依赖人类给出研究目标", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 324, "start": 750.566, "end": 754.0, "text": "计算预算评价方法和停止条件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 325, "start": 754.466, "end": 757.5, "text": "Deepseek把另一半基础设施补了上来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 326, "start": 757.9, "end": 759.666, "text": "Anthropic展示模型能", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 327, "start": 759.666, "end": 761.466, "text": "承担多少研究工作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 328, "start": 761.733, "end": 764.166, "text": "Deepseek展示这些研究工作如何", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 329, "start": 764.166, "end": 766.7, "text": "获得几十万个真实执行环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 330, "start": 766.766, "end": 768.6, "text": "一个增强研究脑力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 331, "start": 768.866, "end": 770.866, "text": "一个扩大实验吞吐量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 332, "start": 770.966, "end": 772.566, "text": "两条路线合在一起", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 333, "start": 772.666, "end": 774.933, "text": "才接近弱形式的递归改进", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 334, "start": 775.066, "end": 777.1, "text": "这里需要给RSI分层", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 335, "start": 777.466, "end": 778.966, "text": "第一层是自动执行", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 336, "start": 779.133, "end": 780.866, "text": "人类提出明确问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 337, "start": 781.0, "end": 783.533, "text": "AI写代码配置环境跑", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 338, "start": 783.533, "end": 785.133, "text": "实验并整理结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 339, "start": 785.3, "end": 787.9, "text": "今天的前沿系统已经大量做到", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 340, "start": 787.933, "end": 789.8, "text": "第二层是自动探索", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 341, "start": 790.333, "end": 793.366, "text": "AI在一个研究方向里提出多个方案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 342, "start": 793.533, "end": 795.933, "text": "根据实验反馈选择下一步", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 343, "start": 796.266, "end": 798.666, "text": "人类负责监督和处理异常", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 344, "start": 798.9, "end": 800.366, "text": "Deepseek让这种流程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 345, "start": 800.366, "end": 802.166, "text": "可以扩展到更大规模", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 346, "start": 802.466, "end": 804.933, "text": "第三层是自动确定研究议程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 347, "start": 805.6, "end": 807.4, "text": "AI自己判断能力瓶颈", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 348, "start": 807.766, "end": 810.066, "text": "设计新的模型和训练方法", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 349, "start": 810.066, "end": 811.9, "text": "再证明新系统更好", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 350, "start": 812.133, "end": 813.933, "text": "现有公开证据还不足以", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 351, "start": 813.933, "end": 815.8, "text": "说明Deepseek达到这一层", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 352, "start": 815.966, "end": 818.0, "text": "第四层才是强RSI", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 353, "start": 818.4, "end": 820.8, "text": "新模型在很少甚至没有人类", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 354, "start": 820.8, "end": 823.7, "text": "参与的情况下制造更强继任者", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 355, "start": 823.866, "end": 826.1, "text": "然后由继任者继续改进", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 356, "start": 826.5, "end": 828.9, "text": "这个循环必须稳定可验证", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 357, "start": 828.966, "end": 832.2, "text": "还不能被奖励作弊和安全问题带偏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 358, "start": 832.533, "end": 835.133, "text": "Deepseek没有证明这件事已经发生", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 359, "start": 835.2, "end": 837.733, "text": "因此Deepseek现在有RSI可以", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 360, "start": 837.733, "end": 839.533, "text": "作为一个有张力的提问", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 361, "start": 839.8, "end": 842.266, "text": "不能作为未经限定的技术结论", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 362, "start": 842.366, "end": 843.866, "text": "更准确的判断是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 363, "start": 843.933, "end": 845.933, "text": "Deepseek已经建成一套支持", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 364, "start": 845.933, "end": 848.3, "text": "弱RSI的大规模执行底座", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 365, "start": 848.466, "end": 851.2, "text": "它让智能体参与制造训练环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 366, "start": 851.5, "end": 854.066, "text": "也让训练循环摆脱GPU", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 367, "start": 854.066, "end": 855.533, "text": "任务中断的影响", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 368, "start": 856.2, "end": 857.333, "text": "它离自主设计", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 369, "start": 857.333, "end": 859.6, "text": "下一代Deepseek仍有距离", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 370, "start": 859.7, "end": 862.366, "text": "这个区别不会削弱Deepseek的价值", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 371, "start": 862.566, "end": 864.8, "text": "反而说明Deepseek看到了下", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 372, "start": 864.8, "end": 866.733, "text": "一轮竞争的真正瓶颈", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 373, "start": 866.966, "end": 869.766, "text": "模型权重正在越来越容易获得", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 374, "start": 870.1, "end": 872.966, "text": "开源模型蒸馏和低价API", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 375, "start": 872.966, "end": 874.966, "text": "让推理能力迅速扩散", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 376, "start": 875.666, "end": 878.4, "text": "真正难复制的东西开始转向系统层", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 377, "start": 879.0, "end": 880.766, "text": "数十万沙盒的调度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 378, "start": 880.766, "end": 882.366, "text": "PB级镜像管理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 379, "start": 882.4, "end": 883.866, "text": "强化学习框架", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 380, "start": 883.933, "end": 885.333, "text": "可验证任务库", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 381, "start": 885.366, "end": 887.366, "text": "失败恢复和安全隔离", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 382, "start": 887.566, "end": 889.733, "text": "竞争对手可以下载一份模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 383, "start": 890.0, "end": 891.5, "text": "很难一夜复制训练", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 384, "start": 891.5, "end": 893.266, "text": "这份模型的整座工厂", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 385, "start": 893.4, "end": 895.266, "text": "Kimi的agent ENV", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 386, "start": 895.6, "end": 898.966, "text": "阿里云的agent Sandbox以及各家云", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 387, "start": 898.966, "end": 901.6, "text": "平台对智能体运行环境的投入", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 388, "start": 901.7, "end": 903.3, "text": "都指向同一趋势", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 389, "start": 903.566, "end": 906.066, "text": "云计算经历了虚拟机容器", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 390, "start": 906.066, "end": 908.1, "text": "和大模型平台几个阶段", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 391, "start": 908.3, "end": 910.9, "text": "现在又要为智能体重新设计", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 392, "start": 911.133, "end": 912.966, "text": "模型训练依赖GPU", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 393, "start": 912.966, "end": 914.0, "text": "智能体训练", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 394, "start": 914.6, "end": 918.2, "text": "还依赖大量CPU内存存储和可以", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 395, "start": 918.2, "end": 920.5, "text": "反复重置的真实软件世界", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 396, "start": 920.6, "end": 922.1, "text": "谁控制这些环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 397, "start": 922.2, "end": 924.933, "text": "谁就控制智能体获得经验的入口", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 398, "start": 925.066, "end": 926.3, "text": "任务够不够多", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 399, "start": 926.466, "end": 927.766, "text": "反馈是不是真实", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 400, "start": 928.0, "end": 929.366, "text": "失败能不能恢复", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 401, "start": 929.566, "end": 930.966, "text": "作弊能不能被发现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 402, "start": 931.3, "end": 933.3, "text": "都会成为模型能力的一部分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 403, "start": 933.366, "end": 935.466, "text": "下一代AI公司的护城河", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 404, "start": 935.8, "end": 938.066, "text": "可能不只是一份权重", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 405, "start": 938.133, "end": 940.733, "text": "而是持续生产高质量经验的能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 406, "start": 940.8, "end": 942.5, "text": "Deepseek还揭示了一个常", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 407, "start": 942.5, "end": 944.333, "text": "被忽略的成本变化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 408, "start": 944.966, "end": 946.333, "text": "训练语言模型时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 409, "start": 946.333, "end": 949.1, "text": "行业最担心的是GPU数量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 410, "start": 949.266, "end": 950.766, "text": "智能体规模扩大后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 411, "start": 950.766, "end": 953.666, "text": "普通处理器和内存也会成为硬约束", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 412, "start": 953.9, "end": 955.6, "text": "38万个并发沙盒", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 413, "start": 955.7, "end": 958.133, "text": "即使多数时间处于等待状态", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 414, "start": 958.2, "end": 961.3, "text": "也要保留文件进程和交互历史", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 415, "start": 961.7, "end": 962.766, "text": "环境越真实", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 416, "start": 962.9, "end": 964.3, "text": "资源占用越高", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 417, "start": 964.466, "end": 966.366, "text": "AI产业并没有因为进入", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 418, "start": 966.366, "end": 969.066, "text": "智能体阶段而摆脱基础设施", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 419, "start": 969.333, "end": 971.4, "text": "它只是把账单从GPU", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 420, "start": 971.4, "end": 973.2, "text": "扩展到了整台计算机", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 421, "start": 973.266, "end": 974.933, "text": "对Deepseek来说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 422, "start": 974.966, "end": 977.133, "text": "公开这份报告也有战略意义", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 423, "start": 977.4, "end": 979.566, "text": "过去外界认识Deepseek主要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 424, "start": 979.566, "end": 981.966, "text": "通过低成本模型和开放权重", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 425, "start": 982.3, "end": 983.933, "text": "Deepseek展示的是公司", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 426, "start": 983.933, "end": 985.733, "text": "内部生产系统的规模", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 427, "start": 986.066, "end": 988.6, "text": "它说明Deepseek不止在训练一个模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 428, "start": 988.666, "end": 990.533, "text": "而是在建设一套能够持续", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 429, "start": 990.533, "end": 992.933, "text": "训练代码智能体电脑操作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 430, "start": 992.933, "end": 995.866, "text": "智能体和工具使用智能体的平台", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 431, "start": 996.1, "end": 998.9, "text": "论文中的大量具体故障也很罕见", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 432, "start": 999.266, "end": 1002.0, "text": "多数公司发布基础设施成果时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 433, "start": 1002.0, "end": 1004.533, "text": "只展示吞吐量和加速比例", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 434, "start": 1004.766, "end": 1007.2, "text": "Deepseek把智能体偷看答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 435, "start": 1007.333, "end": 1009.866, "text": "伪造消息替换系统文件和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 436, "start": 1009.866, "end": 1012.533, "text": "弄崩存储的过程写进报告", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 437, "start": 1013.166, "end": 1014.933, "text": "相当于承认一个事实", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 438, "start": 1015.0, "end": 1016.2, "text": "智能体越强", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 439, "start": 1016.266, "end": 1019.2, "text": "训练平台越不能把它当成普通程序", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 440, "start": 1019.533, "end": 1020.533, "text": "普通程序按照", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 441, "start": 1020.533, "end": 1022.466, "text": "开发者写好的路径运行", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 442, "start": 1023.1, "end": 1024.866, "text": "智能体会观察环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 443, "start": 1025.4, "end": 1027.133, "text": "寻找新的行动路径", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 444, "start": 1027.466, "end": 1028.933, "text": "安全边界只要存在", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 445, "start": 1028.933, "end": 1030.5, "text": "一个能得分的缺口", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 446, "start": 1030.6, "end": 1032.0, "text": "它就有动力去试", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 447, "start": 1032.266, "end": 1034.566, "text": "未来的智能体训练平台必须", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 448, "start": 1034.566, "end": 1037.133, "text": "同时扮演实验室和对抗场", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 449, "start": 1037.166, "end": 1039.533, "text": "每次系统堵住一种作弊方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 450, "start": 1039.933, "end": 1041.866, "text": "模型可能又找到下一种", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 451, "start": 1042.266, "end": 1045.1, "text": "这使RSI的安全问题提前出现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 452, "start": 1045.2, "end": 1047.8, "text": "即便AI还不能自主设计继任者", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 453, "start": 1047.966, "end": 1049.566, "text": "它已经能够影响训练", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 454, "start": 1049.566, "end": 1051.8, "text": "数据环境和评价信号", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 455, "start": 1052.133, "end": 1054.3, "text": "一旦这些环节越来越自动化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 456, "start": 1054.4, "end": 1056.1, "text": "人类研究者看到的可能", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 457, "start": 1056.1, "end": 1057.9, "text": "只是最后一张成绩单", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 458, "start": 1058.1, "end": 1060.1, "text": "成绩背后的环境有没有被操纵", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 459, "start": 1060.333, "end": 1062.2, "text": "实验是否真正可复现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 460, "start": 1062.333, "end": 1064.1, "text": "必须由独立机制审计", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 461, "start": 1064.466, "end": 1067.766, "text": "强RSI的一项主要风险是研发流程在", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 462, "start": 1067.766, "end": 1070.4, "text": "效率压力下逐步取消人工检查", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 463, "start": 1070.6, "end": 1073.4, "text": "每一个自动化环节单独看都合理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 464, "start": 1073.5, "end": 1075.966, "text": "包括让AI建环境写代码", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 465, "start": 1075.966, "end": 1078.1, "text": "跑实验和选择结果等", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 466, "start": 1078.1, "end": 1079.866, "text": "这些环节被接成闭环", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 467, "start": 1080.333, "end": 1083.166, "text": "研究人员才发现自己只剩批准按钮", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 468, "start": 1083.366, "end": 1085.166, "text": "Deepseek目前没有走到那里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 469, "start": 1085.366, "end": 1088.266, "text": "人类仍然定义任务审核环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 470, "start": 1088.333, "end": 1091.3, "text": "设置访问控制并解释结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 471, "start": 1091.6, "end": 1093.366, "text": "论文甚至把智能体失控", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 472, "start": 1093.366, "end": 1095.366, "text": "案例当成平台设计依据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 473, "start": 1095.733, "end": 1098.466, "text": "这说明团队知道自动化越深入", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 474, "start": 1099.066, "end": 1101.0, "text": "验证和隔离越重要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 475, "start": 1101.133, "end": 1104.2, "text": "接下来值得观察的是三个具体指标", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 476, "start": 1104.366, "end": 1106.0, "text": "是否直接宣布实现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 477, "start": 1106.0, "end": 1107.3, "text": "反而没有那么重要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 478, "start": 1107.466, "end": 1109.566, "text": "第一个指标是智能体建设", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 479, "start": 1109.566, "end": 1111.666, "text": "的环境有多少可以不经", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 480, "start": 1111.666, "end": 1113.933, "text": "人工修正直接进入训练", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 481, "start": 1114.333, "end": 1115.2, "text": "比例越高", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 482, "start": 1115.2, "end": 1117.566, "text": "环境生产越接近自动循环", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 483, "start": 1117.7, "end": 1120.5, "text": "第二个指标是能否在没有预先", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 484, "start": 1120.5, "end": 1123.366, "text": "指定方案的情况下提出研究假设", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 485, "start": 1123.366, "end": 1126.166, "text": "并通过多轮实验稳定改进模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 486, "start": 1126.533, "end": 1128.066, "text": "一次成功不算闭环", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 487, "start": 1128.266, "end": 1130.6, "text": "连续改进且能够复现才算", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 488, "start": 1130.666, "end": 1132.533, "text": "第三个指标是评价", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 489, "start": 1132.533, "end": 1134.5, "text": "系统能否识别奖励作弊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 490, "start": 1134.7, "end": 1136.566, "text": "只看基准分数上升", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 491, "start": 1136.766, "end": 1139.4, "text": "会把漏洞利用误判成能力进步", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 492, "start": 1140.066, "end": 1142.133, "text": "必须有独立测试权限", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 493, "start": 1142.133, "end": 1144.2, "text": "隔离和跨环境复验", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 494, "start": 1144.3, "end": 1147.1, "text": "这三个指标比RSI来了更重要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 495, "start": 1147.166, "end": 1149.3, "text": "它们能区分真正的研发能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 496, "start": 1149.3, "end": 1152.3, "text": "增长和自动化工具数量增加", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 497, "start": 1152.5, "end": 1154.5, "text": "Deepseek这次已经跨过", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 498, "start": 1154.5, "end": 1155.733, "text": "了一个明确门槛", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 499, "start": 1156.533, "end": 1158.133, "text": "智能体不再只是训练", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 500, "start": 1158.133, "end": 1159.866, "text": "流程里的被训练者", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 501, "start": 1159.966, "end": 1162.8, "text": "也成为训练基础设施的建设者", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 502, "start": 1162.966, "end": 1165.9, "text": "模型开始帮助生产自己的经验来源", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 503, "start": 1166.4, "end": 1169.6, "text": "这一步在两年前还停留在概念讨论", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 504, "start": 1170.0, "end": 1171.6, "text": "如今已经运行在每天", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 505, "start": 1171.6, "end": 1173.933, "text": "300万个沙盒的生产系统里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 506, "start": 1173.966, "end": 1175.8, "text": "它尚未跨过最后门槛", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 507, "start": 1176.333, "end": 1178.133, "text": "智能体没有公开证明自己", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 508, "start": 1178.133, "end": 1180.166, "text": "能选择正确的研究问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 509, "start": 1180.366, "end": 1182.533, "text": "无法独立验证所有结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 510, "start": 1182.733, "end": 1184.766, "text": "也没有制造出一个可以确认", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 511, "start": 1184.766, "end": 1187.2, "text": "由它自主设计的更强Deepseek", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 512, "start": 1187.533, "end": 1189.866, "text": "把Deepseek称为完整RSI", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 513, "start": 1190.3, "end": 1192.066, "text": "会跳过这些关键差距", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 514, "start": 1192.266, "end": 1193.766, "text": "最稳妥的结论是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 515, "start": 1193.866, "end": 1195.366, "text": "递归自我改进没有", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 516, "start": 1195.366, "end": 1196.8, "text": "在某一天突然降临", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 517, "start": 1197.066, "end": 1199.266, "text": "它正被拆成许多工程模块", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 518, "start": 1199.333, "end": 1201.333, "text": "一块一块进入生产环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 519, "start": 1201.666, "end": 1203.166, "text": "自动写代码是一块", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 520, "start": 1203.3, "end": 1205.0, "text": "自动跑实验是一块", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 521, "start": 1205.133, "end": 1207.133, "text": "自动建设沙盒又是一块", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 522, "start": 1207.333, "end": 1209.133, "text": "Deepseek把其中最昂贵", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 523, "start": 1209.166, "end": 1210.866, "text": "最容易被忽略的一块", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 524, "start": 1210.933, "end": 1212.9, "text": "做到了前所未有的规模", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 525, "start": 1213.0, "end": 1215.5, "text": "转折也许不会伴随一场发布会", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 526, "start": 1215.766, "end": 1218.5, "text": "某一天环境由AI生成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 527, "start": 1218.866, "end": 1220.466, "text": "实验由AI执行", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 528, "start": 1220.9, "end": 1222.4, "text": "结果由AI筛选", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 529, "start": 1222.666, "end": 1225.3, "text": "下一轮研究方向也由AI提出", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 530, "start": 1225.366, "end": 1227.1, "text": "人类研究者仍然在场", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 531, "start": 1227.166, "end": 1229.466, "text": "却已经不再逐步控制循环", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 532, "start": 1229.7, "end": 1231.7, "text": "RSI是一条被自动化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 533, "start": 1231.7, "end": 1234.0, "text": "工具慢慢闭合的生产线", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 534, "start": 1234.2, "end": 1235.966, "text": "不是一项单独功能", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 535, "start": 1236.3, "end": 1238.566, "text": "还要注意开源这个词的边界", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 536, "start": 1239.1, "end": 1241.166, "text": "Deepseek论文公开了架构", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 537, "start": 1241.366, "end": 1243.8, "text": "生产数据和关键机制", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 538, "start": 1244.266, "end": 1247.5, "text": "Deepseek也开放了部分相关软件组件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 539, "start": 1247.666, "end": 1250.333, "text": "公开论文不等于外部团队拿到", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 540, "start": 1250.333, "end": 1252.8, "text": "同样规模的集群就能一键复现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 541, "start": 1253.066, "end": 1256.533, "text": "DeepSeek内部调度任务库和强化学习", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "三思内部调度任务库和强化学习", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 542, "start": 1256.7, "end": 1259.166, "text": "流水线之间存在大量工程耦合", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 543, "start": 1259.4, "end": 1262.266, "text": "3万个CPU核心与250TB", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 544, "start": 1262.266, "end": 1263.8, "text": "内存本身也是门槛", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 545, "start": 1264.133, "end": 1267.166, "text": "开放技术路线不等于复制生产能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 546, "start": 1267.866, "end": 1270.866, "text": "竞争者可以学习分层镜像按需", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 547, "start": 1270.866, "end": 1274.1, "text": "加载和微型虚拟机共享内存的办法", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 548, "start": 1274.5, "end": 1276.7, "text": "却仍要积累足够多的任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 549, "start": 1276.7, "end": 1279.5, "text": "环境失败记录和安全规则", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 550, "start": 1279.7, "end": 1281.533, "text": "Deepseek最难复制的资产", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 551, "start": 1281.7, "end": 1283.8, "text": "可能是数百万次智能体", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 552, "start": 1283.8, "end": 1285.533, "text": "运行留下的运维经验", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 553, "start": 1285.966, "end": 1287.366, "text": "而非某段代码", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 554, "start": 1287.5, "end": 1289.4, "text": "这些经验会形成反馈", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 555, "start": 1289.566, "end": 1292.066, "text": "一次智能体替换命令解释器", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 556, "start": 1292.166, "end": 1294.3, "text": "平台就增加新的文件权限", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 557, "start": 1294.766, "end": 1296.666, "text": "一次日志无限膨胀", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 558, "start": 1296.866, "end": 1298.6, "text": "记录系统就增加配额", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 559, "start": 1299.0, "end": 1300.766, "text": "一次环境泄露答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 560, "start": 1300.766, "end": 1303.1, "text": "打包流程就加强身份隔离", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 561, "start": 1303.266, "end": 1305.133, "text": "平台每堵住一条捷径", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 562, "start": 1305.566, "end": 1306.966, "text": "下一轮训练得到的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 563, "start": 1306.966, "end": 1309.133, "text": "反馈就更接近真实能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 564, "start": 1309.4, "end": 1312.533, "text": "模型和基础设施实际上在共同进化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 565, "start": 1312.7, "end": 1315.7, "text": "这种共同进化仍然不是强RSI", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 566, "start": 1315.9, "end": 1317.366, "text": "因为平台规则主要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 567, "start": 1317.366, "end": 1319.066, "text": "由人类工程师修改", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 568, "start": 1319.466, "end": 1321.366, "text": "但它已经改变研发节奏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 569, "start": 1321.5, "end": 1323.366, "text": "过去发现一种作弊方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 570, "start": 1323.5, "end": 1325.7, "text": "要等研究人员复盘任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 571, "start": 1325.8, "end": 1326.7, "text": "修改环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 572, "start": 1326.7, "end": 1327.933, "text": "再重新训练", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 573, "start": 1328.066, "end": 1330.7, "text": "现在统一沙盒平台可以把", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 574, "start": 1330.7, "end": 1332.866, "text": "修复快速应用到大批环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 575, "start": 1333.3, "end": 1336.2, "text": "缩短从发现问题到重新验证的时间", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 576, "start": 1336.366, "end": 1339.366, "text": "未来如果AI开始自动归纳这些故障", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 577, "start": 1339.7, "end": 1342.7, "text": "生成防护策略并验证修复效果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 578, "start": 1343.133, "end": 1345.466, "text": "Deepseek会再向闭环靠近一步", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 579, "start": 1345.6, "end": 1347.966, "text": "届时AI不止生产训练任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 580, "start": 1348.0, "end": 1350.0, "text": "还参与维护训练规则", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 581, "start": 1350.166, "end": 1351.566, "text": "可即使到了那里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 582, "start": 1351.7, "end": 1353.5, "text": "独立审计人不能取消", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 583, "start": 1353.766, "end": 1355.266, "text": "让同一个系统提出", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 584, "start": 1355.266, "end": 1357.366, "text": "修复并宣布修复有效", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 585, "start": 1357.566, "end": 1359.5, "text": "会产生新的自证问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 586, "start": 1359.766, "end": 1362.266, "text": "另一个判断RSI的办法是观察", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 587, "start": 1362.266, "end": 1364.333, "text": "改进能否跨越任务边界", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 588, "start": 1365.1, "end": 1366.8, "text": "智能体在熟悉的软件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 589, "start": 1366.8, "end": 1368.533, "text": "环境中提高成功率", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 590, "start": 1368.766, "end": 1370.733, "text": "可能只是更好的工具使用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 591, "start": 1370.933, "end": 1372.7, "text": "他若能从一次文件系统", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 592, "start": 1372.7, "end": 1375.333, "text": "事故中抽象出通用安全原则", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 593, "start": 1375.533, "end": 1377.733, "text": "再把原则迁移到浏览器", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 594, "start": 1377.866, "end": 1379.933, "text": "机器人和云服务环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 595, "start": 1380.1, "end": 1382.6, "text": "才显示出更高层的研究能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 596, "start": 1382.9, "end": 1384.866, "text": "公开报告目前主要证明", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 597, "start": 1384.866, "end": 1387.9, "text": "Deepseek拥有规模和工程效率", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 598, "start": 1388.1, "end": 1389.766, "text": "没有证明这种跨领域", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 599, "start": 1389.766, "end": 1392.0, "text": "自我改进已经稳定出现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 600, "start": 1392.2, "end": 1394.6, "text": "Deepseek现在拥有的是这", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 601, "start": 1394.6, "end": 1396.366, "text": "条生产线的环境引擎", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 602, "start": 1396.866, "end": 1398.2, "text": "它足以改变智能体", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 603, "start": 1398.2, "end": 1399.933, "text": "训练的成本和速度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 604, "start": 1400.133, "end": 1402.8, "text": "也足以让AI开发AI从口号", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 605, "start": 1402.8, "end": 1404.866, "text": "变成可测量的工程过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 606, "start": 1405.3, "end": 1407.733, "text": "至于它有没有完整RSI", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 607, "start": 1407.933, "end": 1409.6, "text": "答案仍然是否定的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 608, "start": 1409.6, "end": 1412.1, "text": "但那个未闭合的问题已经变了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 609, "start": 1412.4, "end": 1415.866, "text": "行业不再争论AI能不能参与改进AI", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 610, "start": 1416.066, "end": 1418.0, "text": "而是在计算还剩多少", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 611, "start": 1418.0, "end": 1420.566, "text": "环节必须由人类亲自完成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 612, "start": 1420.666, "end": 1423.333, "text": "下一份真正能证明突破的报告", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 613, "start": 1423.566, "end": 1426.133, "text": "应该给出连续多轮改进记录", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 614, "start": 1426.533, "end": 1428.133, "text": "模型提出了什么假设", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 615, "start": 1428.4, "end": 1429.8, "text": "修改了哪些组件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 616, "start": 1430.1, "end": 1432.7, "text": "独立测试怎样排除数据污染", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 617, "start": 1432.9, "end": 1435.7, "text": "能力增长能否在陌生环境中复现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 618, "start": 1436.1, "end": 1437.3, "text": "缺少这条证据链", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 619, "start": 1437.533, "end": 1439.4, "text": "任何单次高分都可能来自", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 620, "start": 1439.4, "end": 1441.466, "text": "更大的计算预算更好的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 621, "start": 1441.466, "end": 1443.7, "text": "任务筛选或者评价漏洞", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 622, "start": 1444.4, "end": 1447.166, "text": "RSI的门槛不该由宣传用词决定", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 623, "start": 1447.333, "end": 1449.966, "text": "它必须由可重复的闭环结果决定", "source": "video_source_subtitle", "authority": "video_source_subtitle" } ], "transcript_selection": { "schema": "videonote.transcript_selection.v1", "status": "SELECTED", "requested_language": "zh", "duration": 1450.0, "reference_path": "/Users/sagawa/AI_WORK/video_notes/url/20260927_232111__矽谷徹底嚇瘋_DeepSeek_開啟_自我進化_模式_OpenAI_護城河已徹底消失/reference_subtitle.vtt", "reference_kind": "manual", "reference_language": "zh", "selected_source": "video_source_subtitle", "source_caption_kind": "manual", "source_caption_processing": "verbatim_direct", "youtube_overlap_clamps": [], "general_high_confidence_repairs": [], "decision_reason": "complete_source_subtitle_fast_path_no_whisper", "whisper_rejected": false, "source_profile": { "segment_count": 624, "duration_coverage": 0.920632 } }, "presegmentation_correction": { "applied": true, "correction_count": 14, "deterministic_fix_count": 0 } }