{ "text": "Deepseek刚刚做了一件反常的事\n它发布了新一代V4.1Flash\n却准备让这款名字里带着\n轻量的模型接管上一代V4Pro\n北京时间9月14日中午以后\n用户即使继续调用Pro接口\n请求也会被转给V4.1Flash\n而且只收Flash的价格\n一款便宜模型为什么能\n让旗舰型号暂时退场\n答案藏在这次发布\n最容易被忽略的地方\nDeepseek换掉了计算方式\n也开始重新定义\n什么叫大模型的性价比\nV4.1 Flash是Deepseek下一代架构\n家族公开亮相的第一款产品\n升级范围远超增加\n视觉模块和改动版本号\n官方把它称为这个新\n家族中尺寸最小的成员\n可它拥有5,520亿个参数\n这个小只存在于Deepseek\n规划中的下一代产品线里\n放到普通开源模型市场\n它依然是一台需要数据中心级\n资源才能完整部署的庞然大物\n5,520亿之外\n80亿和160亿更能解释\n这款模型怎样运行\nV4.1Flash继续采用混合专家模型\n但输入与输出阶段\n使用不同的激活规模\n模型读取内容时大约激活80亿参数\n生成答案时大约激活160亿参数\n总参数负责储存能力\n参与每一步运算的\n参数则被压到更小范围\nDeepseek给这种设计起的\n名字是causal encoder decoder\n也就是因果编码器与解码器结构\n为什么要把输入和输出分开算\n因为AI的工作形态已经变了\n聊天机器人时代\n用户问一句\n模型答一段\n输入和输出长度相差不大\n智能体时代完全不同\n一个代码代理可能先读\n几十万个TOKEN的仓库\n检查报错文档和提交记录\n最后只修改几百行代码\n一个研究代理可能翻完上百页资料\n调用十几次工具\n最后只交付一份报告\n大量成本花在看懂现场\n真正写出来的内容反而很短\n传统架构如果用同一套高\n成本路径处理所有TOKEN\n就会在输入阶段浪费大量算力\nV4.1Flash把读取和生成\n做成不对称路径\n让输入侧更轻\n输出侧再投入更多专家能力\n这不保证每个任务都更快\n却准确集中了智能体\n最常见的成本结构\nDeepseek要解决的是代理\n连续工作几小时以后\n账单能否保持可控\n单轮回答的价格已经\n不足以衡量这类产品\n这也解释了为什么Deepseek把\n视觉理解直接放进基础架构\n旧产品线里\n文本版V4Flash和V4Flash vision xlarge分开存在\n新版本发布以后\n这两个旧模型都被下线\n接口名称暂时转向V4.1Flash\n开发者不再需要先判断\n请求属于文本还是图片\n再选择不同模型\n截图图表网页界面和文字\n可以进入同一套推理流程\n这项产品整合对智能体很关键\n现实中的电脑操作\n同时包含文字和画面\n按钮位置会改变\n错误提示藏在截图里\n图表数据没有完整写进网页代码\n远程桌面甚至指向模型提供画面\n一个不能稳定理解视觉环境的代理\n工具再多也只能在\n半个互联网里行动\n原生视觉让V4.1 Flash更\n接近真正的电脑操作模型\n能力范围已经超出调用\n搜索接口的聊天机器人\n深度求索为这套\n架构设定了四个目标\n提高能力上限和推理速度\n扩大吞吐\n再把同一设计延伸到更大的模型\n最后一项尤其值得留意\nV4.1Flash很可能只是这一代\n架构的第一场公开压力测试\n输入激活80亿\n输出激活160亿的设计\n如果能在大规模API流量里保持稳定\n未来V4.1Pro就可以沿着同\n一条路线扩大容量\n而不用让成本跟参数量同步膨胀\n模型架构里还有另\n一条更实际的变化\nKV cache被继续压缩\n长上下文模型会把已经\n处理过的信息保存成缓存\n避免生成下一个TOKEN时从头计算\n上下文越长\n缓存占用越大\n模型跑100万TOKEN\n同时服务大量用户时\n系统瓶颈会同时出现在计算芯片\n显存容量数据搬运速度与外部存储\n上Deepseek公布的数据称\nV4.1Flash对HBM的需求降到上一代的1/4\n对SSD的需求降到1/8\n与Deepseek最早一代架构相比\nKV CACHE累计缩小约437倍\n这些数字需要等待独立部署测试\n但它们揭示了产品真正争夺的资源\n模型公司过去比谁有更多算力\n现在还要比谁能用更少\n显存保存更长的工作记忆\nHBM为什么如此重要\n因为它是高端AI芯片旁边最\n昂贵也最紧缺的部件之一\n模型参数激活值和上下文\n缓存都在争夺同一块高速空间\n缓存膨胀以后\n服务器要么减少并发用户\n要么增加GPU数量\n要么把部分数据\n挪到速度较慢的存储\n三条路都会增加成本\n把缓存压下来\n相当于在不购买同等\n数量新硬件的情况下\n为现有服务器腾出更多座位\nSSD需求下降会减少硬盘投入\n也会缓解长任务的数据交换\n为了维持上下文\n系统可能需要在显存和\n外部存储之间不断搬运数据\n交换越频繁\n延迟越高\n系统也越难保持稳定吞吐\nDeepseek把SSD需求压到1/8\n是在减少智能体\n长时间运行时的存储交通\n用户感受到的可能只是回复更快\n服务商看到的却是每台\n机器能承载更多并发任务\nV4.1Flash的第一层意义是重新安排\n参数显存和存储之间的关系\n基准题多做对几个只是结果之一\n大模型的能力竞争\n进入下一阶段以后\n真正难的工作是把高水平\n推理变成可以长期供应的服务\n模型只在演示中聪明没有用\n必须让数百万个代理同时工作\n企业才可能建立稳定收入\n第二层意义来自它的能力定位\nDeepseek公布的GPQA Diamond成绩是90.9\nHLE是36.8 codeforces评分达到3471\n数学与科学推理依然重要\n但这次最醒目的提升集中在终端\n代码仓库自动化和网络安全任务\nterminal bench 2.1达到90.6\ndeepsw 1.1达到74.2\nNL to Repo bench达到65.4\ncybergem达到88.1\n这些名字对普通用户不够直观\nTerminal bench测试模型能不能在真实\n终端环境里连续执行命令并完成\n任务\nDeepseek和NL to repo关注模型是否能\n理解软件仓库定位问题并修改代码\nAutomation bench考察跨工具自动化\ncyberjail与Seckbench则把模型\n放进网络安全任务中\n他们共同测试模型能否把事情做完\n聊天是否流畅只占很小一部分\n过去一款Flash模型往往\n意味着速度快价格低\n复杂任务需要交给Pro\nV4.1Flash想打破这套分工\n官方称它在性能价格速度和\n任务总耗时上都超过V4Pro\n因此决定有序下线旧Pro\n最有冲击力的是Deepseek\n愿意把实际Pro流量交给Flash\n单个分数领先反倒居于其次\n如果新模型在生产环境里撑不住\n这种路由会直接影响付费客户\n不过厂商敢切换流量不等于\n所有性能结论已经得到独立证明\n代码代理基准采用\nDeepseek Harness的minimal模式\n并使用最大推理强度\n和百万TOKEN上下文\n不同模型如果采用不同\n工具框架提示方式和预算\n最终成绩很难完全横向比较\n基准能证明V4.1Flash有\n很强的代理能力\n却不能证明它在任何代码\n库里都比所有旗舰模型可靠\n稀疏激活本身也不会\n自动兑现成低延迟\n混合专家模型需要先判断\n每个TOKEN该送给哪些专家\n再让不同设备交换结果\n专家分布不均时\n部分GPU会排队\n另一些GPU却在等待\n夸卡通信如果处理不好\n省下来的乘法运算\n可能被数据搬运吃掉\nV4.1 Flash能否跑得快\n取决于路由算法并行策略和\n推理框架能不能共同配合\n架构给出了效率空间\n工程系统负责把空间变成\n用户真正感受到的速度\n这也让官方速度与\n本地速度成为两个问题\nDeepseek自己的API可以围绕\n固定硬件做高度优化\n社区部署却要面对不同GPU\n不同带宽和不同量化方式\n一套在大型集群上\n表现优秀的专家路由\n移到规模较小的服务器后\n未必保持同样吞吐\n开放权重让外界有机会验证\n却不会让官方服务的成本优势\n自动复制到每一家企业机房\n还有一个常见误区\n任务分数提高\n不代表模型的每一步操作都更可信\n智能体需要连续执行\n几十甚至上百个动作\n每一步只有很小的错误概率\n累积到长链任务里也会明显放大\n终端基准做到90分\n进入企业生产环境以后\n仍要面对权限管理依赖\n冲突过时文档和不可逆操作\n模型越能行动\n安全边界就越重要V4.1Flash\n在cyberjams上得到88.1\n也会带来双重影响\n防守团队可以用它审计\n代码复现漏洞和处理告警\n供给者同样可以让\n模型自动寻找入口\n开源权重和廉价API降低了\n先进能力的使用门槛\n也会降低滥用门槛\n评价这款模型需要同时\n检查任务能力与部署治理\n后者包括限制权限记录\n操作并阻止高风险行为\n第三层意义是价格\nV4.1Flash上线后\nDeepseek继续采用峰谷计费\n闲时每百万TOKEN的缓存\n命中输入只收0.02元\n未命中输入1元\n输出4元\n高峰时段分别是0.04元2元和8元\n缓存命中价格低到几乎是在\n鼓励开发者重复携带长上下文\n这套定价为智能体量身定做\n一个代理每天处理同一套公司规则\n同一个代码仓库和同一批长期记忆\n大部分输入会反复出现\n缓存命中以后\n服务商不需要重新完成全部计算\n客户也获得更低价格\n只有新问题新网页和新任务\n产生的内容按未命中计费\n上下文越稳定\n成本优势越明显\n便宜的缓存还会改变产品设计\n过去开发者为了省钱\n会主动删掉历史记录\n把复杂任务切成短对话\n这样虽然减少TOKEN\n却容易丢失目标约束和中间结果\n缓存价格降到每\n百万TOKEN几分钱以后\n保留长期记忆变得更可行\n产品可以让代理记住项目状态\n而不用每次启动都重新解释背景\n输出价格依然远高于缓存命中输入\n这也在引导开发者\n减少无意义长回答\n代理最理想的工作方式是大量读取\n少量决策\n精确行动\nV4.1 Flash的非对称激活和API\n价格指向同一个方向\n输入可以很长\n输出应该服务于任务\n那些依靠生成大段\n文本制造工作量的产品\n反而无法充分利用这套成本结构\n峰谷计费则暴露出\nDeepseek的另一项现实约束\n闲时价格只有高峰的一半\n说明算力需求仍然\n会在特定时间集中\nDeepseek用价格把非紧急\n任务推向空闲时段\n让数据中心保持更稳定的利用率\n批量代码检查离线研究和\n夜间数据处理可以延后执行\n即时客服与在线编程\n则继续支付高峰成本\n这套价格承担着算力调度功能\n传统云服务按机器使用时间收费\nDeepseek把电力服务器利用率和\n请求时段直接写进TOKEN价格\n用户看到的是一张价格表\nDeepseek考虑的是怎样\n减少昂贵GPU的空转\n模型架构压低单次推理成本\n峰谷价格在平衡总流量\n两者组合起来才形成真正的低价\n腾讯的work Buddy code Buddy和open code\n已经成为首批官方合作伙伴\n这组名单也说明V4.1Flash瞄准的市场\n办公代理和代码工具\n成为首批展示场景\n娱乐聊天与图片问答被放到后面\n开发者会在这些\n产品里连续调用模型\n使用量大\n对延迟和成本极其敏感\n也最容易验证Flash能否\n承担原本属于Pro的任务\nDeepseek为什么愿意让Pro请求转到Flash\n一方面新架构可能确实把\n相近能力做得更便宜\n另一方面\n统一流量也能减少\n维护多款旧模型的成本\n每多保留一个生产模型\n服务商都要准备独立容量监控\n质量修补安全问题并处理接口兼容\n把旧Flash\n视觉实验版和Pro流量集中到V4.1Flash\n可以提高集群利用率\n这种统一也有风险\n部分客户选择Pro\n可能依赖它在特殊姿势长\n推理或固定输出风格上的表现\n接口名称不变\n底层模型却换了\n自动化流程就可能\n出现难以发现的漂移\nDeepseek需要让API返回明确版本信息\n并给企业足够的迁移窗口\n一次回答贵几分钱\n通常可以计算生产\n结果突然改变才是企业\n客户最难控制的成本\n模型权重仓库同步出现在hugging face\n意味着社区能够继续\n进行量化适配和私有部署\n开放权重会让更多\n团队检查模型行为\n也能推动SG Lang VLLM等推理\n框架快速支持新架构\n但5,520亿总参数提醒了一个现实\n普通游戏电脑无法\n顺畅运行完整模型\n部署仍需要大规模GPU和存储集群\n所以Flash这个名字\n不能被误读为消费级\n他描述Deepseek产品线中的计算效率\n绝对硬件门槛依然很高\n输入激活只有80亿\n总权重仍有5,520亿\n激活参数决定单步计算量\n总参数则决定权重存储规模\n很多宣传会把两者混在一起\n制造80亿参数达到旗舰水平的错觉\n这并不准确\n对企业来说\nAPI可能比自建集群更现实\n私有部署能够控制数据和成本\n但需要承担模型切分推理\n优化故障恢复与硬件采购\n只有调用量足够大\n数据足够敏感\n完整部署才可能划算\nDeepseek留下大规模部署合作入口\n说明它也知道开放权重不\n等于普通客户能够独立跑起来\n企业还要计算一笔常被忽略的账\n模型权重免费并不\n代表每次调用免费\n自建集群需要为峰值流量预留硬件\n低谷时机器又可能闲置\n官方API可以把不同\n客户的波峰摊在一起\n若企业每天调用量不稳定\n购买GPU通常比按TOKEN付费更贵\n只有数据不能离开内网\n延迟要求极高\n或者调用规模长期稳定时\n自建部署的控制权才\n可能抵消运维成本\n私有部署仍有独特价值\n团队可以固定版本\n避免接口悄悄换模\n也能针对自己的文档\n和代码建立专用缓存\n高度监管行业还可以把日志权限\n和数据保留策略掌握在自己手里\nV4.1Flash\n真正扩大的是选择空间\n企业可以先用低价API验证任务\n再决定哪些高频流程值得搬回本地\n而不用一开始就为整套集群下注\nV4.1Flash给闭源模型带来的最大压力\n也许来自价格参照\n绝对能力只是其中一部分\n只要他能用较低成本完成\n大部分代码代理和自动化任务\n客户就会开始追问\n更昂贵的旗舰模型\n到底多解决了哪些问题\n如果差距只存在于少数极难任务\n企业会把日常工作交给Flash\n只在失败时升级到高价模型\n这种分层路由可能成为\n下一阶段的标准架构\n简单任务先由低成本模型处理\n复杂任务再转给更强模型\nV4.1 Flash现在更进一步\n它试图把原本属于高价\n层的大部分工作吞掉\n旗舰模型不会消失\n但调用频率可能下降\n角色从默认入口变成异常处理器\n这会迫使其他厂商重新定价\n模型公司过去用更大\n上下文更高基准分数吸引客户\n接下来必须证明每\n一美元能完成多少有效任务\n只公布输入单价不够\n还要计算完成一次\n代码修改需要几轮对话\n多少工具调用和多长时间\n便宜但反复返工的模型并不便宜\n单价较高却一次做对\n的模型也未必昂贵\nDeepseek强调总用时超过V4Pro\n就是在争夺这项新指标\n智能体成本由TOKEN价格\n生成长度执行轮数失败\n重试和等待时间共同组成\n一款模型每百万TOKEN便宜一半\n如果需要多跑三轮\n最终账单仍可能更高\nV4.1Flash若能同时减少\n生成量和任务轮数\n价格优势才会真正落到客户手里\n新架构的输入输出不对称\n也可能影响模型行为\n输出阶段激活更多参数\n理论上可以在生成决策\n时调动更丰富的专家\n输入阶段更精简\n则要确保压缩以后\n不会漏掉关键细节\n代码库里一个不起眼的配置文件\n可能决定整个任务是否成功\n如何在降低输入成本\n的同时保留关键信息\n是独立测试最需要关注的问题\n百万TOKEN上下文需要同时\n检查容量和信息利用率\n模型能够装下100万TOKEN\n不等于能稳定使用其中每一处信息\n长上下文常见问题包括中间\n信息被忽视跨文件关系判断\n错误以及旧指令与新指令发生冲突\nkv CACHE减少了保存成本\n却没有自动解决检索与注意力质量\n真正的测试应该观察模型在超长\n任务末尾还能否准确引用早期约束\n视觉能力也需要经过相同审查\n原生多模态说明模型可以处理图片\n不代表它已经适合无人\n监督地操作所有界面\n网页按钮可能被遮挡\n确认框会改变状态\n恶意页面还能把提示注入\n藏在图片中\n视觉代理如果同时拥有\n终端浏览器和支付权限\n错误的代价远高于识别错一张照片\n这也是V4.1Flash最值得讨论的矛盾\nDeepseek正在把能力速度和价格\n同时推向智能体大规模应用\n可智能体的治理\n能力并没有自动跟上\n模型越便宜\n部署数量越大\n部署数量越大\n偶发错误就越容易变成系统性事故\n行业需要降低推理成本\n也要同步建立权限隔离日志\n审计人工批准和事故响应\n从竞争角度看\n这款模型也透露出Deepseek的节奏\nV4在4月亮相\n7月底更新Flash\n9月又推出V4.1Flash\n更新速度越来越快\n旧模型生命周期越来越短\n开发者获得了更强能力\n也承担更频繁的迁移\n模型厂商如果每隔几周\n改变接口背后的行为\n软件工程中的稳定\n版本观念就会受到挑战\nDeepseek用兼容路由\n减少了接口迁移成本\n旧名称仍能调用\n程序不会立即报错\n但行为兼容比接口兼容更难\n相同提示在新模型上\n可能生成不同长度\n调用不同工具\n或采用不同决策顺序\n企业需要保存\n回归测试\n用自己的任务验证升级\n官方总榜只能充当初步参考\nV4.1Flash的发布也说明\n开源与闭源竞争正在从\n模型权重扩展到服务体系\n权重可以下载\n真正好用\n还需要推理框架缓存系统\n工具协议监控平台和稳定API\nDeepseek一边开放仓库\n一边通过极低API价格吸引流量\n形成双重策略\n社区帮助它适配硬件\n官方服务则把最怕\n麻烦的客户留在平台上\n价格低到一定程度以后\n模型本身可能变成入口\n作为利润中心的重要性会下降\n开发者围绕Deepseek建立\n代理插件和工作流\n平台就能积累真实\n任务数据与生态影响力\n腾讯和open code把V4.1Flash放进\n开发者每天使用的工具里\n这层入口价值已经\n超过普通渠道合作\n谁掌握默认模型\n谁就更容易影响\n下一代软件的技术选择\n可是低价也会给\nDeepseek自己施加压力\nAPI收入必须覆盖GPU折旧电力带宽\n运维和研发缓存压缩与稀疏激活\n可以降低成本价格战却可能把\n节省下来的空间迅速让给客户\n只要竞争对手继续跟价\n模型服务就会越来越\n像云计算基础设施\n规模和利用率将决定利润\n发布当天的热度很快就会消退\n缓存命中价尤其容易制造误解\n每百万TOKEN2分钱只\n适用于系统已经保存\n再次复用的内容\n首次送入的新资料仍\n按未命中价格计算\n模型输出也按更高价格收费\n代理如果每轮都改写提示\n重新排序文档\n哪怕信息大致相同\n也未必形成稳定缓存\n真正省钱的产品需要把\n系统规则知识库片段和\n项目上下文设计成可重复前缀\n价格优势最终会奖励\n良好的上下文工程\n这会催生一种新的产品分工\n模型厂商负责把\n缓存做得更密更便宜\n应用开发者负责让\n任务产生可复用内容\n企业则要判断哪些\n信息适合长期保存\n缓存保留越多\n隐私和数据生命周期问题也越突出\n低价上下文不能无限保存客户秘密\n服务商必须说明\n缓存是否跨请求复用\n保存多久\n以及用户能否要求删除\n这正是V4.1Flash比一张\n跑分表更重要的原因\n它把模型竞争里的\n三条路线压到了一起\n用非对称架构减少计算\n用缓存压缩释放硬件\n再用价格和接口迁移扩大调用量\n任何一条单独成立都不够\n三条同时成立\nDeepseek才有机会让Flash真正替代Pro\n接下来需要观察四件事\n第一独立评测能否复现\n代码和智能体成绩\n第二百万上下文中的信息\n保持能力是否稳定\n第三视觉工具调用在真实\n网页里能否控制错误\n第四Pro流量切换以后\n企业客户是否出现质量回退\n这些结果比发布会上的领先\n百分比更能决定V4.1Flash的位置\n对普通用户而言\n最直接的变化可能没有那么戏剧化\n应用响应会更快\n更多产品会加入图片理解\n长任务不再轻易因为费用中断\n用户甚至未必知道\n底层模型已经更换\n真正剧烈的变化发生在开发者后台\n同样预算能够运行更多代理\n过去不值得自动化的小\n任务开始变得有经济意义\n一家公司每天有\n10万份文档需要分类\n模型价格下降一点\n全年成本就会出现巨大差异\n一个代码平台每天\n处理数百万次仓库查询\n缓存需求减少3/4\n硬件规划也会改变\nV4.1Flash争夺的就是这些看起来无聊\n调用量却极大的任务\n消费级聊天带来知名度\n重复发生的企业\n流程才带来持续流量\n对国内模型行业来说\nDeepseek还在提高低价旗舰的门槛\n其他厂商无法只拿\n一个更小模型参加竞争\n还得同时支持长上下文\n视觉工具调用和代码代理\n能力追平以后\n价格又必须足够低\n这样的竞争会加速行业整合\n拥有算力工程和分发\n能力的平台更容易留下\n单纯依靠模型演示\n的公司会越来越难\n对海外厂商而言\n压力来自另一种方向\n闭源模型仍可能在\n最高难度任务上领先\n也拥有成熟企业合规与服务支持\n但V4.1Flash会迫使客户进行任务拆分\n大量日常调用如果能转向便宜模型\n高价厂商只能从最\n复杂的少数请求中赚钱\n他们需要证明的不再是模型最好\n而是领先部分足以抵消价格差\nDeepseek这次没有发布一款\n真正意义上的小模型\n他发布的是一台把庞大参数藏\n在稀疏计算后面的工业机器\n5,520亿参数提供容量\n80亿和160亿激活规模控制成本\n压缩缓存负责承载长期任务\n原生视觉负责进入真实电脑环境\n最后再用API低价把这些\n能力送进开发者工具\n开头那个问题现在有了答案\nV4.1Flash能够暂时接管Pro\n靠的是把旗舰能力装进\n更低成本的运行结构\nFlash这个名称并没有失去意义\nDeepseek想淘汰一款旧模型\n也在挑战能力越强就\n一定越贵的行业惯性\n这场测试才刚开始\n官方基准可以证明方向\n生产流量才能证明结果\n如果9月14日以后的Pro请求真的在\n用户无感的情况下由Flash完成\nV4.1就会成为一次重要的架构转折\n如果质量回退\n长任务失控或视觉代理频繁犯错\n这次激进迁移也会\n暴露新架构的边界\nDeepseek把最强的承诺\n写进了路由系统\n宣传页面上的文字反倒居于其次\n便宜模型要接管昂贵模型的工作\n最终裁判不会是排行榜\n裁判是每一次完成的任务\n每一笔实际账单\n以及企业敢不敢把关键流程交给他", "language": "zh", "duration": 1712.0, "segments": [ { "id": 0, "start": 0.166, "end": 2.766, "text": "Deepseek刚刚做了一件反常的事", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 1, "start": 2.9, "end": 5.4, "text": "它发布了新一代V4.1Flash", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 2, "start": 5.566, "end": 7.6, "text": "却准备让这款名字里带着", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 3, "start": 7.6, "end": 10.6, "text": "轻量的模型接管上一代V4Pro", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 4, "start": 10.966, "end": 13.366, "text": "北京时间9月14日中午以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 5, "start": 13.5, "end": 15.866, "text": "用户即使继续调用Pro接口", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 6, "start": 16.0, "end": 18.666, "text": "请求也会被转给V4.1Flash", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 7, "start": 18.8, "end": 20.733, "text": "而且只收Flash的价格", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 8, "start": 20.933, "end": 22.666, "text": "一款便宜模型为什么能", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 9, "start": 22.666, "end": 24.766, "text": "让旗舰型号暂时退场", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 10, "start": 24.9, "end": 26.466, "text": "答案藏在这次发布", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 11, "start": 26.466, "end": 28.133, "text": "最容易被忽略的地方", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 12, "start": 28.566, "end": 30.7, "text": "Deepseek换掉了计算方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 13, "start": 30.766, "end": 32.133, "text": "也开始重新定义", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 14, "start": 32.133, "end": 34.1, "text": "什么叫大模型的性价比", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 15, "start": 34.3, "end": 38.266, "text": "V4.1 Flash是Deepseek下一代架构", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 16, "start": 38.266, "end": 40.533, "text": "家族公开亮相的第一款产品", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 17, "start": 40.866, "end": 42.5, "text": "升级范围远超增加", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 18, "start": 42.5, "end": 44.566, "text": "视觉模块和改动版本号", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 19, "start": 44.933, "end": 46.366, "text": "官方把它称为这个新", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 20, "start": 46.366, "end": 48.666, "text": "家族中尺寸最小的成员", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 21, "start": 48.866, "end": 51.7, "text": "可它拥有5,520亿个参数", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 22, "start": 51.966, "end": 53.9, "text": "这个小只存在于Deepseek", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 23, "start": 54.066, "end": 56.2, "text": "规划中的下一代产品线里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 24, "start": 56.266, "end": 58.2, "text": "放到普通开源模型市场", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 25, "start": 58.266, "end": 60.5, "text": "它依然是一台需要数据中心级", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 26, "start": 60.5, "end": 63.066, "text": "资源才能完整部署的庞然大物", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 27, "start": 63.166, "end": 64.966, "text": "5,520亿之外", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 28, "start": 65.0, "end": 67.666, "text": "80亿和160亿更能解释", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 29, "start": 67.666, "end": 69.466, "text": "这款模型怎样运行", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 30, "start": 69.733, "end": 72.966, "text": "V4.1Flash继续采用混合专家模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 31, "start": 73.2, "end": 74.933, "text": "但输入与输出阶段", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 32, "start": 74.933, "end": 76.7, "text": "使用不同的激活规模", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 33, "start": 76.933, "end": 80.166, "text": "模型读取内容时大约激活80亿参数", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 34, "start": 80.466, "end": 83.6, "text": "生成答案时大约激活160亿参数", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 35, "start": 83.766, "end": 85.766, "text": "总参数负责储存能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 36, "start": 86.2, "end": 87.5, "text": "参与每一步运算的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 37, "start": 87.5, "end": 89.566, "text": "参数则被压到更小范围", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 38, "start": 89.766, "end": 91.666, "text": "Deepseek给这种设计起的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 39, "start": 91.666, "end": 94.2, "text": "名字是causal encoder decoder", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 40, "start": 94.6, "end": 97.6, "text": "也就是因果编码器与解码器结构", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 41, "start": 97.8, "end": 100.4, "text": "为什么要把输入和输出分开算", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 42, "start": 100.666, "end": 103.166, "text": "因为AI的工作形态已经变了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 43, "start": 103.533, "end": 104.933, "text": "聊天机器人时代", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 44, "start": 105.066, "end": 106.2, "text": "用户问一句", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 45, "start": 106.2, "end": 107.4, "text": "模型答一段", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 46, "start": 107.6, "end": 110.133, "text": "输入和输出长度相差不大", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 47, "start": 110.266, "end": 112.166, "text": "智能体时代完全不同", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 48, "start": 112.166, "end": 113.933, "text": "一个代码代理可能先读", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 49, "start": 113.933, "end": 115.533, "text": "几十万个TOKEN的仓库", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 50, "start": 115.8, "end": 118.4, "text": "检查报错文档和提交记录", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 51, "start": 118.7, "end": 120.766, "text": "最后只修改几百行代码", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 52, "start": 120.766, "end": 123.8, "text": "一个研究代理可能翻完上百页资料", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 53, "start": 124.0, "end": 125.533, "text": "调用十几次工具", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 54, "start": 125.966, "end": 127.733, "text": "最后只交付一份报告", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 55, "start": 127.933, "end": 130.2, "text": "大量成本花在看懂现场", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 56, "start": 130.4, "end": 132.9, "text": "真正写出来的内容反而很短", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 57, "start": 133.0, "end": 135.133, "text": "传统架构如果用同一套高", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 58, "start": 135.133, "end": 137.2, "text": "成本路径处理所有TOKEN", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 59, "start": 137.466, "end": 140.2, "text": "就会在输入阶段浪费大量算力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 60, "start": 140.466, "end": 143.4, "text": "V4.1Flash把读取和生成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 61, "start": 143.4, "end": 144.866, "text": "做成不对称路径", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 62, "start": 145.3, "end": 147.0, "text": "让输入侧更轻", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 63, "start": 147.133, "end": 149.9, "text": "输出侧再投入更多专家能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 64, "start": 149.933, "end": 152.166, "text": "这不保证每个任务都更快", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 65, "start": 152.333, "end": 154.1, "text": "却准确集中了智能体", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 66, "start": 154.1, "end": 155.7, "text": "最常见的成本结构", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 67, "start": 155.933, "end": 157.933, "text": "Deepseek要解决的是代理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 68, "start": 157.933, "end": 159.7, "text": "连续工作几小时以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 69, "start": 159.866, "end": 161.566, "text": "账单能否保持可控", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 70, "start": 161.9, "end": 163.466, "text": "单轮回答的价格已经", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 71, "start": 163.466, "end": 165.266, "text": "不足以衡量这类产品", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 72, "start": 165.366, "end": 167.7, "text": "这也解释了为什么Deepseek把", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 73, "start": 167.7, "end": 170.2, "text": "视觉理解直接放进基础架构", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 74, "start": 170.966, "end": 172.133, "text": "旧产品线里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 75, "start": 172.133, "end": 176.866, "text": "文本版V4Flash和V4Flash vision xlarge分开存在", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 76, "start": 177.066, "end": 178.533, "text": "新版本发布以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 77, "start": 178.566, "end": 180.533, "text": "这两个旧模型都被下线", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 78, "start": 180.933, "end": 183.7, "text": "接口名称暂时转向V4.1Flash", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 79, "start": 183.933, "end": 185.7, "text": "开发者不再需要先判断", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 80, "start": 185.7, "end": 187.933, "text": "请求属于文本还是图片", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 81, "start": 188.133, "end": 189.733, "text": "再选择不同模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 82, "start": 190.0, "end": 193.566, "text": "截图图表网页界面和文字", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 83, "start": 193.666, "end": 195.9, "text": "可以进入同一套推理流程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 84, "start": 195.966, "end": 198.533, "text": "这项产品整合对智能体很关键", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 85, "start": 199.366, "end": 200.7, "text": "现实中的电脑操作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 86, "start": 200.7, "end": 202.6, "text": "同时包含文字和画面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 87, "start": 202.766, "end": 204.1, "text": "按钮位置会改变", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 88, "start": 204.1, "end": 205.766, "text": "错误提示藏在截图里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 89, "start": 205.766, "end": 208.533, "text": "图表数据没有完整写进网页代码", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 90, "start": 208.533, "end": 211.733, "text": "远程桌面甚至指向模型提供画面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 91, "start": 211.8, "end": 214.733, "text": "一个不能稳定理解视觉环境的代理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 92, "start": 214.966, "end": 216.333, "text": "工具再多也只能在", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 93, "start": 216.333, "end": 217.9, "text": "半个互联网里行动", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 94, "start": 218.2, "end": 221.166, "text": "原生视觉让V4.1 Flash更", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 95, "start": 221.166, "end": 223.6, "text": "接近真正的电脑操作模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 96, "start": 224.133, "end": 226.066, "text": "能力范围已经超出调用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 97, "start": 226.066, "end": 228.1, "text": "搜索接口的聊天机器人", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 98, "start": 228.366, "end": 229.466, "text": "深度求索为这套", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 99, "start": 229.466, "end": 231.1, "text": "架构设定了四个目标", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 100, "start": 231.366, "end": 233.533, "text": "提高能力上限和推理速度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 101, "start": 233.6, "end": 234.566, "text": "扩大吞吐", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 102, "start": 234.7, "end": 237.3, "text": "再把同一设计延伸到更大的模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 103, "start": 237.5, "end": 239.333, "text": "最后一项尤其值得留意", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 104, "start": 239.7, "end": 242.333, "text": "V4.1Flash很可能只是这一代", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 105, "start": 242.333, "end": 244.7, "text": "架构的第一场公开压力测试", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 106, "start": 244.866, "end": 246.166, "text": "输入激活80亿", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 107, "start": 246.166, "end": 248.333, "text": "输出激活160亿的设计", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 108, "start": 248.7, "end": 251.733, "text": "如果能在大规模API流量里保持稳定", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 109, "start": 251.966, "end": 254.966, "text": "未来V4.1Pro就可以沿着同", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 110, "start": 254.966, "end": 256.6, "text": "一条路线扩大容量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 111, "start": 256.766, "end": 259.566, "text": "而不用让成本跟参数量同步膨胀", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 112, "start": 260.0, "end": 261.066, "text": "模型架构里还有另", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 113, "start": 261.066, "end": 262.6, "text": "一条更实际的变化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 114, "start": 262.766, "end": 264.733, "text": "KV cache被继续压缩", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 115, "start": 265.0, "end": 266.766, "text": "长上下文模型会把已经", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 116, "start": 266.766, "end": 268.933, "text": "处理过的信息保存成缓存", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 117, "start": 269.2, "end": 272.133, "text": "避免生成下一个TOKEN时从头计算", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 118, "start": 272.333, "end": 273.666, "text": "上下文越长", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 119, "start": 273.7, "end": 275.1, "text": "缓存占用越大", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 120, "start": 275.266, "end": 276.933, "text": "模型跑100万TOKEN", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 121, "start": 277.266, "end": 279.133, "text": "同时服务大量用户时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 122, "start": 279.3, "end": 282.266, "text": "系统瓶颈会同时出现在计算芯片", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 123, "start": 282.566, "end": 285.866, "text": "显存容量数据搬运速度与外部存储", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 124, "start": 285.866, "end": 288.566, "text": "上Deepseek公布的数据称", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 125, "start": 289.133, "end": 294.2, "text": "V4.1Flash对HBM的需求降到上一代的1/4", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 126, "start": 294.3, "end": 297.0, "text": "对SSD的需求降到1/8", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 127, "start": 297.0, "end": 299.533, "text": "与Deepseek最早一代架构相比", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 128, "start": 299.6, "end": 302.766, "text": "KV CACHE累计缩小约437倍", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 129, "start": 303.266, "end": 305.8, "text": "这些数字需要等待独立部署测试", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 130, "start": 305.933, "end": 308.9, "text": "但它们揭示了产品真正争夺的资源", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 131, "start": 309.1, "end": 311.7, "text": "模型公司过去比谁有更多算力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 132, "start": 311.933, "end": 313.766, "text": "现在还要比谁能用更少", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 133, "start": 313.766, "end": 316.1, "text": "显存保存更长的工作记忆", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 134, "start": 316.5, "end": 318.466, "text": "HBM为什么如此重要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 135, "start": 318.566, "end": 321.133, "text": "因为它是高端AI芯片旁边最", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 136, "start": 321.133, "end": 323.4, "text": "昂贵也最紧缺的部件之一", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 137, "start": 323.733, "end": 326.4, "text": "模型参数激活值和上下文", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 138, "start": 326.4, "end": 329.133, "text": "缓存都在争夺同一块高速空间", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 139, "start": 329.366, "end": 330.733, "text": "缓存膨胀以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 140, "start": 331.133, "end": 333.333, "text": "服务器要么减少并发用户", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 141, "start": 333.4, "end": 335.3, "text": "要么增加GPU数量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 142, "start": 335.4, "end": 336.9, "text": "要么把部分数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 143, "start": 336.933, "end": 338.7, "text": "挪到速度较慢的存储", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 144, "start": 339.1, "end": 340.9, "text": "三条路都会增加成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 145, "start": 340.966, "end": 342.5, "text": "把缓存压下来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 146, "start": 342.6, "end": 344.266, "text": "相当于在不购买同等", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 147, "start": 344.266, "end": 346.0, "text": "数量新硬件的情况下", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 148, "start": 346.1, "end": 348.733, "text": "为现有服务器腾出更多座位", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 149, "start": 348.966, "end": 352.2, "text": "SSD需求下降会减少硬盘投入", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 150, "start": 352.2, "end": 354.533, "text": "也会缓解长任务的数据交换", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 151, "start": 354.933, "end": 356.266, "text": "为了维持上下文", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 152, "start": 356.266, "end": 358.366, "text": "系统可能需要在显存和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 153, "start": 358.366, "end": 360.8, "text": "外部存储之间不断搬运数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 154, "start": 361.0, "end": 362.133, "text": "交换越频繁", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 155, "start": 362.5, "end": 363.533, "text": "延迟越高", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 156, "start": 363.566, "end": 365.8, "text": "系统也越难保持稳定吞吐", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 157, "start": 366.066, "end": 369.166, "text": "Deepseek把SSD需求压到1/8", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 158, "start": 369.166, "end": 370.666, "text": "是在减少智能体", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 159, "start": 370.666, "end": 372.866, "text": "长时间运行时的存储交通", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 160, "start": 373.166, "end": 375.8, "text": "用户感受到的可能只是回复更快", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 161, "start": 376.0, "end": 377.933, "text": "服务商看到的却是每台", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 162, "start": 377.933, "end": 380.4, "text": "机器能承载更多并发任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 163, "start": 380.5, "end": 384.1, "text": "V4.1Flash的第一层意义是重新安排", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 164, "start": 384.1, "end": 386.9, "text": "参数显存和存储之间的关系", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 165, "start": 387.166, "end": 390.1, "text": "基准题多做对几个只是结果之一", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 166, "start": 390.133, "end": 392.0, "text": "大模型的能力竞争", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 167, "start": 392.0, "end": 393.6, "text": "进入下一阶段以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 168, "start": 394.0, "end": 396.1, "text": "真正难的工作是把高水平", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 169, "start": 396.1, "end": 398.666, "text": "推理变成可以长期供应的服务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 170, "start": 398.8, "end": 401.566, "text": "模型只在演示中聪明没有用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 171, "start": 402.066, "end": 404.666, "text": "必须让数百万个代理同时工作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 172, "start": 404.866, "end": 407.066, "text": "企业才可能建立稳定收入", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 173, "start": 407.133, "end": 409.666, "text": "第二层意义来自它的能力定位", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 174, "start": 409.9, "end": 414.933, "text": "Deepseek公布的GPQA Diamond成绩是90.9", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 175, "start": 415.2, "end": 420.533, "text": "HLE是36.8 codeforces评分达到3471", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 176, "start": 420.666, "end": 423.066, "text": "数学与科学推理依然重要", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 177, "start": 423.7, "end": 426.4, "text": "但这次最醒目的提升集中在终端", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 178, "start": 426.6, "end": 430.266, "text": "代码仓库自动化和网络安全任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 179, "start": 430.5, "end": 433.133, "text": "terminal bench 2.1达到90.6", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 180, "start": 433.366, "end": 436.366, "text": "deepsw 1.1达到74.2", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 181, "start": 436.766, "end": 439.666, "text": "NL to Repo bench达到65.4", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 182, "start": 439.966, "end": 442.1, "text": "cybergem达到88.1", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 183, "start": 442.266, "end": 444.966, "text": "这些名字对普通用户不够直观", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 184, "start": 445.933, "end": 448.766, "text": "Terminal bench测试模型能不能在真实", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 185, "start": 448.766, "end": 451.6, "text": "终端环境里连续执行命令并完成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 186, "start": 451.6, "end": 452.4, "text": "任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 187, "start": 452.666, "end": 456.5, "text": "Deepseek和NL to repo关注模型是否能", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 188, "start": 456.5, "end": 459.666, "text": "理解软件仓库定位问题并修改代码", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 189, "start": 460.0, "end": 463.333, "text": "Automation bench考察跨工具自动化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 190, "start": 463.566, "end": 466.5, "text": "cyberjail与Seckbench则把模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 191, "start": 466.5, "end": 468.333, "text": "放进网络安全任务中", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 192, "start": 468.666, "end": 471.466, "text": "他们共同测试模型能否把事情做完", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 193, "start": 471.5, "end": 474.333, "text": "聊天是否流畅只占很小一部分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 194, "start": 474.466, "end": 476.2, "text": "过去一款Flash模型往往", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 195, "start": 476.2, "end": 478.3, "text": "意味着速度快价格低", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 196, "start": 478.466, "end": 480.3, "text": "复杂任务需要交给Pro", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 197, "start": 480.7, "end": 483.2, "text": "V4.1Flash想打破这套分工", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 198, "start": 483.3, "end": 486.0, "text": "官方称它在性能价格速度和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 199, "start": 486.0, "end": 488.6, "text": "任务总耗时上都超过V4Pro", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 200, "start": 488.8, "end": 491.1, "text": "因此决定有序下线旧Pro", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 201, "start": 491.466, "end": 493.1, "text": "最有冲击力的是Deepseek", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 202, "start": 493.366, "end": 495.9, "text": "愿意把实际Pro流量交给Flash", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 203, "start": 496.133, "end": 498.866, "text": "单个分数领先反倒居于其次", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 204, "start": 498.966, "end": 501.5, "text": "如果新模型在生产环境里撑不住", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 205, "start": 501.5, "end": 504.1, "text": "这种路由会直接影响付费客户", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 206, "start": 504.133, "end": 506.266, "text": "不过厂商敢切换流量不等于", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 207, "start": 506.266, "end": 508.9, "text": "所有性能结论已经得到独立证明", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 208, "start": 509.6, "end": 511.066, "text": "代码代理基准采用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 209, "start": 511.066, "end": 513.3, "text": "Deepseek Harness的minimal模式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 210, "start": 513.7, "end": 515.366, "text": "并使用最大推理强度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 211, "start": 515.366, "end": 517.3, "text": "和百万TOKEN上下文", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 212, "start": 517.366, "end": 519.0, "text": "不同模型如果采用不同", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 213, "start": 519.0, "end": 521.6, "text": "工具框架提示方式和预算", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 214, "start": 521.7, "end": 524.166, "text": "最终成绩很难完全横向比较", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 215, "start": 524.4, "end": 526.9, "text": "基准能证明V4.1Flash有", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 216, "start": 526.9, "end": 528.266, "text": "很强的代理能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 217, "start": 528.6, "end": 530.566, "text": "却不能证明它在任何代码", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 218, "start": 530.566, "end": 533.066, "text": "库里都比所有旗舰模型可靠", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 219, "start": 533.3, "end": 534.9, "text": "稀疏激活本身也不会", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 220, "start": 534.9, "end": 536.566, "text": "自动兑现成低延迟", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 221, "start": 537.166, "end": 539.166, "text": "混合专家模型需要先判断", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 222, "start": 539.166, "end": 541.366, "text": "每个TOKEN该送给哪些专家", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 223, "start": 541.5, "end": 543.666, "text": "再让不同设备交换结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 224, "start": 543.866, "end": 545.466, "text": "专家分布不均时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 225, "start": 545.466, "end": 547.1, "text": "部分GPU会排队", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 226, "start": 547.2, "end": 549.133, "text": "另一些GPU却在等待", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 227, "start": 549.466, "end": 551.533, "text": "夸卡通信如果处理不好", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 228, "start": 551.7, "end": 553.133, "text": "省下来的乘法运算", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 229, "start": 553.133, "end": 555.0, "text": "可能被数据搬运吃掉", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 230, "start": 555.333, "end": 558.266, "text": "V4.1 Flash能否跑得快", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 231, "start": 558.5, "end": 561.566, "text": "取决于路由算法并行策略和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 232, "start": 561.566, "end": 564.066, "text": "推理框架能不能共同配合", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 233, "start": 564.333, "end": 566.1, "text": "架构给出了效率空间", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 234, "start": 566.5, "end": 568.7, "text": "工程系统负责把空间变成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 235, "start": 568.7, "end": 570.9, "text": "用户真正感受到的速度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 236, "start": 570.966, "end": 572.733, "text": "这也让官方速度与", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 237, "start": 572.733, "end": 574.9, "text": "本地速度成为两个问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 238, "start": 575.166, "end": 577.5, "text": "Deepseek自己的API可以围绕", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 239, "start": 577.5, "end": 579.466, "text": "固定硬件做高度优化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 240, "start": 579.933, "end": 582.666, "text": "社区部署却要面对不同GPU", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 241, "start": 582.666, "end": 585.133, "text": "不同带宽和不同量化方式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 242, "start": 585.266, "end": 586.9, "text": "一套在大型集群上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 243, "start": 586.9, "end": 588.6, "text": "表现优秀的专家路由", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 244, "start": 588.866, "end": 591.1, "text": "移到规模较小的服务器后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 245, "start": 591.133, "end": 592.9, "text": "未必保持同样吞吐", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 246, "start": 593.3, "end": 595.5, "text": "开放权重让外界有机会验证", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 247, "start": 595.6, "end": 598.1, "text": "却不会让官方服务的成本优势", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 248, "start": 598.2, "end": 600.533, "text": "自动复制到每一家企业机房", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 249, "start": 600.6, "end": 602.0, "text": "还有一个常见误区", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 250, "start": 602.333, "end": 603.7, "text": "任务分数提高", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 251, "start": 603.7, "end": 606.5, "text": "不代表模型的每一步操作都更可信", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 252, "start": 606.733, "end": 608.366, "text": "智能体需要连续执行", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 253, "start": 608.366, "end": 610.3, "text": "几十甚至上百个动作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 254, "start": 610.466, "end": 612.7, "text": "每一步只有很小的错误概率", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 255, "start": 613.366, "end": 616.266, "text": "累积到长链任务里也会明显放大", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 256, "start": 616.4, "end": 618.333, "text": "终端基准做到90分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 257, "start": 618.4, "end": 620.133, "text": "进入企业生产环境以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 258, "start": 620.133, "end": 622.166, "text": "仍要面对权限管理依赖", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 259, "start": 622.166, "end": 624.566, "text": "冲突过时文档和不可逆操作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 260, "start": 624.666, "end": 625.9, "text": "模型越能行动", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 261, "start": 625.9, "end": 629.066, "text": "安全边界就越重要V4.1Flash", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 262, "start": 629.4, "end": 631.866, "text": "在cyberjams上得到88.1", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 263, "start": 631.866, "end": 633.533, "text": "也会带来双重影响", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 264, "start": 633.866, "end": 635.533, "text": "防守团队可以用它审计", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 265, "start": 635.533, "end": 638.0, "text": "代码复现漏洞和处理告警", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 266, "start": 638.4, "end": 639.733, "text": "供给者同样可以让", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 267, "start": 639.733, "end": 641.366, "text": "模型自动寻找入口", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 268, "start": 641.466, "end": 644.166, "text": "开源权重和廉价API降低了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 269, "start": 644.166, "end": 645.8, "text": "先进能力的使用门槛", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 270, "start": 645.933, "end": 647.733, "text": "也会降低滥用门槛", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 271, "start": 647.966, "end": 649.766, "text": "评价这款模型需要同时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 272, "start": 649.766, "end": 651.966, "text": "检查任务能力与部署治理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 273, "start": 652.166, "end": 654.666, "text": "后者包括限制权限记录", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 274, "start": 654.666, "end": 657.0, "text": "操作并阻止高风险行为", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 275, "start": 657.366, "end": 658.9, "text": "第三层意义是价格", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 276, "start": 659.166, "end": 661.533, "text": "V4.1Flash上线后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 277, "start": 661.6, "end": 664.2, "text": "Deepseek继续采用峰谷计费", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 278, "start": 664.5, "end": 666.3, "text": "闲时每百万TOKEN的缓存", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 279, "start": 666.3, "end": 668.7, "text": "命中输入只收0.02元", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 280, "start": 669.266, "end": 670.666, "text": "未命中输入1元", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 281, "start": 670.8, "end": 671.866, "text": "输出4元", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 282, "start": 672.266, "end": 676.566, "text": "高峰时段分别是0.04元2元和8元", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 283, "start": 676.7, "end": 678.9, "text": "缓存命中价格低到几乎是在", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 284, "start": 678.9, "end": 681.966, "text": "鼓励开发者重复携带长上下文", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 285, "start": 682.4, "end": 684.866, "text": "这套定价为智能体量身定做", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 286, "start": 684.866, "end": 687.733, "text": "一个代理每天处理同一套公司规则", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 287, "start": 687.9, "end": 691.0, "text": "同一个代码仓库和同一批长期记忆", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 288, "start": 691.566, "end": 693.7, "text": "大部分输入会反复出现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 289, "start": 693.966, "end": 695.266, "text": "缓存命中以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 290, "start": 695.366, "end": 698.066, "text": "服务商不需要重新完成全部计算", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 291, "start": 698.2, "end": 700.0, "text": "客户也获得更低价格", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 292, "start": 700.666, "end": 703.133, "text": "只有新问题新网页和新任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 293, "start": 703.133, "end": 705.533, "text": "产生的内容按未命中计费", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 294, "start": 705.866, "end": 707.3, "text": "上下文越稳定", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 295, "start": 707.366, "end": 708.933, "text": "成本优势越明显", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 296, "start": 709.1, "end": 711.666, "text": "便宜的缓存还会改变产品设计", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 297, "start": 712.2, "end": 714.0, "text": "过去开发者为了省钱", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 298, "start": 714.0, "end": 715.766, "text": "会主动删掉历史记录", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 299, "start": 715.766, "end": 717.9, "text": "把复杂任务切成短对话", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 300, "start": 718.5, "end": 720.0, "text": "这样虽然减少TOKEN", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 301, "start": 720.0, "end": 723.2, "text": "却容易丢失目标约束和中间结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 302, "start": 723.466, "end": 724.733, "text": "缓存价格降到每", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 303, "start": 724.733, "end": 726.666, "text": "百万TOKEN几分钱以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 304, "start": 726.7, "end": 729.066, "text": "保留长期记忆变得更可行", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 305, "start": 729.666, "end": 732.133, "text": "产品可以让代理记住项目状态", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 306, "start": 732.2, "end": 735.0, "text": "而不用每次启动都重新解释背景", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 307, "start": 735.2, "end": 738.566, "text": "输出价格依然远高于缓存命中输入", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 308, "start": 738.733, "end": 740.3, "text": "这也在引导开发者", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 309, "start": 740.3, "end": 741.933, "text": "减少无意义长回答", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 310, "start": 742.466, "end": 745.333, "text": "代理最理想的工作方式是大量读取", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 311, "start": 745.333, "end": 746.333, "text": "少量决策", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 312, "start": 746.366, "end": 747.4, "text": "精确行动", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 313, "start": 747.533, "end": 751.366, "text": "V4.1 Flash的非对称激活和API", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 314, "start": 751.366, "end": 752.966, "text": "价格指向同一个方向", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 315, "start": 753.4, "end": 754.6, "text": "输入可以很长", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 316, "start": 754.6, "end": 756.333, "text": "输出应该服务于任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 317, "start": 756.366, "end": 757.966, "text": "那些依靠生成大段", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 318, "start": 757.966, "end": 760.066, "text": "文本制造工作量的产品", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 319, "start": 760.266, "end": 762.933, "text": "反而无法充分利用这套成本结构", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 320, "start": 763.1, "end": 764.8, "text": "峰谷计费则暴露出", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 321, "start": 764.8, "end": 767.133, "text": "Deepseek的另一项现实约束", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 322, "start": 767.766, "end": 770.066, "text": "闲时价格只有高峰的一半", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 323, "start": 770.133, "end": 771.766, "text": "说明算力需求仍然", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 324, "start": 771.766, "end": 773.5, "text": "会在特定时间集中", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 325, "start": 773.9, "end": 776.366, "text": "Deepseek用价格把非紧急", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 326, "start": 776.366, "end": 778.266, "text": "任务推向空闲时段", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 327, "start": 778.4, "end": 781.366, "text": "让数据中心保持更稳定的利用率", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 328, "start": 781.5, "end": 783.766, "text": "批量代码检查离线研究和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 329, "start": 783.766, "end": 786.166, "text": "夜间数据处理可以延后执行", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 330, "start": 786.566, "end": 788.533, "text": "即时客服与在线编程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 331, "start": 788.533, "end": 790.5, "text": "则继续支付高峰成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 332, "start": 790.733, "end": 793.566, "text": "这套价格承担着算力调度功能", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 333, "start": 793.866, "end": 796.766, "text": "传统云服务按机器使用时间收费", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 334, "start": 797.0, "end": 799.733, "text": "Deepseek把电力服务器利用率和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 335, "start": 799.733, "end": 802.3, "text": "请求时段直接写进TOKEN价格", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 336, "start": 802.566, "end": 804.7, "text": "用户看到的是一张价格表", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 337, "start": 805.133, "end": 806.733, "text": "Deepseek考虑的是怎样", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 338, "start": 806.733, "end": 808.966, "text": "减少昂贵GPU的空转", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 339, "start": 809.133, "end": 811.7, "text": "模型架构压低单次推理成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 340, "start": 811.933, "end": 814.1, "text": "峰谷价格在平衡总流量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 341, "start": 814.2, "end": 817.066, "text": "两者组合起来才形成真正的低价", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 342, "start": 817.2, "end": 820.766, "text": "腾讯的work Buddy code Buddy和open code", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 343, "start": 821.1, "end": 823.4, "text": "已经成为首批官方合作伙伴", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 344, "start": 823.5, "end": 827.2, "text": "这组名单也说明V4.1Flash瞄准的市场", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 345, "start": 827.466, "end": 829.3, "text": "办公代理和代码工具", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 346, "start": 829.333, "end": 831.133, "text": "成为首批展示场景", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 347, "start": 831.5, "end": 834.5, "text": "娱乐聊天与图片问答被放到后面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 348, "start": 834.566, "end": 835.733, "text": "开发者会在这些", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 349, "start": 835.733, "end": 837.5, "text": "产品里连续调用模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 350, "start": 837.733, "end": 838.8, "text": "使用量大", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 351, "start": 838.8, "end": 841.2, "text": "对延迟和成本极其敏感", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 352, "start": 841.333, "end": 843.666, "text": "也最容易验证Flash能否", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 353, "start": 843.666, "end": 845.7, "text": "承担原本属于Pro的任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 354, "start": 845.933, "end": 849.0, "text": "Deepseek为什么愿意让Pro请求转到Flash", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 355, "start": 849.2, "end": 851.333, "text": "一方面新架构可能确实把", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 356, "start": 851.333, "end": 853.1, "text": "相近能力做得更便宜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 357, "start": 853.566, "end": 854.366, "text": "另一方面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 358, "start": 854.533, "end": 855.933, "text": "统一流量也能减少", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 359, "start": 855.933, "end": 857.866, "text": "维护多款旧模型的成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 360, "start": 857.933, "end": 860.1, "text": "每多保留一个生产模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 361, "start": 860.466, "end": 863.466, "text": "服务商都要准备独立容量监控", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 362, "start": 863.466, "end": 866.9, "text": "质量修补安全问题并处理接口兼容", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 363, "start": 867.066, "end": 868.366, "text": "把旧Flash", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 364, "start": 868.933, "end": 872.533, "text": "视觉实验版和Pro流量集中到V4.1Flash", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 365, "start": 872.733, "end": 874.466, "text": "可以提高集群利用率", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 366, "start": 874.5, "end": 876.066, "text": "这种统一也有风险", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 367, "start": 876.533, "end": 878.0, "text": "部分客户选择Pro", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 368, "start": 878.266, "end": 880.733, "text": "可能依赖它在特殊姿势长", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 369, "start": 880.733, "end": 883.066, "text": "推理或固定输出风格上的表现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 370, "start": 883.2, "end": 884.466, "text": "接口名称不变", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 371, "start": 884.7, "end": 886.166, "text": "底层模型却换了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 372, "start": 886.2, "end": 887.666, "text": "自动化流程就可能", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 373, "start": 887.666, "end": 889.533, "text": "出现难以发现的漂移", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 374, "start": 889.766, "end": 893.266, "text": "Deepseek需要让API返回明确版本信息", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 375, "start": 893.266, "end": 895.566, "text": "并给企业足够的迁移窗口", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 376, "start": 895.733, "end": 897.4, "text": "一次回答贵几分钱", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 377, "start": 897.4, "end": 898.966, "text": "通常可以计算生产", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 378, "start": 899.366, "end": 901.066, "text": "结果突然改变才是企业", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 379, "start": 901.066, "end": 902.766, "text": "客户最难控制的成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 380, "start": 902.933, "end": 906.1, "text": "模型权重仓库同步出现在hugging face", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 381, "start": 906.533, "end": 908.333, "text": "意味着社区能够继续", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 382, "start": 908.333, "end": 911.466, "text": "进行量化适配和私有部署", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 383, "start": 911.9, "end": 913.266, "text": "开放权重会让更多", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 384, "start": 913.266, "end": 914.933, "text": "团队检查模型行为", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 385, "start": 915.0, "end": 918.3, "text": "也能推动SG Lang VLLM等推理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 386, "start": 918.3, "end": 920.3, "text": "框架快速支持新架构", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 387, "start": 920.366, "end": 923.766, "text": "但5,520亿总参数提醒了一个现实", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 388, "start": 924.6, "end": 926.066, "text": "普通游戏电脑无法", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 389, "start": 926.066, "end": 927.8, "text": "顺畅运行完整模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 390, "start": 928.3, "end": 931.533, "text": "部署仍需要大规模GPU和存储集群", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 391, "start": 931.6, "end": 933.0, "text": "所以Flash这个名字", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 392, "start": 933.0, "end": 934.7, "text": "不能被误读为消费级", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 393, "start": 935.166, "end": 938.4, "text": "他描述Deepseek产品线中的计算效率", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 394, "start": 938.6, "end": 940.7, "text": "绝对硬件门槛依然很高", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 395, "start": 940.8, "end": 942.733, "text": "输入激活只有80亿", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 396, "start": 942.866, "end": 945.366, "text": "总权重仍有5,520亿", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 397, "start": 945.733, "end": 947.933, "text": "激活参数决定单步计算量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 398, "start": 948.1, "end": 950.7, "text": "总参数则决定权重存储规模", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 399, "start": 950.866, "end": 953.333, "text": "很多宣传会把两者混在一起", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 400, "start": 953.333, "end": 956.533, "text": "制造80亿参数达到旗舰水平的错觉", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 401, "start": 956.566, "end": 957.766, "text": "这并不准确", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 402, "start": 957.8, "end": 959.133, "text": "对企业来说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 403, "start": 959.4, "end": 962.066, "text": "API可能比自建集群更现实", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 404, "start": 962.5, "end": 964.8, "text": "私有部署能够控制数据和成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 405, "start": 964.866, "end": 967.166, "text": "但需要承担模型切分推理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 406, "start": 967.166, "end": 969.666, "text": "优化故障恢复与硬件采购", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 407, "start": 969.866, "end": 971.5, "text": "只有调用量足够大", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 408, "start": 971.5, "end": 972.766, "text": "数据足够敏感", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 409, "start": 972.9, "end": 974.666, "text": "完整部署才可能划算", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 410, "start": 974.9, "end": 977.966, "text": "Deepseek留下大规模部署合作入口", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 411, "start": 978.066, "end": 980.066, "text": "说明它也知道开放权重不", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 412, "start": 980.066, "end": 982.3, "text": "等于普通客户能够独立跑起来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 413, "start": 982.533, "end": 985.0, "text": "企业还要计算一笔常被忽略的账", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 414, "start": 985.4, "end": 987.133, "text": "模型权重免费并不", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 415, "start": 987.133, "end": 988.933, "text": "代表每次调用免费", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 416, "start": 989.266, "end": 992.1, "text": "自建集群需要为峰值流量预留硬件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 417, "start": 992.1, "end": 994.1, "text": "低谷时机器又可能闲置", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 418, "start": 994.3, "end": 995.7, "text": "官方API可以把不同", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 419, "start": 995.7, "end": 997.5, "text": "客户的波峰摊在一起", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 420, "start": 997.866, "end": 999.966, "text": "若企业每天调用量不稳定", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 421, "start": 1000.066, "end": 1003.133, "text": "购买GPU通常比按TOKEN付费更贵", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 422, "start": 1003.2, "end": 1005.266, "text": "只有数据不能离开内网", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 423, "start": 1005.5, "end": 1006.966, "text": "延迟要求极高", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 424, "start": 1007.066, "end": 1009.666, "text": "或者调用规模长期稳定时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 425, "start": 1009.8, "end": 1011.933, "text": "自建部署的控制权才", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 426, "start": 1011.933, "end": 1013.6, "text": "可能抵消运维成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 427, "start": 1013.8, "end": 1015.7, "text": "私有部署仍有独特价值", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 428, "start": 1016.4, "end": 1017.8, "text": "团队可以固定版本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 429, "start": 1017.8, "end": 1019.466, "text": "避免接口悄悄换模", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 430, "start": 1019.666, "end": 1021.133, "text": "也能针对自己的文档", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 431, "start": 1021.133, "end": 1022.966, "text": "和代码建立专用缓存", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 432, "start": 1023.2, "end": 1025.8, "text": "高度监管行业还可以把日志权限", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 433, "start": 1025.8, "end": 1028.766, "text": "和数据保留策略掌握在自己手里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 434, "start": 1028.933, "end": 1030.1, "text": "V4.1Flash", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 435, "start": 1030.5, "end": 1032.533, "text": "真正扩大的是选择空间", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 436, "start": 1032.733, "end": 1035.6, "text": "企业可以先用低价API验证任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 437, "start": 1035.733, "end": 1039.133, "text": "再决定哪些高频流程值得搬回本地", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 438, "start": 1039.3, "end": 1042.166, "text": "而不用一开始就为整套集群下注", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 439, "start": 1042.466, "end": 1045.933, "text": "V4.1Flash给闭源模型带来的最大压力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 440, "start": 1046.266, "end": 1048.0, "text": "也许来自价格参照", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 441, "start": 1048.166, "end": 1050.3, "text": "绝对能力只是其中一部分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 442, "start": 1050.566, "end": 1052.533, "text": "只要他能用较低成本完成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 443, "start": 1052.533, "end": 1055.066, "text": "大部分代码代理和自动化任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 444, "start": 1055.266, "end": 1056.9, "text": "客户就会开始追问", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 445, "start": 1057.2, "end": 1058.566, "text": "更昂贵的旗舰模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 446, "start": 1058.566, "end": 1060.4, "text": "到底多解决了哪些问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 447, "start": 1060.5, "end": 1063.466, "text": "如果差距只存在于少数极难任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 448, "start": 1063.733, "end": 1066.166, "text": "企业会把日常工作交给Flash", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 449, "start": 1066.366, "end": 1069.0, "text": "只在失败时升级到高价模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 450, "start": 1069.133, "end": 1070.933, "text": "这种分层路由可能成为", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 451, "start": 1070.933, "end": 1072.7, "text": "下一阶段的标准架构", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 452, "start": 1073.4, "end": 1076.066, "text": "简单任务先由低成本模型处理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 453, "start": 1076.2, "end": 1078.466, "text": "复杂任务再转给更强模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 454, "start": 1078.666, "end": 1081.3, "text": "V4.1 Flash现在更进一步", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 455, "start": 1081.366, "end": 1083.266, "text": "它试图把原本属于高价", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 456, "start": 1083.266, "end": 1085.066, "text": "层的大部分工作吞掉", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 457, "start": 1085.266, "end": 1086.666, "text": "旗舰模型不会消失", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 458, "start": 1086.666, "end": 1088.533, "text": "但调用频率可能下降", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 459, "start": 1088.566, "end": 1091.266, "text": "角色从默认入口变成异常处理器", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 460, "start": 1091.533, "end": 1093.866, "text": "这会迫使其他厂商重新定价", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 461, "start": 1094.0, "end": 1095.6, "text": "模型公司过去用更大", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 462, "start": 1095.6, "end": 1098.533, "text": "上下文更高基准分数吸引客户", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 463, "start": 1098.666, "end": 1100.0, "text": "接下来必须证明每", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 464, "start": 1100.0, "end": 1102.3, "text": "一美元能完成多少有效任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 465, "start": 1102.5, "end": 1104.533, "text": "只公布输入单价不够", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 466, "start": 1104.866, "end": 1106.333, "text": "还要计算完成一次", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 467, "start": 1106.333, "end": 1108.7, "text": "代码修改需要几轮对话", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 468, "start": 1108.866, "end": 1111.4, "text": "多少工具调用和多长时间", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 469, "start": 1111.933, "end": 1114.4, "text": "便宜但反复返工的模型并不便宜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 470, "start": 1114.6, "end": 1116.3, "text": "单价较高却一次做对", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 471, "start": 1116.3, "end": 1117.9, "text": "的模型也未必昂贵", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 472, "start": 1118.133, "end": 1121.366, "text": "Deepseek强调总用时超过V4Pro", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 473, "start": 1121.366, "end": 1123.333, "text": "就是在争夺这项新指标", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 474, "start": 1123.5, "end": 1125.566, "text": "智能体成本由TOKEN价格", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 475, "start": 1125.6, "end": 1128.333, "text": "生成长度执行轮数失败", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 476, "start": 1128.333, "end": 1130.666, "text": "重试和等待时间共同组成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 477, "start": 1130.9, "end": 1133.366, "text": "一款模型每百万TOKEN便宜一半", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 478, "start": 1133.7, "end": 1135.366, "text": "如果需要多跑三轮", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 479, "start": 1135.4, "end": 1137.466, "text": "最终账单仍可能更高", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 480, "start": 1137.6, "end": 1140.266, "text": "V4.1Flash若能同时减少", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 481, "start": 1140.266, "end": 1142.133, "text": "生成量和任务轮数", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 482, "start": 1142.766, "end": 1145.466, "text": "价格优势才会真正落到客户手里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 483, "start": 1145.533, "end": 1147.866, "text": "新架构的输入输出不对称", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 484, "start": 1148.133, "end": 1150.1, "text": "也可能影响模型行为", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 485, "start": 1150.266, "end": 1152.333, "text": "输出阶段激活更多参数", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 486, "start": 1152.566, "end": 1154.333, "text": "理论上可以在生成决策", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 487, "start": 1154.333, "end": 1156.466, "text": "时调动更丰富的专家", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 488, "start": 1156.866, "end": 1158.333, "text": "输入阶段更精简", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 489, "start": 1158.533, "end": 1160.066, "text": "则要确保压缩以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 490, "start": 1160.066, "end": 1161.666, "text": "不会漏掉关键细节", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 491, "start": 1161.9, "end": 1164.3, "text": "代码库里一个不起眼的配置文件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 492, "start": 1164.4, "end": 1166.766, "text": "可能决定整个任务是否成功", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 493, "start": 1167.466, "end": 1169.0, "text": "如何在降低输入成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 494, "start": 1169.0, "end": 1171.1, "text": "的同时保留关键信息", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 495, "start": 1171.366, "end": 1174.1, "text": "是独立测试最需要关注的问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 496, "start": 1174.3, "end": 1176.933, "text": "百万TOKEN上下文需要同时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 497, "start": 1176.933, "end": 1179.066, "text": "检查容量和信息利用率", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 498, "start": 1179.733, "end": 1181.8, "text": "模型能够装下100万TOKEN", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 499, "start": 1182.1, "end": 1185.0, "text": "不等于能稳定使用其中每一处信息", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 500, "start": 1185.366, "end": 1187.866, "text": "长上下文常见问题包括中间", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 501, "start": 1187.866, "end": 1190.566, "text": "信息被忽视跨文件关系判断", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 502, "start": 1190.566, "end": 1193.9, "text": "错误以及旧指令与新指令发生冲突", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 503, "start": 1194.166, "end": 1196.166, "text": "kv CACHE减少了保存成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 504, "start": 1196.166, "end": 1199.066, "text": "却没有自动解决检索与注意力质量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 505, "start": 1199.1, "end": 1201.566, "text": "真正的测试应该观察模型在超长", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 506, "start": 1201.566, "end": 1204.766, "text": "任务末尾还能否准确引用早期约束", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 507, "start": 1205.1, "end": 1207.7, "text": "视觉能力也需要经过相同审查", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 508, "start": 1208.366, "end": 1211.4, "text": "原生多模态说明模型可以处理图片", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 509, "start": 1211.4, "end": 1213.2, "text": "不代表它已经适合无人", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 510, "start": 1213.2, "end": 1215.133, "text": "监督地操作所有界面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 511, "start": 1215.333, "end": 1217.166, "text": "网页按钮可能被遮挡", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 512, "start": 1217.6, "end": 1219.2, "text": "确认框会改变状态", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 513, "start": 1219.5, "end": 1221.533, "text": "恶意页面还能把提示注入", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 514, "start": 1221.533, "end": 1222.7, "text": "藏在图片中", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 515, "start": 1222.866, "end": 1224.466, "text": "视觉代理如果同时拥有", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 516, "start": 1224.466, "end": 1227.066, "text": "终端浏览器和支付权限", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 517, "start": 1227.3, "end": 1230.2, "text": "错误的代价远高于识别错一张照片", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 518, "start": 1230.3, "end": 1233.566, "text": "这也是V4.1Flash最值得讨论的矛盾", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 519, "start": 1234.466, "end": 1237.6, "text": "Deepseek正在把能力速度和价格", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 520, "start": 1237.6, "end": 1240.066, "text": "同时推向智能体大规模应用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 521, "start": 1240.166, "end": 1241.4, "text": "可智能体的治理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 522, "start": 1241.4, "end": 1243.133, "text": "能力并没有自动跟上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 523, "start": 1243.3, "end": 1244.466, "text": "模型越便宜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 524, "start": 1244.8, "end": 1246.3, "text": "部署数量越大", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 525, "start": 1246.6, "end": 1247.9, "text": "部署数量越大", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 526, "start": 1248.066, "end": 1250.933, "text": "偶发错误就越容易变成系统性事故", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 527, "start": 1251.1, "end": 1253.166, "text": "行业需要降低推理成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 528, "start": 1253.2, "end": 1256.0, "text": "也要同步建立权限隔离日志", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 529, "start": 1256.0, "end": 1258.8, "text": "审计人工批准和事故响应", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 530, "start": 1258.966, "end": 1260.466, "text": "从竞争角度看", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 531, "start": 1260.533, "end": 1263.366, "text": "这款模型也透露出Deepseek的节奏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 532, "start": 1263.9, "end": 1265.466, "text": "V4在4月亮相", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 533, "start": 1265.666, "end": 1267.266, "text": "7月底更新Flash", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 534, "start": 1267.466, "end": 1269.8, "text": "9月又推出V4.1Flash", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 535, "start": 1270.266, "end": 1271.966, "text": "更新速度越来越快", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 536, "start": 1272.166, "end": 1274.5, "text": "旧模型生命周期越来越短", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 537, "start": 1274.8, "end": 1276.666, "text": "开发者获得了更强能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 538, "start": 1276.866, "end": 1278.866, "text": "也承担更频繁的迁移", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 539, "start": 1279.066, "end": 1280.866, "text": "模型厂商如果每隔几周", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 540, "start": 1280.866, "end": 1282.566, "text": "改变接口背后的行为", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 541, "start": 1282.9, "end": 1284.333, "text": "软件工程中的稳定", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 542, "start": 1284.333, "end": 1286.166, "text": "版本观念就会受到挑战", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 543, "start": 1286.6, "end": 1288.533, "text": "Deepseek用兼容路由", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 544, "start": 1288.766, "end": 1290.666, "text": "减少了接口迁移成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 545, "start": 1290.8, "end": 1292.4, "text": "旧名称仍能调用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 546, "start": 1292.5, "end": 1294.066, "text": "程序不会立即报错", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 547, "start": 1294.333, "end": 1296.966, "text": "但行为兼容比接口兼容更难", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 548, "start": 1297.066, "end": 1298.966, "text": "相同提示在新模型上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 549, "start": 1298.966, "end": 1300.6, "text": "可能生成不同长度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 550, "start": 1300.733, "end": 1301.933, "text": "调用不同工具", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 551, "start": 1301.933, "end": 1303.766, "text": "或采用不同决策顺序", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 552, "start": 1304.066, "end": 1305.3, "text": "企业需要保存", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 553, "start": 1305.3, "end": 1306.266, "text": "回归测试", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 554, "start": 1306.4, "end": 1308.566, "text": "用自己的任务验证升级", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 555, "start": 1308.8, "end": 1311.333, "text": "官方总榜只能充当初步参考", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 556, "start": 1311.7, "end": 1314.066, "text": "V4.1Flash的发布也说明", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 557, "start": 1314.333, "end": 1316.5, "text": "开源与闭源竞争正在从", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 558, "start": 1316.5, "end": 1318.933, "text": "模型权重扩展到服务体系", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 559, "start": 1319.266, "end": 1320.366, "text": "权重可以下载", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 560, "start": 1320.766, "end": 1321.766, "text": "真正好用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 561, "start": 1321.866, "end": 1324.6, "text": "还需要推理框架缓存系统", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 562, "start": 1324.7, "end": 1328.2, "text": "工具协议监控平台和稳定API", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 563, "start": 1328.4, "end": 1330.6, "text": "Deepseek一边开放仓库", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 564, "start": 1330.666, "end": 1333.8, "text": "一边通过极低API价格吸引流量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 565, "start": 1333.8, "end": 1335.2, "text": "形成双重策略", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 566, "start": 1335.733, "end": 1337.7, "text": "社区帮助它适配硬件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 567, "start": 1337.733, "end": 1339.166, "text": "官方服务则把最怕", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 568, "start": 1339.166, "end": 1341.166, "text": "麻烦的客户留在平台上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 569, "start": 1341.266, "end": 1343.3, "text": "价格低到一定程度以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 570, "start": 1343.5, "end": 1345.966, "text": "模型本身可能变成入口", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 571, "start": 1346.366, "end": 1349.333, "text": "作为利润中心的重要性会下降", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 572, "start": 1349.533, "end": 1351.566, "text": "开发者围绕Deepseek建立", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 573, "start": 1351.566, "end": 1353.966, "text": "代理插件和工作流", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 574, "start": 1354.066, "end": 1355.533, "text": "平台就能积累真实", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 575, "start": 1355.533, "end": 1357.666, "text": "任务数据与生态影响力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 576, "start": 1357.866, "end": 1360.866, "text": "腾讯和open code把V4.1Flash放进", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 577, "start": 1360.866, "end": 1363.0, "text": "开发者每天使用的工具里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 578, "start": 1363.366, "end": 1364.733, "text": "这层入口价值已经", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 579, "start": 1364.733, "end": 1366.466, "text": "超过普通渠道合作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 580, "start": 1366.7, "end": 1368.3, "text": "谁掌握默认模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 581, "start": 1368.466, "end": 1369.7, "text": "谁就更容易影响", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 582, "start": 1369.7, "end": 1371.666, "text": "下一代软件的技术选择", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 583, "start": 1371.766, "end": 1373.133, "text": "可是低价也会给", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 584, "start": 1373.133, "end": 1374.9, "text": "Deepseek自己施加压力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 585, "start": 1375.4, "end": 1380.066, "text": "API收入必须覆盖GPU折旧电力带宽", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 586, "start": 1380.066, "end": 1383.366, "text": "运维和研发缓存压缩与稀疏激活", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 587, "start": 1383.366, "end": 1386.066, "text": "可以降低成本价格战却可能把", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 588, "start": 1386.066, "end": 1388.533, "text": "节省下来的空间迅速让给客户", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 589, "start": 1388.866, "end": 1391.066, "text": "只要竞争对手继续跟价", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 590, "start": 1391.3, "end": 1392.766, "text": "模型服务就会越来越", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 591, "start": 1392.766, "end": 1394.8, "text": "像云计算基础设施", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 592, "start": 1395.166, "end": 1397.6, "text": "规模和利用率将决定利润", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 593, "start": 1398.2, "end": 1400.733, "text": "发布当天的热度很快就会消退", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 594, "start": 1401.0, "end": 1403.766, "text": "缓存命中价尤其容易制造误解", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 595, "start": 1404.0, "end": 1405.966, "text": "每百万TOKEN2分钱只", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 596, "start": 1405.966, "end": 1407.766, "text": "适用于系统已经保存", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 597, "start": 1407.8, "end": 1409.333, "text": "再次复用的内容", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 598, "start": 1409.9, "end": 1411.7, "text": "首次送入的新资料仍", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 599, "start": 1411.7, "end": 1413.533, "text": "按未命中价格计算", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 600, "start": 1414.2, "end": 1416.733, "text": "模型输出也按更高价格收费", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 601, "start": 1416.8, "end": 1419.0, "text": "代理如果每轮都改写提示", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 602, "start": 1419.3, "end": 1420.666, "text": "重新排序文档", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 603, "start": 1420.966, "end": 1422.866, "text": "哪怕信息大致相同", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 604, "start": 1422.966, "end": 1424.966, "text": "也未必形成稳定缓存", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 605, "start": 1425.366, "end": 1427.1, "text": "真正省钱的产品需要把", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 606, "start": 1427.1, "end": 1429.566, "text": "系统规则知识库片段和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 607, "start": 1429.566, "end": 1432.266, "text": "项目上下文设计成可重复前缀", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 608, "start": 1432.566, "end": 1434.466, "text": "价格优势最终会奖励", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 609, "start": 1434.466, "end": 1436.066, "text": "良好的上下文工程", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 610, "start": 1436.166, "end": 1438.533, "text": "这会催生一种新的产品分工", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 611, "start": 1438.933, "end": 1440.3, "text": "模型厂商负责把", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 612, "start": 1440.3, "end": 1442.366, "text": "缓存做得更密更便宜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 613, "start": 1442.733, "end": 1444.166, "text": "应用开发者负责让", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 614, "start": 1444.166, "end": 1446.2, "text": "任务产生可复用内容", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 615, "start": 1446.766, "end": 1448.3, "text": "企业则要判断哪些", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 616, "start": 1448.3, "end": 1450.066, "text": "信息适合长期保存", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 617, "start": 1450.333, "end": 1451.733, "text": "缓存保留越多", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 618, "start": 1452.2, "end": 1455.1, "text": "隐私和数据生命周期问题也越突出", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 619, "start": 1455.266, "end": 1458.166, "text": "低价上下文不能无限保存客户秘密", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 620, "start": 1458.733, "end": 1459.966, "text": "服务商必须说明", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 621, "start": 1459.966, "end": 1462.1, "text": "缓存是否跨请求复用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 622, "start": 1462.1, "end": 1463.166, "text": "保存多久", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 623, "start": 1463.266, "end": 1465.366, "text": "以及用户能否要求删除", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 624, "start": 1465.4, "end": 1467.7, "text": "这正是V4.1Flash比一张", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 625, "start": 1467.7, "end": 1469.6, "text": "跑分表更重要的原因", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 626, "start": 1469.733, "end": 1471.133, "text": "它把模型竞争里的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 627, "start": 1471.133, "end": 1473.1, "text": "三条路线压到了一起", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 628, "start": 1473.3, "end": 1475.666, "text": "用非对称架构减少计算", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 629, "start": 1475.866, "end": 1478.066, "text": "用缓存压缩释放硬件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 630, "start": 1478.6, "end": 1481.5, "text": "再用价格和接口迁移扩大调用量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 631, "start": 1481.566, "end": 1483.5, "text": "任何一条单独成立都不够", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 632, "start": 1483.6, "end": 1485.066, "text": "三条同时成立", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 633, "start": 1485.3, "end": 1488.666, "text": "Deepseek才有机会让Flash真正替代Pro", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 634, "start": 1488.733, "end": 1490.733, "text": "接下来需要观察四件事", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 635, "start": 1490.9, "end": 1493.3, "text": "第一独立评测能否复现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 636, "start": 1493.3, "end": 1494.9, "text": "代码和智能体成绩", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 637, "start": 1495.466, "end": 1497.766, "text": "第二百万上下文中的信息", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 638, "start": 1497.766, "end": 1499.533, "text": "保持能力是否稳定", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 639, "start": 1499.7, "end": 1502.366, "text": "第三视觉工具调用在真实", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 640, "start": 1502.366, "end": 1504.166, "text": "网页里能否控制错误", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 641, "start": 1504.4, "end": 1506.4, "text": "第四Pro流量切换以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 642, "start": 1506.533, "end": 1508.933, "text": "企业客户是否出现质量回退", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 643, "start": 1509.1, "end": 1511.166, "text": "这些结果比发布会上的领先", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 644, "start": 1511.166, "end": 1514.366, "text": "百分比更能决定V4.1Flash的位置", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 645, "start": 1514.466, "end": 1515.866, "text": "对普通用户而言", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 646, "start": 1515.866, "end": 1518.666, "text": "最直接的变化可能没有那么戏剧化", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 647, "start": 1519.1, "end": 1520.566, "text": "应用响应会更快", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 648, "start": 1520.666, "end": 1522.9, "text": "更多产品会加入图片理解", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 649, "start": 1523.166, "end": 1525.566, "text": "长任务不再轻易因为费用中断", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 650, "start": 1525.866, "end": 1527.533, "text": "用户甚至未必知道", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 651, "start": 1527.533, "end": 1529.3, "text": "底层模型已经更换", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 652, "start": 1529.4, "end": 1532.366, "text": "真正剧烈的变化发生在开发者后台", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 653, "start": 1532.666, "end": 1535.0, "text": "同样预算能够运行更多代理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 654, "start": 1535.366, "end": 1537.066, "text": "过去不值得自动化的小", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 655, "start": 1537.066, "end": 1539.4, "text": "任务开始变得有经济意义", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 656, "start": 1539.533, "end": 1540.733, "text": "一家公司每天有", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 657, "start": 1540.733, "end": 1542.533, "text": "10万份文档需要分类", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 658, "start": 1542.733, "end": 1544.3, "text": "模型价格下降一点", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 659, "start": 1544.6, "end": 1546.866, "text": "全年成本就会出现巨大差异", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 660, "start": 1547.0, "end": 1548.566, "text": "一个代码平台每天", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 661, "start": 1548.566, "end": 1551.0, "text": "处理数百万次仓库查询", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 662, "start": 1551.366, "end": 1553.566, "text": "缓存需求减少3/4", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 663, "start": 1553.7, "end": 1555.333, "text": "硬件规划也会改变", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 664, "start": 1555.5, "end": 1559.533, "text": "V4.1Flash争夺的就是这些看起来无聊", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 665, "start": 1559.766, "end": 1561.6, "text": "调用量却极大的任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 666, "start": 1561.966, "end": 1563.933, "text": "消费级聊天带来知名度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 667, "start": 1564.166, "end": 1565.4, "text": "重复发生的企业", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 668, "start": 1565.4, "end": 1567.4, "text": "流程才带来持续流量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 669, "start": 1567.466, "end": 1569.333, "text": "对国内模型行业来说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 670, "start": 1569.533, "end": 1572.5, "text": "Deepseek还在提高低价旗舰的门槛", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 671, "start": 1572.933, "end": 1574.533, "text": "其他厂商无法只拿", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 672, "start": 1574.533, "end": 1576.466, "text": "一个更小模型参加竞争", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 673, "start": 1576.8, "end": 1579.1, "text": "还得同时支持长上下文", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 674, "start": 1579.266, "end": 1581.966, "text": "视觉工具调用和代码代理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 675, "start": 1582.166, "end": 1583.6, "text": "能力追平以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 676, "start": 1583.766, "end": 1585.6, "text": "价格又必须足够低", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 677, "start": 1585.866, "end": 1588.166, "text": "这样的竞争会加速行业整合", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 678, "start": 1588.4, "end": 1590.6, "text": "拥有算力工程和分发", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 679, "start": 1590.6, "end": 1592.7, "text": "能力的平台更容易留下", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 680, "start": 1592.733, "end": 1594.333, "text": "单纯依靠模型演示", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 681, "start": 1594.333, "end": 1596.066, "text": "的公司会越来越难", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 682, "start": 1596.133, "end": 1597.7, "text": "对海外厂商而言", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 683, "start": 1597.8, "end": 1599.566, "text": "压力来自另一种方向", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 684, "start": 1599.966, "end": 1601.566, "text": "闭源模型仍可能在", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 685, "start": 1601.566, "end": 1603.466, "text": "最高难度任务上领先", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 686, "start": 1603.566, "end": 1606.566, "text": "也拥有成熟企业合规与服务支持", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 687, "start": 1607.1, "end": 1610.766, "text": "但V4.1Flash会迫使客户进行任务拆分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 688, "start": 1610.866, "end": 1613.766, "text": "大量日常调用如果能转向便宜模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 689, "start": 1613.8, "end": 1615.333, "text": "高价厂商只能从最", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 690, "start": 1615.333, "end": 1617.3, "text": "复杂的少数请求中赚钱", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 691, "start": 1617.666, "end": 1620.333, "text": "他们需要证明的不再是模型最好", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 692, "start": 1620.366, "end": 1622.966, "text": "而是领先部分足以抵消价格差", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 693, "start": 1623.266, "end": 1625.366, "text": "Deepseek这次没有发布一款", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 694, "start": 1625.366, "end": 1627.2, "text": "真正意义上的小模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 695, "start": 1627.533, "end": 1630.333, "text": "他发布的是一台把庞大参数藏", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 696, "start": 1630.333, "end": 1633.0, "text": "在稀疏计算后面的工业机器", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 697, "start": 1633.1, "end": 1635.8, "text": "5,520亿参数提供容量", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 698, "start": 1635.9, "end": 1639.333, "text": "80亿和160亿激活规模控制成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 699, "start": 1639.533, "end": 1642.166, "text": "压缩缓存负责承载长期任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 700, "start": 1642.333, "end": 1645.333, "text": "原生视觉负责进入真实电脑环境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 701, "start": 1645.733, "end": 1647.666, "text": "最后再用API低价把这些", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 702, "start": 1647.666, "end": 1649.533, "text": "能力送进开发者工具", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 703, "start": 1649.7, "end": 1651.866, "text": "开头那个问题现在有了答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 704, "start": 1652.133, "end": 1655.5, "text": "V4.1Flash能够暂时接管Pro", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 705, "start": 1655.566, "end": 1657.566, "text": "靠的是把旗舰能力装进", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 706, "start": 1657.566, "end": 1659.466, "text": "更低成本的运行结构", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 707, "start": 1659.933, "end": 1662.3, "text": "Flash这个名称并没有失去意义", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 708, "start": 1662.533, "end": 1665.2, "text": "Deepseek想淘汰一款旧模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 709, "start": 1665.333, "end": 1667.566, "text": "也在挑战能力越强就", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 710, "start": 1667.566, "end": 1669.4, "text": "一定越贵的行业惯性", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 711, "start": 1669.5, "end": 1671.133, "text": "这场测试才刚开始", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 712, "start": 1671.4, "end": 1673.5, "text": "官方基准可以证明方向", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 713, "start": 1673.566, "end": 1675.666, "text": "生产流量才能证明结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 714, "start": 1675.933, "end": 1678.7, "text": "如果9月14日以后的Pro请求真的在", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 715, "start": 1678.7, "end": 1681.766, "text": "用户无感的情况下由Flash完成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 716, "start": 1682.1, "end": 1685.3, "text": "V4.1就会成为一次重要的架构转折", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 717, "start": 1685.466, "end": 1686.8, "text": "如果质量回退", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 718, "start": 1687.0, "end": 1689.933, "text": "长任务失控或视觉代理频繁犯错", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 719, "start": 1689.966, "end": 1691.566, "text": "这次激进迁移也会", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 720, "start": 1691.566, "end": 1693.0, "text": "暴露新架构的边界", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 721, "start": 1693.4, "end": 1695.133, "text": "Deepseek把最强的承诺", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 722, "start": 1695.133, "end": 1696.6, "text": "写进了路由系统", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 723, "start": 1696.933, "end": 1699.733, "text": "宣传页面上的文字反倒居于其次", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 724, "start": 1700.166, "end": 1702.7, "text": "便宜模型要接管昂贵模型的工作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 725, "start": 1702.866, "end": 1704.8, "text": "最终裁判不会是排行榜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 726, "start": 1705.133, "end": 1707.0, "text": "裁判是每一次完成的任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 727, "start": 1707.066, "end": 1708.6, "text": "每一笔实际账单", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 728, "start": 1708.666, "end": 1711.566, "text": "以及企业敢不敢把关键流程交给他", "source": "video_source_subtitle", "authority": "video_source_subtitle" } ], "transcript_selection": { "schema": "videonote.transcript_selection.v1", "status": "SELECTED", "requested_language": "zh", "duration": 1712.0, "reference_path": "/Users/sagawa/AI_WORK/video_notes/url/20260911_151335__Deepseek_瘋了_最強旗艦_Pro_竟然被自己下架_AI_界歷史性的一天_OpenAI_這次真的要跪了.../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": 729, "duration_coverage": 0.926347 } } }