{ "text": "google正在补齐\nai\n成为长期研究伙伴所需要的三种能力\nai已经能写代码读论文\n也能在几秒钟内\n给出一套看起来完整的研究方案\n但只要把任务从十分钟拉长到十小时\n问题马上就会暴露\n他会忘记前面验证过什么\n会沿着错误方向反复试探\n会把过时信息当成当前状态\n也会在遇到专业数据\n时突然失去判断力\n模型回答的像专家\n距离真正参与研究\n还有很长的路要走\n仍然隔着一整套运行系统\nGoogle最近公布的三项进展\n刚好对准了这三个缺口\nantigravity teamwork负责组织多个\n智能体长期协作skill state试图\n解决长任务里的状态失控timesfm\n3.0则把基础模型推进到时间序列\n预测这种专业场景三者尚未被\nGoogle宣布为一套已经集成的产品\n但它们放在一起\n透露出的方向很清楚\n下一轮AI竞争重点开始从\n模型一次能达多聪明\n转向系统能不能连续\n几天做对一件复杂的事\n这也是研究伙伴\n四个字真正难的地方\n研究工作很少靠一次灵感直接完成\n一个数学证明可能要\n同时探索十几条路线\n硬件模拟器需要\n反复编译运行和对照\n企业预测还要理解变量之间的关系\n人类团队会分工会质疑会记录失败\n也会在关键节点换方法\n单个聊天窗口没有这些机制\n再长的上下文也只是把\n更多材料塞给同一个决策者\nteamwork先动的是组织方式\n普通的多智能体系统\n通常由一个主智能体拆任务\n再把工作交给几个子智能体\n这个办法处理资料整理\n和代码分工已经够用\n碰到开放研究却\n很容易出现危险局面\n第一个智能体提出了一个方向\n后面的智能体默认它是对的\n所有人开始在同一条错误路线加速\n参与者越多\n错误反而包装的越完整\nteamwork给不同任务设置\n了不同的协作模式\n不能拆开的代码问题进入\n反复测试和修改的循环\n可以拆开的工程任务交给\n多个工作单元并行完成\n再由批评者检查\n数学证明则同时生成多套候选策略\n每套策略旁边都安排\n一个专门找漏洞的角色\n系统还会根据任务\n进展临时增加智能体\n或者改变团队结构\n任务开始时判断简单\n做到一半发现复杂度上升\n它可以继续扩张\n而不是被最初的分工锁死\n这里最值得注意的功能叫long proof\n开放Google数学问题\n经常出现一种情况\n某条路线在前10步都很漂亮\n做到第50步才发现\n一个无法补救的漏洞\n传统智能体可能已经把\n这条路线写进长上下文\n之后不断受到它的影响\nlong proof会让多条路线并行竞争\n每一条都接受反驳者的压力测试\n被否定的路线不会彻底消失\n其中有用的局部想法会留下\n失败原因则进入共享的陷阱记录\n接下来系统把选中的策略拆成\n带有依赖关系的子问题\n能够同时处理的部分并行推进\n必须等待前置结论的\n部分按照顺序执行\n每个子问题内部还会展开一轮候选\n批评和综合\n研究过程由此从一条不断变长\n的思维链变成一张可以检查回\n退和重新分配资源的工作网络\nGoogle声称\n这套方法参与处理了7项数学\n和理论计算机科学开放问题\n涉及稀疏凸优化\n子空间近似向量嵌入\n量化和矩阵分解等方向\n5项结果形成了公开论文\n大部分结论经过人类专家审查\nnew cycles相关证明则\n使用Lin进行形式化验证\nGoogle还报告Gemini3.7Flash和3.1Pro\n配合long proof在TCS bench上达到71%\n此前报告的成绩为67.7%七\n七项成果的验证方式各不相同\n题目数量反而次要\nnutcycles可以把证明翻译进Lin\n让形式系统逐步检查逻辑是否成立\n稀疏凸优化和子空间近似则\n需要领域专家判断问题设定\n证明边界和既有文献之间的关系\nAI能够生成一份很长的证明\n不代表它自动知道自己\n是否遗漏了学界默认条件\n不同研究对象需要不同验收工具\n这也是研究系统无法只靠\n一个总评分解决的原因long proof\n保留失败路线还有\n一个容易被低估的价值\n科研里的失败通常很贵\n同一条思路可能被\n不同研究者重复尝试\n直到很晚才发现障碍\n如果系统能够记录某\n个引理为何不成立\n哪个边界条件造成反例\n下一轮搜索就能\n绕开已经确认的死路\n这个能力不会直接出现在答对率里\n却可能决定系统能否在\n几天后继续有效工作\n但失败记录也会产生新的风险\n一条路线被早期反驳\n不代表它在修改假设后仍然无效\n一个批评智能体给出的否定\n本身也可能是错的\nteamwork需要保留意义\n又不能让错误意义\n永久封死搜索空间\n他公开展示了竞争\n批评和综合的框架\n怎样给不同意见分配权重\n怎样在后续证据出现时恢复旧路线\n仍是决定系统上限的细节\n这些成绩有分量\n也需要正确理解\n71% 是在Google内部测试中的最高层级\n部分研究任务使用的并行\n规模高于公众版本的默认配置\n公开用户可以在付费\nAntigravity计划里调用teamwork预览\n但不能据此认定普通\n账户会复现所有研究结果\n解决7个开放问题同样不该被\n简化成AI独立完成7次数学突破\n题目选择\n资料边界结果验收和最终\n责任仍然掌握在人类专家手中\n它在工程任务上的\n表现更容易看清价值\nGoogle让teamwork从头构建了一个\n乱序执行的risk v周期级模拟器\n这个模拟器可以启动XV6操作系统\n并运行100多项标准基准\n研究团队把它与\nboom处理器的参考结果比较\n在未见过的工作负载上\n平均周期对齐误差为0.71%\n为了防止系统只是在输出\n一个看起来能跑的项目\n开发过程加入了两阶段验证\n还使用与主环境隔离的\nSpike进行持续所部对照\n这类任务不靠一句高质量回答取胜\n一个处理器模拟器涉及取指解码\n乱序调度缓存异常处理和周期精度\n任何一个模块的错误都可能\n在几十万条指令之后才出现\nteamwork能发挥作用\n靠的是让不同角色负责\n实现测试质疑和众合\n并把验证嵌进工作流\n模型本身提供推理能力\n协作机制负责把这种\n能力约束在工程流程里\n更有说服力的证据来自开源项目\nteamwork对Eigen和Polyhash\n提出了性能优化\n部分改动通过正常\n代码审查进入上游项目\n实验报告里的漂亮数字可以\n受到环境和选择方式影响\n外部维护者愿意接收代码\n至少说明改动不只存在于演示页面\n他开始碰到真实\n软件世界最麻烦的部分\n旧代码兼容性测试\n要求和他人的审查意见\nPolyhash的结果还展示了研究\n伙伴与代码助手的差别\nGoogle报告相关改动让64线程下的\n初始插入吞吐量提高约两倍\n单线程整体性能提高约1.5倍\n与flat hash map相比\n内存还减少约25%\n性能优化往往存在交换条件\n一处更快可能换来另一处更慢\n一种数据分布有效\n也可能在另一种负载下失灵\n只有通过基准代码\n审查和维护者判断\n优化才有机会进入真实项目\n这也是teamwork模式比自动生成\n整个项目更值得关注的地方\n研究和工程的价值\n很少由代码行数决定\n系统需要提出假设\n选择测量方法\n识别性能瓶颈\n再接受外部审查\n生成代码只是其中一步\nGoogle把成果送进现有开源社区\n相当于让智能体进入一个\n由人类规则历史兼容性和\n公开讨论组成的环境\n难度远高于在封闭\n演示里让测试全部变绿\nteamwork解决了谁和谁一起做的问题\n长任务还有另一个隐蔽故障\n智能体每执行一步\n通常都会把观察结果工具\n输出和中间推理追加到对话\n10步以后还能管理\n100步以后\n上下文就会变成混合了当前\n事实过期结论和失败尝试的仓库\n模型每次行动之前\n都要重新从这堆\n文字里恢复世界状态\n人类做长期项目时\n不会这样工作\n工程师不会每天把几个月\n的聊天记录从头读一遍\n再猜代码现在编译到哪一步\n团队会维护任务看板\n测试状态\n风险清单和当前分值\n历史记录可以留档\n下一步决策依赖的却是\n一份清楚的当前状态\n把这个尝试放进智能体运行时\n每一步只给模型三类信息\n固定的任务规范当前结构化\n状态以及环境刚刚返回的新观察\n模型完成推理后提出\n状态修改和下一步动作\n运行时先验证修改是否合法\n再写入状态\n那一轮中间推理随后被丢弃\n不再跟着任务无限增长\n这与传统的上下文压缩有根本区别\n摘要方法仍然需要把\n过去重新写成一段文字\n信息可能在压缩时变形\n模型也要判断摘要里\n哪些内容仍然有效skill\nstate直接规定哪些\n字段代表当前世界\n例如安全挑战可以只\n保留已经发现的标记\n测试过的假设\n活跃文件\n当前目录和命令摘要\n下一步行动读取这些字段\n不必回放\n全部终端输出\n论文给出的复杂度变化很直接\n传统追加历史的方式单次\n输入会随着任务长度增长\n累计TOKEN消耗可能\n接近任务步数的平方\n只要状态结构和最新\n观察的大小受到限制\nskill state每一步的提示\n规模可以保持在固定范围\n累计消耗随步数线性增加\n任务从几十步拉到几百步时\n这种差异会迅速放大\n在一项100步的仓库任务里\nGemini3Flash使用普通react\n方法取得84%的成功率\n累计消耗约124万TOKEN\nskill state把成功率提高到94%\n累计TOKEN约6.5万\n到了200步\nreact成功率下降到74%\nskill state仍保持94%\n论文在intercode CTF安全\n任务上的结果也显示\nskill state完成率为54.2%\n它高于几种对照运行方式\n同时把平均提示规模\n和总token压了下来\n更关键的实验是噪声测试\n研究人员向任务环境加入\n与状态无关的遥测信息\n普通提示在噪声变多后明显恶化\n因为无用信息也会一直留在历史里\nskill state只把影响后续\n行动的内容写进状态\n所以成功率保持在接近完整水平\n它展示了一条很实用的原则\n长任务的可靠性不只\n取决于模型能记住多少\n还取决于系统敢不敢\n丢掉已经没用的东西\n这个方案也有明显代价\n状态结构必须由人提前设计\n如果某条早期信息当时看起来无关\n后来才发现它很重要\n而它又没有进入状态字段\n这条线索可能已经丢失\n需要完整审计轨迹\n解释研究过程\n或者回看思路变化的任务\n也不能只保留当前状态\n论文目前主要验证单智能体\n多个智能体同时修改同一份状态时\n会不会覆盖\n冲突又该由谁裁决\n仍然没有充分答案\n论文对小模型的\n失败分析进一步说明\n结构化状态并不会自动带来可靠性\nGemma开放权重模型在100步\n任务中只有42%的得分\n68%的失败来自过早\n覆盖或删除重要状态\n20%属于结构和类型错误\n还有12%是JSON语法问题\n运行时可以拒绝格式错误的更新\n却无法保证模型对什么\n值得保留的判断正确\n状态越短\n写错一个字段造成\n的影响可能越集中\n承受系统因此需要\n同时保存两层记录\n决策层读取精简的当前状态\n审计层保存完整日志\n出现异常时能够追溯\n前者控制成本和注意力\n后者承担调查合规和复现\nskilled state论文强调丢弃中间推理\n适合追求执行效率\n医疗金融和关键基础设施\n可能要求更严格的留痕\n怎样在隐私成本与可追责之间取舍\n会决定它能进入哪些行业\n状态结构的设计也\n可能成为新的专业岗位\n过去的提示工程关注\n怎样把问题问清楚\n长任务系统更需要\n定义任务进行到哪一步\n应该有哪些字段表示\n供应链任务要记录订单\n交付风险和库存约束\n代码任务要记录分支测试和依赖\n科研任务则要记录\n假设证据和未解决问题\n同一个模板无法覆盖所有领域\n领域专家会重新进入系统设计中心\n现有证据还不足以把skill state写成\nteamwork已经采用的底层记忆系统\n风控材料没有给出这种集成关系\n它们分别击中了长期\n智能体的两处故障\nteamwork防止多个智能体集体\n沿着错误方向自我强化\nskill state防止任务\n历史把当前判断淹没\n未来的研究系统很可能\n同时需要这两种能力\nGoogle是否会把它们合并\n仍要看后续产品和论文\n第三项能力来自Timesfm 3.0\n它看起来没有数学证明和\n多智能体团队那么戏剧化\n却可能最接近企业每天\n真正要花钱解决的问题\n大量现实决策建立在时间序列上\n零售商要预测下周每家门店的销量\n电网要估计明天每小时的负荷\n云平台要提前分配容量\n工厂要判断设备何时可能异常\n语言模型擅长处理文字\n时间序列要求模型理解趋势\n周期突变变量联动和外部条件\n一个预测错几个百分点\n可能对应库存积压\n能源浪费或者服务中断\n传统做法往往给不同业务单独建模\n某个商品用一套模型\n另一个地区重新调参数\n遇到节假日促销和\n天气还要增加外生变量\n这个流程需要数据\n科学团队长期维护\n模型迁移到新场景时也容易失效\n时间序列基础模型\n希望学习更通用的规律\n用一个预训练模型\n直接处理大量陌生任务\nTIMESFM3.0大约有3亿参数\n支持单变量和多变量预测\n也能接收协变量\n所谓多变量\n是让模型同时理解\n几个彼此影响的序列\n例如价格销量库存和客流\n协变量则可以分成两类\n一类只有过去数据\n另一类在未来已经知道\n比如节假日安排电价\n计划或者促销日程\n这些信息进入预测后\n模型才能区分正常\n季节变化和真正异常\n它采用Stack mixing Transformer架构\n引入varied attention处理\n不同变量之间的关系\n归一化机制则用来减少\n不同序列尺度带来的干扰\n训练材料包括去除评测\n重叠后的时间序列数据\n也纳入截至2023年11月的\nWikipedia页面访问量\n截至2022年底的Google Trends热门查询\n合成数据和增强数据同样进入训练\n这些细节说明\n所谓通用预测能力依然\n建立在大规模多来源序列上\n时间序列模型最怕数据泄露\n训练集如果包含评测其信息\n或者同一序列的片段\n同时进入训练和测试\n排名会显得异常漂亮\nGoogle特别说明\ngift Evals训练数据排除了\n与Fbench重叠的数据集\n目的就是降低这种问题\n即便如此\n三个基准第一仍然属于发布方报告\n企业部署前还要在自己的\n时间跨度异常条件和业务\n损失函数上重新验证\n预测准确率也不能\n单独决定业务价值\n库存系统更关心缺货和\n积压哪一种错误代价更高\n电网预测必须关注极端峰值\n金融公开则可能在意尾部事件\n一个模型平均误差更低\n未必在最昂贵的错误上表现更好\nTimesfm提供分位数预测\n让使用者看到不同可能区间\n这比只交付一个单\n点数字更适合决策\n但最后仍需业务规则解释风险\nGoogle列出了Fbench time Benchmark\n和JIFF Evals三个主要\n时间序列基础模型评测\n并称TIMESFM3.0均排名第一\nFbench包含100项多样化现实预测任务\ntimebenchmark覆盖50个领域\n数据集和98项评估任务\n这个结果说明基础模型路线正在从\n文本图像继续向企业数字世界扩张\ntimesfm的商业价值不止\n来自一次预测精度\n统一模型可以减少每个部门\n重复训练和维护小模型的成本\n也更适合被智能体调用\n一个负责供应链的智能体\n如果能够读取订单库存和物流数据\n再调用时间序列模型生成预测\n它才有机会提出可执行的采购方案\n缺少这一层专业工具\n通用智能体往往只能根据\n文字描述给出宽泛建议\nGoogle在这个方向还有分发优势\ntimesfm较早版本已经进入\nbigqueryml和googlesheets等产品生态\n企业不需要先搭一套全新\n平台便可以在熟悉的数据\n工具中接触预测能力\nTIMESFM3.0公开了Pytorch权重和使用方式\n但现有资料没有证明3.0权重已经\n同步替换这些产品里的所有版本\n产品整合能走多快\n仍需以Google后续公告为准\n许可证也是一个容易被忽略的限制\ntimesfm代码仓库采用\nApache2.03.0预训练权重\n却使用单独的非商业许可证\n研究者可以下载和实验\n默认权重目前不允许用于商业生产\n把它简单称为完全开源\n企业拿来就能用\n会误导观众\nGoogle开放了能力展示和研究入口\n同时保留了商业化控制权\n把这三项工作放回同一张图\nGoogle的策略开始显形\nteamwork负责扩大搜索空间\n让多个智能体竞争和互相纠错\nskilled state试图控制长\n任务的状态和成本\ntimesfm 3.0提供专业预测能力\n协作状态专业工具\n构成AI参与长期研究和\n工程工作的三根支柱\n这条路线也解释了为什么单纯\n扩大模型参数带来的惊喜正在减少\n模型可以在基准上更聪明\n复杂项目仍然会被组织问题拖垮\n给一个人更高智商\n不代表他自动拥有实验室\n项目管理制度和专业仪器\nAI系统也一样\n模型负责提出和判断\n编排系统负责调度\n状态系统负责记录当前事实\n专业模型负责处理通用\n语言能力覆盖不到的数据\nGoogle在这一轮竞争里的优势\n来自这些组件之间潜在的闭环\n它有Gemini模型\n也有antigravity这个智能体工作界面\ncloud和bigquery可以承接企业数据\n专业能力则覆盖科学代码和预测\n只要这些产品真正连起来\nGoogle出售的将是一套能进入研究\n流程和生产系统的工作环境\n聊天机器人只是其中一个入口\n这个闭环真正形成后\n用户的购买单位也会变化\n企业过去按模型调用\n次数购买智能能力\n未来可能按完成的研究活动\n工程任务或预测工作流付费\n模型选择会退到系统内部\n简单步骤调用Flash\n困难判断切换Pro\n时间序列交给timesfm\n代码验证交给测试工具\n客户关心的是任务有没有完成\n证据是否充分\n成本能否预测\nGoogle已经在antigravity里提供\n动态子智能体异步任务\n定时执行和自定义角色\nteamwork继续向上增加协作模式\n他还把antigravity扩展到\n命令行IDE和企业平台\n每一个入口都在降低智能体\n进入现有工作的迁移成本\n研究人员不必为了使用\n智能体放弃代码仓库\n企业也不必把全部\n数据搬进一个聊天网页\n竞争对手当然会走同一方向\n多智能体编排结构化\n状态和专业模型都没有\n只属于Google的理论壁垒\nGoogle较难复制的部分是搜索\n云办公套件开发工具和自\n研模型形成的分发网络\n技术领先可以被追赶\n把能力送到企业已经\n存在的数据和流程里\n通常需要更长时间\n潜力越大\n成本问题越不能跳过\nteamwork的高质量结果依赖\n并行搜索和多轮验证\n同一个问题同时运行十几条路线\n会消耗更多推理资源Gemini 3.7\nFlash的速度和价格可以降低门槛\n却不会让算力成本消失\n公开预览版本刻意在\n能力和费用之间做平衡\n也说明实验室里最大化智能的配置\n很难直接成为每个用户的默认服务\n企业还要面对另一种成本\nskill点state需要设计状态结构\ntimesfm需要干净且有意义的数据\nteamwork需要明确验收标准\n公司如果连流程指标和\n数据权限都没有理清\n接入更多智能体只会把混乱自动化\n研究伙伴能够加快一套好流程\n也能更快执行一套错误流程\n责任边界同样没有消失\n数学证明可以交给Ling验证\n代码可以用测试和上游审查把关\n时间序列预测却经常\n影响库存能源和资金配置\n模型给出结果以后\n谁确认数据没有偏差\n谁判断异常是否来自制度变化\n谁为错误决策承担后果\n这些问题不能交给\n一个置信度数字解决\nGoogle在teamwork介绍中反复强调\n人类仍然掌握目标和最终验收\n这句话点出了当前系统成立的前提\n智能体可以扩大候选方案数量\n也能更快找漏洞\n却不知道一个组织\n真正愿意承担什么风险\n研究方向的价值证据是否\n足够结果何时可以进入现实\n都需要有人签字\n所以AI成为长期研究\n伙伴的标志会很安静\n他开始连续工作几天\n知道当前项目处在哪一步\n他能主动找人反驳自己的方案\n也会调用合适的专业模型\n避免用语言猜数字\n他留下可检查的结果\n也接受外部验证\n某个模型宣布自己达到专家水平\n证明不了这些能力\nGoogle正在补齐的正是这些基础能力\nteamwork让AI学会组织争论\nskill state让它在长任务中保持清醒\nTimes FM3.0让它进入专业预测\n三项技术还没有合成一个无缝系统\n距离可信赖的研究伙伴也差\n着权限安全责任和成本管理\n但竞争的目标已经变了\n过去的AI比赛看谁能在\n一分钟里给出最惊艳的答案\n接下来的比赛\n要看谁能在100个步骤之后\n仍然知道自己为什么这么做\n能把结果交给专家复核\n还能把整套流程控制\n在可承受的成本里\nGoogle这次释放的信号很明确\n模型负责起步\n产品要让智能长期\n稳定并且专业的工作", "language": "zh", "duration": 1562.0, "segments": [ { "id": 0, "start": 0.0, "end": 1.04, "text": "google正在补齐", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 1, "start": 1.04, "end": 1.7, "text": "ai", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 2, "start": 1.7, "end": 5.14, "text": "成为长期研究伙伴所需要的三种能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 3, "start": 5.24, "end": 7.8, "text": "ai已经能写代码读论文", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 4, "start": 7.8, "end": 9.04, "text": "也能在几秒钟内", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 5, "start": 9.04, "end": 11.64, "text": "给出一套看起来完整的研究方案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 6, "start": 11.64, "end": 15.04, "text": "但只要把任务从十分钟拉长到十小时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 7, "start": 15.1, "end": 16.78, "text": "问题马上就会暴露", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 8, "start": 16.78, "end": 19.24, "text": "他会忘记前面验证过什么", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 9, "start": 19.4, "end": 21.74, "text": "会沿着错误方向反复试探", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 10, "start": 21.94, "end": 24.6, "text": "会把过时信息当成当前状态", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 11, "start": 24.78, "end": 26.3, "text": "也会在遇到专业数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 12, "start": 26.3, "end": 28.28, "text": "时突然失去判断力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 13, "start": 28.5, "end": 30.3, "text": "模型回答的像专家", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 14, "start": 30.5, "end": 31.94, "text": "距离真正参与研究", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 15, "start": 31.94, "end": 32.8, "text": "还有很长的路要走", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 16, "start": 32.8, "end": 34.98, "text": "仍然隔着一整套运行系统", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 17, "start": 35.2, "end": 37.34, "text": "Google最近公布的三项进展", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 18, "start": 37.58, "end": 39.5, "text": "刚好对准了这三个缺口", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 19, "start": 40.14, "end": 42.74, "text": "antigravity teamwork负责组织多个", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 20, "start": 42.74, "end": 45.94, "text": "智能体长期协作skill state试图", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 21, "start": 45.94, "end": 48.98, "text": "解决长任务里的状态失控timesfm", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 22, "start": 48.98, "end": 52.54, "text": "3.0则把基础模型推进到时间序列", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 23, "start": 52.54, "end": 55.5, "text": "预测这种专业场景三者尚未被", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 24, "start": 55.5, "end": 58.38, "text": "Google宣布为一套已经集成的产品", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 25, "start": 58.68, "end": 60.08, "text": "但它们放在一起", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 26, "start": 60.2, "end": 62.08, "text": "透露出的方向很清楚", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 27, "start": 62.2, "end": 65.48, "text": "下一轮AI竞争重点开始从", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 28, "start": 65.48, "end": 67.3, "text": "模型一次能达多聪明", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 29, "start": 67.88, "end": 69.48, "text": "转向系统能不能连续", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 30, "start": 69.48, "end": 71.5, "text": "几天做对一件复杂的事", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 31, "start": 71.6, "end": 72.9, "text": "这也是研究伙伴", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 32, "start": 72.9, "end": 74.74, "text": "四个字真正难的地方", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 33, "start": 75.14, "end": 78.18, "text": "研究工作很少靠一次灵感直接完成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 34, "start": 78.34, "end": 79.8, "text": "一个数学证明可能要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 35, "start": 79.8, "end": 81.7, "text": "同时探索十几条路线", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 36, "start": 82.2, "end": 83.48, "text": "硬件模拟器需要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 37, "start": 83.48, "end": 85.74, "text": "反复编译运行和对照", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 38, "start": 86.18, "end": 89.08, "text": "企业预测还要理解变量之间的关系", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 39, "start": 89.2, "end": 93.1, "text": "人类团队会分工会质疑会记录失败", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 40, "start": 93.28, "end": 95.38, "text": "也会在关键节点换方法", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 41, "start": 95.48, "end": 97.5, "text": "单个聊天窗口没有这些机制", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 42, "start": 98.08, "end": 99.9, "text": "再长的上下文也只是把", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 43, "start": 99.9, "end": 102.18, "text": "更多材料塞给同一个决策者", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 44, "start": 102.4, "end": 104.98, "text": "teamwork先动的是组织方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 45, "start": 105.2, "end": 107.14, "text": "普通的多智能体系统", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 46, "start": 107.28, "end": 109.68, "text": "通常由一个主智能体拆任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 47, "start": 109.8, "end": 112.2, "text": "再把工作交给几个子智能体", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 48, "start": 112.38, "end": 114.14, "text": "这个办法处理资料整理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 49, "start": 114.14, "end": 115.98, "text": "和代码分工已经够用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 50, "start": 116.38, "end": 117.6, "text": "碰到开放研究却", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 51, "start": 117.6, "end": 119.3, "text": "很容易出现危险局面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 52, "start": 119.4, "end": 121.38, "text": "第一个智能体提出了一个方向", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 53, "start": 121.48, "end": 123.78, "text": "后面的智能体默认它是对的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 54, "start": 124.38, "end": 127.0, "text": "所有人开始在同一条错误路线加速", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 55, "start": 127.6, "end": 128.7, "text": "参与者越多", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 56, "start": 128.7, "end": 130.54, "text": "错误反而包装的越完整", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 57, "start": 130.78, "end": 132.48, "text": "teamwork给不同任务设置", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 58, "start": 132.48, "end": 133.94, "text": "了不同的协作模式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 59, "start": 134.48, "end": 136.38, "text": "不能拆开的代码问题进入", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 60, "start": 136.38, "end": 138.4, "text": "反复测试和修改的循环", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 61, "start": 138.6, "end": 140.58, "text": "可以拆开的工程任务交给", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 62, "start": 140.58, "end": 142.74, "text": "多个工作单元并行完成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 63, "start": 143.14, "end": 144.7, "text": "再由批评者检查", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 64, "start": 144.88, "end": 147.7, "text": "数学证明则同时生成多套候选策略", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 65, "start": 147.94, "end": 149.48, "text": "每套策略旁边都安排", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 66, "start": 149.48, "end": 151.3, "text": "一个专门找漏洞的角色", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 67, "start": 151.6, "end": 152.94, "text": "系统还会根据任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 68, "start": 152.94, "end": 155.08, "text": "进展临时增加智能体", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 69, "start": 155.08, "end": 156.8, "text": "或者改变团队结构", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 70, "start": 156.98, "end": 158.88, "text": "任务开始时判断简单", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 71, "start": 158.94, "end": 161.14, "text": "做到一半发现复杂度上升", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 72, "start": 161.54, "end": 162.8, "text": "它可以继续扩张", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 73, "start": 163.1, "end": 165.2, "text": "而不是被最初的分工锁死", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 74, "start": 165.3, "end": 168.0, "text": "这里最值得注意的功能叫long proof", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 75, "start": 168.54, "end": 169.7, "text": "开放Google数学问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 76, "start": 169.7, "end": 171.28, "text": "经常出现一种情况", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 77, "start": 171.5, "end": 174.18, "text": "某条路线在前10步都很漂亮", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 78, "start": 174.3, "end": 176.0, "text": "做到第50步才发现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 79, "start": 176.0, "end": 177.6, "text": "一个无法补救的漏洞", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 80, "start": 177.98, "end": 179.3, "text": "传统智能体可能已经把", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 81, "start": 179.3, "end": 181.48, "text": "这条路线写进长上下文", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 82, "start": 181.6, "end": 183.4, "text": "之后不断受到它的影响", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 83, "start": 183.7, "end": 186.7, "text": "long proof会让多条路线并行竞争", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 84, "start": 186.94, "end": 189.38, "text": "每一条都接受反驳者的压力测试", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 85, "start": 189.58, "end": 192.0, "text": "被否定的路线不会彻底消失", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 86, "start": 192.4, "end": 194.78, "text": "其中有用的局部想法会留下", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 87, "start": 194.94, "end": 198.1, "text": "失败原因则进入共享的陷阱记录", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 88, "start": 198.28, "end": 201.1, "text": "接下来系统把选中的策略拆成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 89, "start": 201.1, "end": 203.08, "text": "带有依赖关系的子问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 90, "start": 203.18, "end": 205.94, "text": "能够同时处理的部分并行推进", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 91, "start": 206.0, "end": 207.74, "text": "必须等待前置结论的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 92, "start": 207.74, "end": 209.54, "text": "部分按照顺序执行", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 93, "start": 209.6, "end": 212.18, "text": "每个子问题内部还会展开一轮候选", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 94, "start": 212.4, "end": 213.48, "text": "批评和综合", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 95, "start": 213.7, "end": 216.5, "text": "研究过程由此从一条不断变长", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 96, "start": 216.5, "end": 220.08, "text": "的思维链变成一张可以检查回", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 97, "start": 220.08, "end": 223.2, "text": "退和重新分配资源的工作网络", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 98, "start": 223.38, "end": 224.38, "text": "Google声称", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 99, "start": 224.78, "end": 227.14, "text": "这套方法参与处理了7项数学", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 100, "start": 227.14, "end": 229.6, "text": "和理论计算机科学开放问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 101, "start": 229.98, "end": 231.58, "text": "涉及稀疏凸优化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 102, "start": 231.9, "end": 234.28, "text": "子空间近似向量嵌入", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 103, "start": 234.34, "end": 236.48, "text": "量化和矩阵分解等方向", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 104, "start": 236.58, "end": 238.98, "text": "5项结果形成了公开论文", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 105, "start": 239.08, "end": 241.74, "text": "大部分结论经过人类专家审查", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 106, "start": 241.98, "end": 244.1, "text": "new cycles相关证明则", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 107, "start": 244.1, "end": 246.38, "text": "使用Lin进行形式化验证", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 108, "start": 246.78, "end": 250.38, "text": "Google还报告Gemini3.7Flash和3.1Pro", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 109, "start": 250.38, "end": 254.74, "text": "配合long proof在TCS bench上达到71%", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 110, "start": 254.8, "end": 258.0, "text": "此前报告的成绩为67.7%七", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 111, "start": 258.0, "end": 260.58, "text": "七项成果的验证方式各不相同", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 112, "start": 260.9, "end": 262.9, "text": "题目数量反而次要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 113, "start": 263.68, "end": 266.48, "text": "nutcycles可以把证明翻译进Lin", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 114, "start": 266.8, "end": 270.0, "text": "让形式系统逐步检查逻辑是否成立", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 115, "start": 270.34, "end": 273.68, "text": "稀疏凸优化和子空间近似则", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 116, "start": 273.68, "end": 276.2, "text": "需要领域专家判断问题设定", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 117, "start": 276.68, "end": 279.58, "text": "证明边界和既有文献之间的关系", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 118, "start": 279.9, "end": 282.3, "text": "AI能够生成一份很长的证明", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 119, "start": 282.3, "end": 283.98, "text": "不代表它自动知道自己", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 120, "start": 283.98, "end": 286.2, "text": "是否遗漏了学界默认条件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 121, "start": 286.8, "end": 289.5, "text": "不同研究对象需要不同验收工具", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 122, "start": 290.0, "end": 292.14, "text": "这也是研究系统无法只靠", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 123, "start": 292.14, "end": 294.94, "text": "一个总评分解决的原因long proof", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 124, "start": 295.58, "end": 297.2, "text": "保留失败路线还有", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 125, "start": 297.2, "end": 298.94, "text": "一个容易被低估的价值", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 126, "start": 299.08, "end": 301.28, "text": "科研里的失败通常很贵", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 127, "start": 301.5, "end": 303.08, "text": "同一条思路可能被", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 128, "start": 303.08, "end": 304.98, "text": "不同研究者重复尝试", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 129, "start": 305.28, "end": 307.18, "text": "直到很晚才发现障碍", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 130, "start": 307.68, "end": 308.9, "text": "如果系统能够记录某", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 131, "start": 308.9, "end": 310.6, "text": "个引理为何不成立", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 132, "start": 310.74, "end": 312.78, "text": "哪个边界条件造成反例", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 133, "start": 312.94, "end": 314.38, "text": "下一轮搜索就能", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 134, "start": 314.38, "end": 316.18, "text": "绕开已经确认的死路", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 135, "start": 316.2, "end": 319.14, "text": "这个能力不会直接出现在答对率里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 136, "start": 319.5, "end": 321.28, "text": "却可能决定系统能否在", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 137, "start": 321.28, "end": 323.34, "text": "几天后继续有效工作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 138, "start": 323.4, "end": 326.0, "text": "但失败记录也会产生新的风险", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 139, "start": 326.14, "end": 328.08, "text": "一条路线被早期反驳", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 140, "start": 328.28, "end": 331.14, "text": "不代表它在修改假设后仍然无效", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 141, "start": 331.28, "end": 333.9, "text": "一个批评智能体给出的否定", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 142, "start": 334.0, "end": 335.7, "text": "本身也可能是错的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 143, "start": 336.0, "end": 337.7, "text": "teamwork需要保留意义", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 144, "start": 337.8, "end": 339.3, "text": "又不能让错误意义", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 145, "start": 339.3, "end": 341.18, "text": "永久封死搜索空间", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 146, "start": 341.3, "end": 342.98, "text": "他公开展示了竞争", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 147, "start": 343.1, "end": 344.68, "text": "批评和综合的框架", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 148, "start": 344.9, "end": 346.98, "text": "怎样给不同意见分配权重", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 149, "start": 347.1, "end": 350.08, "text": "怎样在后续证据出现时恢复旧路线", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 150, "start": 350.14, "end": 352.38, "text": "仍是决定系统上限的细节", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 151, "start": 352.54, "end": 353.98, "text": "这些成绩有分量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 152, "start": 353.98, "end": 355.54, "text": "也需要正确理解", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 153, "start": 355.98, "end": 359.6, "text": "71% 是在Google内部测试中的最高层级", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 154, "start": 359.94, "end": 362.0, "text": "部分研究任务使用的并行", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 155, "start": 362.0, "end": 364.6, "text": "规模高于公众版本的默认配置", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 156, "start": 364.8, "end": 366.34, "text": "公开用户可以在付费", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 157, "start": 366.34, "end": 369.1, "text": "Antigravity计划里调用teamwork预览", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 158, "start": 369.54, "end": 371.38, "text": "但不能据此认定普通", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 159, "start": 371.38, "end": 373.68, "text": "账户会复现所有研究结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 160, "start": 373.9, "end": 376.4, "text": "解决7个开放问题同样不该被", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 161, "start": 376.4, "end": 379.74, "text": "简化成AI独立完成7次数学突破", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 162, "start": 379.8, "end": 380.78, "text": "题目选择", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 163, "start": 381.2, "end": 384.18, "text": "资料边界结果验收和最终", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 164, "start": 384.18, "end": 386.98, "text": "责任仍然掌握在人类专家手中", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 165, "start": 387.08, "end": 388.28, "text": "它在工程任务上的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 166, "start": 388.28, "end": 390.18, "text": "表现更容易看清价值", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 167, "start": 390.54, "end": 393.0, "text": "Google让teamwork从头构建了一个", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 168, "start": 393.0, "end": 395.98, "text": "乱序执行的risk v周期级模拟器", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 169, "start": 396.14, "end": 399.28, "text": "这个模拟器可以启动XV6操作系统", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 170, "start": 399.58, "end": 402.1, "text": "并运行100多项标准基准", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 171, "start": 402.28, "end": 403.68, "text": "研究团队把它与", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 172, "start": 403.68, "end": 405.94, "text": "boom处理器的参考结果比较", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 173, "start": 406.2, "end": 408.48, "text": "在未见过的工作负载上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 174, "start": 408.5, "end": 411.94, "text": "平均周期对齐误差为0.71%", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 175, "start": 411.94, "end": 413.9, "text": "为了防止系统只是在输出", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 176, "start": 413.9, "end": 416.1, "text": "一个看起来能跑的项目", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 177, "start": 416.3, "end": 418.74, "text": "开发过程加入了两阶段验证", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 178, "start": 419.08, "end": 420.88, "text": "还使用与主环境隔离的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 179, "start": 420.88, "end": 423.18, "text": "Spike进行持续所部对照", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 180, "start": 423.24, "end": 426.4, "text": "这类任务不靠一句高质量回答取胜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 181, "start": 427.08, "end": 430.2, "text": "一个处理器模拟器涉及取指解码", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 182, "start": 430.28, "end": 434.08, "text": "乱序调度缓存异常处理和周期精度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 183, "start": 434.2, "end": 436.18, "text": "任何一个模块的错误都可能", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 184, "start": 436.18, "end": 438.54, "text": "在几十万条指令之后才出现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 185, "start": 438.88, "end": 440.58, "text": "teamwork能发挥作用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 186, "start": 440.6, "end": 442.54, "text": "靠的是让不同角色负责", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 187, "start": 442.54, "end": 445.14, "text": "实现测试质疑和众合", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 188, "start": 445.58, "end": 447.6, "text": "并把验证嵌进工作流", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 189, "start": 448.0, "end": 450.08, "text": "模型本身提供推理能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 190, "start": 450.18, "end": 451.8, "text": "协作机制负责把这种", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 191, "start": 451.8, "end": 453.94, "text": "能力约束在工程流程里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 192, "start": 453.94, "end": 456.54, "text": "更有说服力的证据来自开源项目", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 193, "start": 456.8, "end": 459.54, "text": "teamwork对Eigen和Polyhash", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 194, "start": 459.78, "end": 461.28, "text": "提出了性能优化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 195, "start": 461.34, "end": 463.18, "text": "部分改动通过正常", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 196, "start": 463.18, "end": 465.34, "text": "代码审查进入上游项目", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 197, "start": 465.48, "end": 467.28, "text": "实验报告里的漂亮数字可以", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 198, "start": 467.28, "end": 469.48, "text": "受到环境和选择方式影响", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 199, "start": 469.6, "end": 471.9, "text": "外部维护者愿意接收代码", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 200, "start": 472.0, "end": 475.28, "text": "至少说明改动不只存在于演示页面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 201, "start": 475.54, "end": 476.58, "text": "他开始碰到真实", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 202, "start": 476.58, "end": 478.68, "text": "软件世界最麻烦的部分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 203, "start": 478.88, "end": 481.1, "text": "旧代码兼容性测试", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 204, "start": 481.1, "end": 483.08, "text": "要求和他人的审查意见", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 205, "start": 483.34, "end": 485.98, "text": "Polyhash的结果还展示了研究", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 206, "start": 485.98, "end": 488.08, "text": "伙伴与代码助手的差别", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 207, "start": 488.8, "end": 491.58, "text": "Google报告相关改动让64线程下的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 208, "start": 491.58, "end": 494.18, "text": "初始插入吞吐量提高约两倍", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 209, "start": 494.3, "end": 497.18, "text": "单线程整体性能提高约1.5倍", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 210, "start": 497.54, "end": 499.5, "text": "与flat hash map相比", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 211, "start": 499.54, "end": 501.88, "text": "内存还减少约25%", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 212, "start": 501.98, "end": 504.54, "text": "性能优化往往存在交换条件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 213, "start": 504.74, "end": 507.5, "text": "一处更快可能换来另一处更慢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 214, "start": 507.8, "end": 509.4, "text": "一种数据分布有效", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 215, "start": 509.4, "end": 511.94, "text": "也可能在另一种负载下失灵", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 216, "start": 512.18, "end": 513.88, "text": "只有通过基准代码", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 217, "start": 513.88, "end": 515.54, "text": "审查和维护者判断", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 218, "start": 515.74, "end": 518.1, "text": "优化才有机会进入真实项目", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 219, "start": 518.14, "end": 520.74, "text": "这也是teamwork模式比自动生成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 220, "start": 520.74, "end": 523.0, "text": "整个项目更值得关注的地方", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 221, "start": 523.54, "end": 524.94, "text": "研究和工程的价值", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 222, "start": 524.94, "end": 526.88, "text": "很少由代码行数决定", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 223, "start": 527.1, "end": 528.78, "text": "系统需要提出假设", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 224, "start": 528.98, "end": 530.54, "text": "选择测量方法", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 225, "start": 530.68, "end": 532.0, "text": "识别性能瓶颈", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 226, "start": 532.14, "end": 533.78, "text": "再接受外部审查", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 227, "start": 533.9, "end": 536.0, "text": "生成代码只是其中一步", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 228, "start": 536.3, "end": 538.88, "text": "Google把成果送进现有开源社区", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 229, "start": 539.14, "end": 540.9, "text": "相当于让智能体进入一个", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 230, "start": 540.9, "end": 543.58, "text": "由人类规则历史兼容性和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 231, "start": 543.58, "end": 545.4, "text": "公开讨论组成的环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 232, "start": 545.7, "end": 547.38, "text": "难度远高于在封闭", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 233, "start": 547.38, "end": 549.5, "text": "演示里让测试全部变绿", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 234, "start": 549.68, "end": 552.34, "text": "teamwork解决了谁和谁一起做的问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 235, "start": 552.78, "end": 554.74, "text": "长任务还有另一个隐蔽故障", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 236, "start": 555.18, "end": 556.74, "text": "智能体每执行一步", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 237, "start": 556.8, "end": 559.2, "text": "通常都会把观察结果工具", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 238, "start": 559.2, "end": 561.98, "text": "输出和中间推理追加到对话", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 239, "start": 562.1, "end": 563.7, "text": "10步以后还能管理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 240, "start": 563.9, "end": 565.0, "text": "100步以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 241, "start": 565.18, "end": 567.38, "text": "上下文就会变成混合了当前", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 242, "start": 567.38, "end": 570.7, "text": "事实过期结论和失败尝试的仓库", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 243, "start": 571.18, "end": 572.8, "text": "模型每次行动之前", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 244, "start": 572.98, "end": 574.48, "text": "都要重新从这堆", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 245, "start": 574.48, "end": 576.18, "text": "文字里恢复世界状态", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 246, "start": 576.34, "end": 577.8, "text": "人类做长期项目时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 247, "start": 577.8, "end": 579.0, "text": "不会这样工作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 248, "start": 579.38, "end": 581.28, "text": "工程师不会每天把几个月", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 249, "start": 581.28, "end": 583.28, "text": "的聊天记录从头读一遍", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 250, "start": 583.34, "end": 585.7, "text": "再猜代码现在编译到哪一步", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 251, "start": 585.98, "end": 587.68, "text": "团队会维护任务看板", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 252, "start": 587.74, "end": 588.78, "text": "测试状态", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 253, "start": 588.98, "end": 590.94, "text": "风险清单和当前分值", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 254, "start": 591.18, "end": 592.88, "text": "历史记录可以留档", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 255, "start": 593.18, "end": 595.2, "text": "下一步决策依赖的却是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 256, "start": 595.2, "end": 597.0, "text": "一份清楚的当前状态", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 257, "start": 597.14, "end": 599.6, "text": "把这个尝试放进智能体运行时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 258, "start": 600.1, "end": 602.18, "text": "每一步只给模型三类信息", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 259, "start": 602.5, "end": 604.88, "text": "固定的任务规范当前结构化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 260, "start": 604.88, "end": 607.88, "text": "状态以及环境刚刚返回的新观察", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 261, "start": 608.1, "end": 610.0, "text": "模型完成推理后提出", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 262, "start": 610.0, "end": 612.14, "text": "状态修改和下一步动作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 263, "start": 612.3, "end": 614.94, "text": "运行时先验证修改是否合法", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 264, "start": 615.08, "end": 616.28, "text": "再写入状态", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 265, "start": 616.54, "end": 618.94, "text": "那一轮中间推理随后被丢弃", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 266, "start": 619.08, "end": 621.1, "text": "不再跟着任务无限增长", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 267, "start": 621.18, "end": 624.1, "text": "这与传统的上下文压缩有根本区别", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 268, "start": 624.7, "end": 626.18, "text": "摘要方法仍然需要把", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 269, "start": 626.18, "end": 628.1, "text": "过去重新写成一段文字", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 270, "start": 628.34, "end": 630.3, "text": "信息可能在压缩时变形", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 271, "start": 630.68, "end": 632.34, "text": "模型也要判断摘要里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 272, "start": 632.34, "end": 634.6, "text": "哪些内容仍然有效skill", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 273, "start": 635.34, "end": 636.48, "text": "state直接规定哪些", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 274, "start": 636.48, "end": 637.94, "text": "字段代表当前世界", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 275, "start": 638.18, "end": 639.98, "text": "例如安全挑战可以只", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 276, "start": 639.98, "end": 641.58, "text": "保留已经发现的标记", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 277, "start": 641.88, "end": 643.18, "text": "测试过的假设", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 278, "start": 643.5, "end": 644.5, "text": "活跃文件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 279, "start": 644.6, "end": 646.68, "text": "当前目录和命令摘要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 280, "start": 646.94, "end": 649.14, "text": "下一步行动读取这些字段", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 281, "start": 649.14, "end": 650.2, "text": "不必回放", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 282, "start": 650.3, "end": 651.8, "text": "全部终端输出", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 283, "start": 651.98, "end": 654.68, "text": "论文给出的复杂度变化很直接", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 284, "start": 655.28, "end": 657.54, "text": "传统追加历史的方式单次", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 285, "start": 657.54, "end": 659.94, "text": "输入会随着任务长度增长", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 286, "start": 660.2, "end": 661.8, "text": "累计TOKEN消耗可能", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 287, "start": 661.8, "end": 663.6, "text": "接近任务步数的平方", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 288, "start": 663.78, "end": 665.58, "text": "只要状态结构和最新", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 289, "start": 665.58, "end": 667.4, "text": "观察的大小受到限制", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 290, "start": 667.98, "end": 669.78, "text": "skill state每一步的提示", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 291, "start": 669.78, "end": 671.94, "text": "规模可以保持在固定范围", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 292, "start": 672.4, "end": 674.88, "text": "累计消耗随步数线性增加", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 293, "start": 675.1, "end": 677.54, "text": "任务从几十步拉到几百步时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 294, "start": 677.6, "end": 679.74, "text": "这种差异会迅速放大", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 295, "start": 679.78, "end": 682.2, "text": "在一项100步的仓库任务里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 296, "start": 682.4, "end": 684.98, "text": "Gemini3Flash使用普通react", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 297, "start": 684.98, "end": 687.9, "text": "方法取得84%的成功率", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 298, "start": 688.14, "end": 690.78, "text": "累计消耗约124万TOKEN", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 299, "start": 691.28, "end": 694.1, "text": "skill state把成功率提高到94%", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 300, "start": 694.38, "end": 696.38, "text": "累计TOKEN约6.5万", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 301, "start": 696.4, "end": 697.54, "text": "到了200步", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 302, "start": 697.74, "end": 700.5, "text": "react成功率下降到74%", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 303, "start": 700.9, "end": 703.3, "text": "skill state仍保持94%", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 304, "start": 703.88, "end": 705.94, "text": "论文在intercode CTF安全", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 305, "start": 705.94, "end": 707.6, "text": "任务上的结果也显示", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 306, "start": 708.1, "end": 711.14, "text": "skill state完成率为54.2%", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 307, "start": 711.2, "end": 713.48, "text": "它高于几种对照运行方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 308, "start": 713.68, "end": 715.4, "text": "同时把平均提示规模", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 309, "start": 715.4, "end": 717.14, "text": "和总token压了下来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 310, "start": 717.2, "end": 719.5, "text": "更关键的实验是噪声测试", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 311, "start": 720.18, "end": 722.2, "text": "研究人员向任务环境加入", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 312, "start": 722.2, "end": 724.34, "text": "与状态无关的遥测信息", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 313, "start": 724.58, "end": 727.34, "text": "普通提示在噪声变多后明显恶化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 314, "start": 727.48, "end": 730.0, "text": "因为无用信息也会一直留在历史里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 315, "start": 730.8, "end": 732.88, "text": "skill state只把影响后续", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 316, "start": 732.88, "end": 734.58, "text": "行动的内容写进状态", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 317, "start": 734.94, "end": 737.7, "text": "所以成功率保持在接近完整水平", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 318, "start": 737.8, "end": 740.1, "text": "它展示了一条很实用的原则", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 319, "start": 740.48, "end": 742.28, "text": "长任务的可靠性不只", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 320, "start": 742.28, "end": 744.48, "text": "取决于模型能记住多少", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 321, "start": 744.58, "end": 746.38, "text": "还取决于系统敢不敢", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 322, "start": 746.38, "end": 748.18, "text": "丢掉已经没用的东西", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 323, "start": 748.3, "end": 750.2, "text": "这个方案也有明显代价", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 324, "start": 750.78, "end": 753.5, "text": "状态结构必须由人提前设计", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 325, "start": 754.08, "end": 757.2, "text": "如果某条早期信息当时看起来无关", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 326, "start": 757.38, "end": 759.28, "text": "后来才发现它很重要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 327, "start": 759.7, "end": 761.7, "text": "而它又没有进入状态字段", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 328, "start": 761.94, "end": 763.98, "text": "这条线索可能已经丢失", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 329, "start": 764.18, "end": 765.88, "text": "需要完整审计轨迹", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 330, "start": 765.98, "end": 767.3, "text": "解释研究过程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 331, "start": 767.3, "end": 769.38, "text": "或者回看思路变化的任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 332, "start": 769.74, "end": 771.7, "text": "也不能只保留当前状态", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 333, "start": 772.08, "end": 774.6, "text": "论文目前主要验证单智能体", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 334, "start": 774.74, "end": 777.9, "text": "多个智能体同时修改同一份状态时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 335, "start": 777.94, "end": 778.94, "text": "会不会覆盖", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 336, "start": 779.14, "end": 781.14, "text": "冲突又该由谁裁决", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 337, "start": 781.3, "end": 782.88, "text": "仍然没有充分答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 338, "start": 783.08, "end": 784.14, "text": "论文对小模型的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 339, "start": 784.14, "end": 785.88, "text": "失败分析进一步说明", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 340, "start": 786.1, "end": 788.8, "text": "结构化状态并不会自动带来可靠性", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 341, "start": 789.1, "end": 791.58, "text": "Gemma开放权重模型在100步", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 342, "start": 791.58, "end": 794.08, "text": "任务中只有42%的得分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 343, "start": 794.28, "end": 796.58, "text": "68%的失败来自过早", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 344, "start": 796.58, "end": 798.54, "text": "覆盖或删除重要状态", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 345, "start": 798.9, "end": 801.88, "text": "20%属于结构和类型错误", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 346, "start": 802.18, "end": 805.2, "text": "还有12%是JSON语法问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 347, "start": 805.48, "end": 808.38, "text": "运行时可以拒绝格式错误的更新", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 348, "start": 809.1, "end": 810.98, "text": "却无法保证模型对什么", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 349, "start": 810.98, "end": 812.74, "text": "值得保留的判断正确", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 350, "start": 813.0, "end": 814.1, "text": "状态越短", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 351, "start": 814.34, "end": 815.94, "text": "写错一个字段造成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 352, "start": 815.94, "end": 817.6, "text": "的影响可能越集中", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 353, "start": 817.8, "end": 819.1, "text": "承受系统因此需要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 354, "start": 819.1, "end": 820.8, "text": "同时保存两层记录", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 355, "start": 821.2, "end": 823.78, "text": "决策层读取精简的当前状态", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 356, "start": 824.1, "end": 826.0, "text": "审计层保存完整日志", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 357, "start": 826.28, "end": 828.28, "text": "出现异常时能够追溯", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 358, "start": 828.74, "end": 830.78, "text": "前者控制成本和注意力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 359, "start": 830.8, "end": 833.54, "text": "后者承担调查合规和复现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 360, "start": 833.94, "end": 837.08, "text": "skilled state论文强调丢弃中间推理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 361, "start": 837.2, "end": 839.0, "text": "适合追求执行效率", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 362, "start": 839.18, "end": 842.28, "text": "医疗金融和关键基础设施", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 363, "start": 842.28, "end": 844.38, "text": "可能要求更严格的留痕", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 364, "start": 844.78, "end": 848.68, "text": "怎样在隐私成本与可追责之间取舍", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 365, "start": 848.88, "end": 851.08, "text": "会决定它能进入哪些行业", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 366, "start": 851.28, "end": 852.8, "text": "状态结构的设计也", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 367, "start": 852.8, "end": 854.88, "text": "可能成为新的专业岗位", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 368, "start": 855.34, "end": 857.2, "text": "过去的提示工程关注", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 369, "start": 857.2, "end": 858.9, "text": "怎样把问题问清楚", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 370, "start": 859.54, "end": 860.9, "text": "长任务系统更需要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 371, "start": 860.9, "end": 862.78, "text": "定义任务进行到哪一步", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 372, "start": 862.78, "end": 864.5, "text": "应该有哪些字段表示", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 373, "start": 864.68, "end": 866.68, "text": "供应链任务要记录订单", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 374, "start": 866.88, "end": 868.88, "text": "交付风险和库存约束", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 375, "start": 869.08, "end": 872.1, "text": "代码任务要记录分支测试和依赖", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 376, "start": 872.3, "end": 873.78, "text": "科研任务则要记录", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 377, "start": 873.78, "end": 876.28, "text": "假设证据和未解决问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 378, "start": 876.34, "end": 878.58, "text": "同一个模板无法覆盖所有领域", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 379, "start": 879.0, "end": 881.78, "text": "领域专家会重新进入系统设计中心", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 380, "start": 881.94, "end": 884.98, "text": "现有证据还不足以把skill state写成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 381, "start": 884.98, "end": 888.1, "text": "teamwork已经采用的底层记忆系统", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 382, "start": 888.7, "end": 891.38, "text": "风控材料没有给出这种集成关系", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 383, "start": 891.74, "end": 893.4, "text": "它们分别击中了长期", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 384, "start": 893.4, "end": 894.9, "text": "智能体的两处故障", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 385, "start": 895.18, "end": 897.5, "text": "teamwork防止多个智能体集体", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 386, "start": 897.5, "end": 899.5, "text": "沿着错误方向自我强化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 387, "start": 899.88, "end": 901.48, "text": "skill state防止任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 388, "start": 901.48, "end": 903.38, "text": "历史把当前判断淹没", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 389, "start": 903.58, "end": 905.38, "text": "未来的研究系统很可能", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 390, "start": 905.38, "end": 907.18, "text": "同时需要这两种能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 391, "start": 907.88, "end": 909.68, "text": "Google是否会把它们合并", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 392, "start": 909.7, "end": 911.9, "text": "仍要看后续产品和论文", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 393, "start": 911.98, "end": 914.94, "text": "第三项能力来自Timesfm 3.0", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 394, "start": 915.1, "end": 916.78, "text": "它看起来没有数学证明和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 395, "start": 916.78, "end": 919.1, "text": "多智能体团队那么戏剧化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 396, "start": 919.18, "end": 921.08, "text": "却可能最接近企业每天", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 397, "start": 921.08, "end": 922.94, "text": "真正要花钱解决的问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 398, "start": 923.08, "end": 925.98, "text": "大量现实决策建立在时间序列上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 399, "start": 926.34, "end": 929.1, "text": "零售商要预测下周每家门店的销量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 400, "start": 929.2, "end": 931.7, "text": "电网要估计明天每小时的负荷", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 401, "start": 932.08, "end": 933.98, "text": "云平台要提前分配容量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 402, "start": 934.18, "end": 936.78, "text": "工厂要判断设备何时可能异常", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 403, "start": 936.98, "end": 939.14, "text": "语言模型擅长处理文字", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 404, "start": 939.34, "end": 942.1, "text": "时间序列要求模型理解趋势", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 405, "start": 942.38, "end": 946.08, "text": "周期突变变量联动和外部条件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 406, "start": 946.78, "end": 948.88, "text": "一个预测错几个百分点", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 407, "start": 948.88, "end": 950.5, "text": "可能对应库存积压", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 408, "start": 950.74, "end": 952.94, "text": "能源浪费或者服务中断", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 409, "start": 953.18, "end": 956.4, "text": "传统做法往往给不同业务单独建模", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 410, "start": 956.58, "end": 958.5, "text": "某个商品用一套模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 411, "start": 958.54, "end": 960.88, "text": "另一个地区重新调参数", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 412, "start": 961.18, "end": 963.1, "text": "遇到节假日促销和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 413, "start": 963.1, "end": 965.38, "text": "天气还要增加外生变量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 414, "start": 965.7, "end": 966.9, "text": "这个流程需要数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 415, "start": 966.9, "end": 968.58, "text": "科学团队长期维护", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 416, "start": 968.88, "end": 971.58, "text": "模型迁移到新场景时也容易失效", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 417, "start": 971.8, "end": 973.48, "text": "时间序列基础模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 418, "start": 973.48, "end": 975.4, "text": "希望学习更通用的规律", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 419, "start": 975.74, "end": 977.2, "text": "用一个预训练模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 420, "start": 977.2, "end": 979.18, "text": "直接处理大量陌生任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 421, "start": 979.4, "end": 982.7, "text": "TIMESFM3.0大约有3亿参数", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 422, "start": 982.7, "end": 985.1, "text": "支持单变量和多变量预测", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 423, "start": 985.14, "end": 986.74, "text": "也能接收协变量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 424, "start": 986.78, "end": 987.94, "text": "所谓多变量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 425, "start": 988.14, "end": 989.54, "text": "是让模型同时理解", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 426, "start": 989.54, "end": 991.4, "text": "几个彼此影响的序列", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 427, "start": 991.6, "end": 994.98, "text": "例如价格销量库存和客流", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 428, "start": 995.18, "end": 997.1, "text": "协变量则可以分成两类", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 429, "start": 997.18, "end": 998.9, "text": "一类只有过去数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 430, "start": 999.1, "end": 1000.98, "text": "另一类在未来已经知道", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 431, "start": 1001.34, "end": 1003.28, "text": "比如节假日安排电价", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 432, "start": 1003.28, "end": 1005.0, "text": "计划或者促销日程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 433, "start": 1005.2, "end": 1007.1, "text": "这些信息进入预测后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 434, "start": 1007.1, "end": 1008.6, "text": "模型才能区分正常", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 435, "start": 1008.6, "end": 1010.54, "text": "季节变化和真正异常", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 436, "start": 1010.7, "end": 1013.18, "text": "它采用Stack mixing Transformer架构", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 437, "start": 1013.4, "end": 1015.28, "text": "引入varied attention处理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 438, "start": 1015.28, "end": 1016.94, "text": "不同变量之间的关系", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 439, "start": 1017.38, "end": 1019.2, "text": "归一化机制则用来减少", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 440, "start": 1019.2, "end": 1021.3, "text": "不同序列尺度带来的干扰", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 441, "start": 1021.6, "end": 1023.94, "text": "训练材料包括去除评测", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 442, "start": 1023.94, "end": 1026.0, "text": "重叠后的时间序列数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 443, "start": 1026.18, "end": 1029.18, "text": "也纳入截至2023年11月的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 444, "start": 1029.18, "end": 1031.14, "text": "Wikipedia页面访问量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 445, "start": 1031.34, "end": 1034.6, "text": "截至2022年底的Google Trends热门查询", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 446, "start": 1034.94, "end": 1038.38, "text": "合成数据和增强数据同样进入训练", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 447, "start": 1038.54, "end": 1039.9, "text": "这些细节说明", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 448, "start": 1039.98, "end": 1041.8, "text": "所谓通用预测能力依然", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 449, "start": 1041.8, "end": 1044.7, "text": "建立在大规模多来源序列上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 450, "start": 1044.7, "end": 1046.88, "text": "时间序列模型最怕数据泄露", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 451, "start": 1047.54, "end": 1049.88, "text": "训练集如果包含评测其信息", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 452, "start": 1050.08, "end": 1051.6, "text": "或者同一序列的片段", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 453, "start": 1051.6, "end": 1053.38, "text": "同时进入训练和测试", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 454, "start": 1053.58, "end": 1055.3, "text": "排名会显得异常漂亮", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 455, "start": 1055.54, "end": 1056.58, "text": "Google特别说明", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 456, "start": 1056.58, "end": 1058.4, "text": "gift Evals训练数据排除了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 457, "start": 1058.4, "end": 1060.54, "text": "与Fbench重叠的数据集", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 458, "start": 1060.88, "end": 1062.7, "text": "目的就是降低这种问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 459, "start": 1062.98, "end": 1063.9, "text": "即便如此", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 460, "start": 1064.34, "end": 1067.28, "text": "三个基准第一仍然属于发布方报告", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 461, "start": 1067.54, "end": 1069.7, "text": "企业部署前还要在自己的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 462, "start": 1069.7, "end": 1072.6, "text": "时间跨度异常条件和业务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 463, "start": 1072.6, "end": 1074.58, "text": "损失函数上重新验证", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 464, "start": 1074.7, "end": 1076.1, "text": "预测准确率也不能", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 465, "start": 1076.1, "end": 1077.68, "text": "单独决定业务价值", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 466, "start": 1078.2, "end": 1080.2, "text": "库存系统更关心缺货和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 467, "start": 1080.2, "end": 1082.4, "text": "积压哪一种错误代价更高", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 468, "start": 1082.78, "end": 1085.38, "text": "电网预测必须关注极端峰值", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 469, "start": 1085.78, "end": 1088.2, "text": "金融公开则可能在意尾部事件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 470, "start": 1088.38, "end": 1090.6, "text": "一个模型平均误差更低", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 471, "start": 1090.74, "end": 1093.48, "text": "未必在最昂贵的错误上表现更好", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 472, "start": 1093.6, "end": 1095.98, "text": "Timesfm提供分位数预测", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 473, "start": 1096.1, "end": 1098.54, "text": "让使用者看到不同可能区间", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 474, "start": 1099.08, "end": 1100.38, "text": "这比只交付一个单", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 475, "start": 1100.38, "end": 1102.18, "text": "点数字更适合决策", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 476, "start": 1102.34, "end": 1105.0, "text": "但最后仍需业务规则解释风险", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 477, "start": 1105.34, "end": 1108.28, "text": "Google列出了Fbench time Benchmark", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 478, "start": 1108.58, "end": 1110.48, "text": "和JIFF Evals三个主要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 479, "start": 1110.48, "end": 1112.5, "text": "时间序列基础模型评测", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 480, "start": 1112.68, "end": 1116.08, "text": "并称TIMESFM3.0均排名第一", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 481, "start": 1116.2, "end": 1119.74, "text": "Fbench包含100项多样化现实预测任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 482, "start": 1120.08, "end": 1122.38, "text": "timebenchmark覆盖50个领域", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 483, "start": 1122.38, "end": 1124.68, "text": "数据集和98项评估任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 484, "start": 1124.9, "end": 1128.14, "text": "这个结果说明基础模型路线正在从", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 485, "start": 1128.14, "end": 1131.9, "text": "文本图像继续向企业数字世界扩张", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 486, "start": 1132.14, "end": 1134.28, "text": "timesfm的商业价值不止", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 487, "start": 1134.28, "end": 1136.0, "text": "来自一次预测精度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 488, "start": 1136.4, "end": 1138.54, "text": "统一模型可以减少每个部门", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 489, "start": 1138.54, "end": 1141.2, "text": "重复训练和维护小模型的成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 490, "start": 1141.38, "end": 1143.58, "text": "也更适合被智能体调用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 491, "start": 1143.8, "end": 1145.68, "text": "一个负责供应链的智能体", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 492, "start": 1145.68, "end": 1149.18, "text": "如果能够读取订单库存和物流数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 493, "start": 1149.28, "end": 1152.0, "text": "再调用时间序列模型生成预测", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 494, "start": 1152.14, "end": 1155.1, "text": "它才有机会提出可执行的采购方案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 495, "start": 1155.4, "end": 1157.18, "text": "缺少这一层专业工具", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 496, "start": 1157.54, "end": 1159.74, "text": "通用智能体往往只能根据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 497, "start": 1159.74, "end": 1161.98, "text": "文字描述给出宽泛建议", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 498, "start": 1162.2, "end": 1164.38, "text": "Google在这个方向还有分发优势", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 499, "start": 1164.7, "end": 1166.9, "text": "timesfm较早版本已经进入", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 500, "start": 1166.9, "end": 1170.2, "text": "bigqueryml和googlesheets等产品生态", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 501, "start": 1170.58, "end": 1172.3, "text": "企业不需要先搭一套全新", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 502, "start": 1172.3, "end": 1174.1, "text": "平台便可以在熟悉的数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 503, "start": 1174.1, "end": 1175.98, "text": "工具中接触预测能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 504, "start": 1176.18, "end": 1180.68, "text": "TIMESFM3.0公开了Pytorch权重和使用方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 505, "start": 1181.0, "end": 1184.0, "text": "但现有资料没有证明3.0权重已经", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 506, "start": 1184.0, "end": 1186.9, "text": "同步替换这些产品里的所有版本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 507, "start": 1186.98, "end": 1188.58, "text": "产品整合能走多快", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 508, "start": 1188.68, "end": 1191.1, "text": "仍需以Google后续公告为准", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 509, "start": 1191.28, "end": 1194.54, "text": "许可证也是一个容易被忽略的限制", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 510, "start": 1195.3, "end": 1197.6, "text": "timesfm代码仓库采用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 511, "start": 1197.74, "end": 1200.98, "text": "Apache2.03.0预训练权重", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 512, "start": 1200.98, "end": 1203.58, "text": "却使用单独的非商业许可证", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 513, "start": 1203.88, "end": 1205.94, "text": "研究者可以下载和实验", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 514, "start": 1206.5, "end": 1209.78, "text": "默认权重目前不允许用于商业生产", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 515, "start": 1209.8, "end": 1212.1, "text": "把它简单称为完全开源", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 516, "start": 1212.3, "end": 1213.54, "text": "企业拿来就能用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 517, "start": 1213.54, "end": 1214.78, "text": "会误导观众", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 518, "start": 1215.58, "end": 1218.2, "text": "Google开放了能力展示和研究入口", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 519, "start": 1218.28, "end": 1220.4, "text": "同时保留了商业化控制权", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 520, "start": 1220.48, "end": 1222.74, "text": "把这三项工作放回同一张图", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 521, "start": 1222.94, "end": 1224.54, "text": "Google的策略开始显形", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 522, "start": 1225.08, "end": 1227.38, "text": "teamwork负责扩大搜索空间", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 523, "start": 1227.38, "end": 1230.3, "text": "让多个智能体竞争和互相纠错", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 524, "start": 1230.6, "end": 1232.38, "text": "skilled state试图控制长", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 525, "start": 1232.38, "end": 1233.94, "text": "任务的状态和成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 526, "start": 1234.2, "end": 1237.38, "text": "timesfm 3.0提供专业预测能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 527, "start": 1237.54, "end": 1240.28, "text": "协作状态专业工具", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 528, "start": 1240.7, "end": 1242.74, "text": "构成AI参与长期研究和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 529, "start": 1242.74, "end": 1244.48, "text": "工程工作的三根支柱", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 530, "start": 1245.0, "end": 1247.18, "text": "这条路线也解释了为什么单纯", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 531, "start": 1247.18, "end": 1250.7, "text": "扩大模型参数带来的惊喜正在减少", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 532, "start": 1251.0, "end": 1253.08, "text": "模型可以在基准上更聪明", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 533, "start": 1253.54, "end": 1256.0, "text": "复杂项目仍然会被组织问题拖垮", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 534, "start": 1256.2, "end": 1257.7, "text": "给一个人更高智商", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 535, "start": 1257.78, "end": 1259.94, "text": "不代表他自动拥有实验室", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 536, "start": 1260.14, "end": 1262.48, "text": "项目管理制度和专业仪器", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 537, "start": 1262.6, "end": 1264.08, "text": "AI系统也一样", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 538, "start": 1264.2, "end": 1266.0, "text": "模型负责提出和判断", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 539, "start": 1266.28, "end": 1267.88, "text": "编排系统负责调度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 540, "start": 1268.1, "end": 1270.54, "text": "状态系统负责记录当前事实", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 541, "start": 1270.98, "end": 1272.58, "text": "专业模型负责处理通用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 542, "start": 1272.58, "end": 1274.58, "text": "语言能力覆盖不到的数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 543, "start": 1274.88, "end": 1277.1, "text": "Google在这一轮竞争里的优势", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 544, "start": 1277.54, "end": 1280.3, "text": "来自这些组件之间潜在的闭环", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 545, "start": 1280.74, "end": 1281.98, "text": "它有Gemini模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 546, "start": 1281.98, "end": 1285.1, "text": "也有antigravity这个智能体工作界面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 547, "start": 1285.34, "end": 1288.2, "text": "cloud和bigquery可以承接企业数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 548, "start": 1288.4, "end": 1291.54, "text": "专业能力则覆盖科学代码和预测", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 549, "start": 1291.7, "end": 1293.94, "text": "只要这些产品真正连起来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 550, "start": 1293.98, "end": 1296.74, "text": "Google出售的将是一套能进入研究", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 551, "start": 1296.74, "end": 1299.1, "text": "流程和生产系统的工作环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 552, "start": 1299.58, "end": 1302.08, "text": "聊天机器人只是其中一个入口", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 553, "start": 1302.08, "end": 1304.0, "text": "这个闭环真正形成后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 554, "start": 1304.08, "end": 1306.34, "text": "用户的购买单位也会变化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 555, "start": 1306.6, "end": 1308.5, "text": "企业过去按模型调用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 556, "start": 1308.5, "end": 1310.2, "text": "次数购买智能能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 557, "start": 1310.38, "end": 1312.68, "text": "未来可能按完成的研究活动", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 558, "start": 1312.7, "end": 1315.7, "text": "工程任务或预测工作流付费", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 559, "start": 1315.9, "end": 1317.9, "text": "模型选择会退到系统内部", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 560, "start": 1318.4, "end": 1319.94, "text": "简单步骤调用Flash", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 561, "start": 1320.14, "end": 1321.7, "text": "困难判断切换Pro", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 562, "start": 1322.14, "end": 1324.28, "text": "时间序列交给timesfm", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 563, "start": 1324.34, "end": 1326.3, "text": "代码验证交给测试工具", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 564, "start": 1326.4, "end": 1328.9, "text": "客户关心的是任务有没有完成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 565, "start": 1329.3, "end": 1330.5, "text": "证据是否充分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 566, "start": 1330.54, "end": 1331.8, "text": "成本能否预测", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 567, "start": 1332.08, "end": 1334.08, "text": "Google已经在antigravity里提供", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 568, "start": 1334.08, "end": 1336.58, "text": "动态子智能体异步任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 569, "start": 1336.7, "end": 1338.98, "text": "定时执行和自定义角色", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 570, "start": 1339.3, "end": 1342.0, "text": "teamwork继续向上增加协作模式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 571, "start": 1342.5, "end": 1344.5, "text": "他还把antigravity扩展到", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 572, "start": 1344.5, "end": 1347.38, "text": "命令行IDE和企业平台", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 573, "start": 1347.54, "end": 1349.9, "text": "每一个入口都在降低智能体", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 574, "start": 1349.9, "end": 1352.18, "text": "进入现有工作的迁移成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 575, "start": 1352.3, "end": 1353.9, "text": "研究人员不必为了使用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 576, "start": 1353.9, "end": 1355.78, "text": "智能体放弃代码仓库", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 577, "start": 1356.38, "end": 1357.7, "text": "企业也不必把全部", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 578, "start": 1357.7, "end": 1359.6, "text": "数据搬进一个聊天网页", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 579, "start": 1359.78, "end": 1362.4, "text": "竞争对手当然会走同一方向", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 580, "start": 1362.54, "end": 1364.8, "text": "多智能体编排结构化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 581, "start": 1364.8, "end": 1367.2, "text": "状态和专业模型都没有", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 582, "start": 1367.48, "end": 1369.54, "text": "只属于Google的理论壁垒", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 583, "start": 1369.78, "end": 1372.74, "text": "Google较难复制的部分是搜索", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 584, "start": 1372.94, "end": 1376.1, "text": "云办公套件开发工具和自", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 585, "start": 1376.1, "end": 1378.0, "text": "研模型形成的分发网络", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 586, "start": 1378.34, "end": 1379.94, "text": "技术领先可以被追赶", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 587, "start": 1380.08, "end": 1381.68, "text": "把能力送到企业已经", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 588, "start": 1381.68, "end": 1383.68, "text": "存在的数据和流程里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 589, "start": 1383.74, "end": 1385.54, "text": "通常需要更长时间", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 590, "start": 1385.78, "end": 1386.74, "text": "潜力越大", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 591, "start": 1386.78, "end": 1388.6, "text": "成本问题越不能跳过", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 592, "start": 1389.2, "end": 1391.48, "text": "teamwork的高质量结果依赖", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 593, "start": 1391.48, "end": 1393.34, "text": "并行搜索和多轮验证", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 594, "start": 1393.5, "end": 1396.28, "text": "同一个问题同时运行十几条路线", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 595, "start": 1396.58, "end": 1400.1, "text": "会消耗更多推理资源Gemini 3.7", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 596, "start": 1400.4, "end": 1403.08, "text": "Flash的速度和价格可以降低门槛", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 597, "start": 1403.14, "end": 1405.18, "text": "却不会让算力成本消失", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 598, "start": 1405.4, "end": 1407.1, "text": "公开预览版本刻意在", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 599, "start": 1407.1, "end": 1409.0, "text": "能力和费用之间做平衡", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 600, "start": 1409.18, "end": 1412.08, "text": "也说明实验室里最大化智能的配置", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 601, "start": 1412.68, "end": 1415.2, "text": "很难直接成为每个用户的默认服务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 602, "start": 1415.4, "end": 1417.74, "text": "企业还要面对另一种成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 603, "start": 1418.08, "end": 1420.6, "text": "skill点state需要设计状态结构", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 604, "start": 1420.7, "end": 1423.9, "text": "timesfm需要干净且有意义的数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 605, "start": 1424.0, "end": 1426.48, "text": "teamwork需要明确验收标准", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 606, "start": 1426.6, "end": 1428.74, "text": "公司如果连流程指标和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 607, "start": 1428.74, "end": 1430.4, "text": "数据权限都没有理清", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 608, "start": 1430.98, "end": 1433.8, "text": "接入更多智能体只会把混乱自动化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 609, "start": 1433.94, "end": 1436.4, "text": "研究伙伴能够加快一套好流程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 610, "start": 1436.7, "end": 1439.18, "text": "也能更快执行一套错误流程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 611, "start": 1439.38, "end": 1441.34, "text": "责任边界同样没有消失", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 612, "start": 1441.58, "end": 1443.74, "text": "数学证明可以交给Ling验证", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 613, "start": 1444.14, "end": 1446.9, "text": "代码可以用测试和上游审查把关", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 614, "start": 1447.48, "end": 1449.08, "text": "时间序列预测却经常", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 615, "start": 1449.08, "end": 1451.8, "text": "影响库存能源和资金配置", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 616, "start": 1452.08, "end": 1453.58, "text": "模型给出结果以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 617, "start": 1453.78, "end": 1455.6, "text": "谁确认数据没有偏差", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 618, "start": 1456.1, "end": 1458.68, "text": "谁判断异常是否来自制度变化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 619, "start": 1458.94, "end": 1461.34, "text": "谁为错误决策承担后果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 620, "start": 1461.48, "end": 1462.68, "text": "这些问题不能交给", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 621, "start": 1462.68, "end": 1464.58, "text": "一个置信度数字解决", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 622, "start": 1465.08, "end": 1467.48, "text": "Google在teamwork介绍中反复强调", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 623, "start": 1467.6, "end": 1470.18, "text": "人类仍然掌握目标和最终验收", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 624, "start": 1470.38, "end": 1473.2, "text": "这句话点出了当前系统成立的前提", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 625, "start": 1473.4, "end": 1476.14, "text": "智能体可以扩大候选方案数量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 626, "start": 1476.3, "end": 1477.88, "text": "也能更快找漏洞", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 627, "start": 1478.3, "end": 1479.54, "text": "却不知道一个组织", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 628, "start": 1479.54, "end": 1481.5, "text": "真正愿意承担什么风险", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 629, "start": 1481.7, "end": 1483.88, "text": "研究方向的价值证据是否", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 630, "start": 1483.88, "end": 1486.54, "text": "足够结果何时可以进入现实", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 631, "start": 1486.54, "end": 1488.08, "text": "都需要有人签字", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 632, "start": 1488.2, "end": 1490.28, "text": "所以AI成为长期研究", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 633, "start": 1490.28, "end": 1492.18, "text": "伙伴的标志会很安静", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 634, "start": 1492.48, "end": 1494.38, "text": "他开始连续工作几天", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 635, "start": 1494.5, "end": 1496.8, "text": "知道当前项目处在哪一步", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 636, "start": 1497.08, "end": 1499.7, "text": "他能主动找人反驳自己的方案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 637, "start": 1499.9, "end": 1502.18, "text": "也会调用合适的专业模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 638, "start": 1502.34, "end": 1504.28, "text": "避免用语言猜数字", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 639, "start": 1504.48, "end": 1506.18, "text": "他留下可检查的结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 640, "start": 1506.2, "end": 1507.74, "text": "也接受外部验证", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 641, "start": 1508.28, "end": 1510.8, "text": "某个模型宣布自己达到专家水平", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 642, "start": 1510.88, "end": 1512.28, "text": "证明不了这些能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 643, "start": 1512.48, "end": 1515.3, "text": "Google正在补齐的正是这些基础能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 644, "start": 1515.58, "end": 1518.28, "text": "teamwork让AI学会组织争论", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 645, "start": 1518.6, "end": 1521.7, "text": "skill state让它在长任务中保持清醒", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 646, "start": 1521.9, "end": 1525.14, "text": "Times FM3.0让它进入专业预测", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 647, "start": 1525.28, "end": 1527.7, "text": "三项技术还没有合成一个无缝系统", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 648, "start": 1528.28, "end": 1530.68, "text": "距离可信赖的研究伙伴也差", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 649, "start": 1530.68, "end": 1534.58, "text": "着权限安全责任和成本管理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 650, "start": 1534.58, "end": 1536.74, "text": "但竞争的目标已经变了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 651, "start": 1536.78, "end": 1538.94, "text": "过去的AI比赛看谁能在", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 652, "start": 1538.94, "end": 1541.2, "text": "一分钟里给出最惊艳的答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 653, "start": 1541.88, "end": 1543.1, "text": "接下来的比赛", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 654, "start": 1543.2, "end": 1545.5, "text": "要看谁能在100个步骤之后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 655, "start": 1545.5, "end": 1547.58, "text": "仍然知道自己为什么这么做", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 656, "start": 1547.74, "end": 1549.8, "text": "能把结果交给专家复核", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 657, "start": 1550.14, "end": 1551.8, "text": "还能把整套流程控制", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 658, "start": 1551.8, "end": 1553.68, "text": "在可承受的成本里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 659, "start": 1553.74, "end": 1556.14, "text": "Google这次释放的信号很明确", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 660, "start": 1556.48, "end": 1557.88, "text": "模型负责起步", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 661, "start": 1558.08, "end": 1559.7, "text": "产品要让智能长期", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 662, "start": 1559.94, "end": 1562.0, "text": "稳定并且专业的工作", "source": "video_source_subtitle", "authority": "video_source_subtitle" } ], "transcript_selection": { "schema": "videonote.transcript_selection.v1", "status": "SELECTED", "requested_language": "zh", "duration": 1562.0, "reference_path": "/Users/sagawa/AI_WORK/video_notes/url/20260902_154043__Google_瘋了_絕密_三王牌_突襲_OpenAI_聊天機器人時代正式終結/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": 663, "duration_coverage": 0.922215 } } }