英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled


AI free token/AI免费积分
0

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

Apple 智能备案新增 iPad 和笔记本

英伟达逆势大跌 4.27%,一夜蒸发 1.29 万亿元

曝苹果调整 iOS 28 开发节奏,首阶段留出六周探索能力

曝玛莎拉蒂拟与尊界合作,新车将由江淮生产

️ Anthropic 发布 Claude Fable 和 Mythos 5.2,两款模型 API 降价

✨ 谷歌发布两套第八代 TPU 平台,分别面向长上下文和保密推理

迪士尼与 Adobe 达成多年合作,将为旗下 IP 定制生成模型

詹姆斯·古恩将卸任 DC 影业联席 CEO,前漫威高管接掌创意

瑞银 CEO:AI 股回调前,市面一直按「完美情景」定价

比亚迪第 1900 万辆新能源汽车下线,方程豹钛 9 同日开订

千问办公 2.0 发布,团队可共享资料并让 AI 按消息触发任务

英伟达推出 GR00T 2,让机器人从人类示范中学习精细操作

✨ 字节跳动发布 Seedream 5.5,生成图片可直接带透明背景

‍ LG 免费开放 K-EXAONE 2.5 模型 API

⚡ 百度披露 ERNIE Evolve,已用于优化文心 5.5 Flash 推理

Meta 推出小模型 Muse Ember,专门处理 AI 助手的重复任务

门店贴出「买烟优先结账」,零食有鸣客服称会核实

️ 12 月起,进口小客车可在「交管 12123」网上办结上牌

Netflix 推进《卡坦岛》真人电影,据报将由尼兄弟编剧执导

郑雨盛主演武打片《杀生簿》国际销售权由 Finecut 拿下

Apple 智能备案新增 iPad 和笔记本

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

网信上海 11 月 9 日发布公告,Apple 智能生成式人工智能服务新增苹果平板和笔记本电脑两类适用情形,已完成备案。变更后的适用范围涵盖苹果手机、平板和笔记本电脑。

公告所列备案主体为苹果工艺开发(上海)有限公司,原备案日期为 7 月 8 日,此次情形变更日期为 11 月 9 日。


英伟达逆势大跌 4.27%,一夜蒸发 1.29 万亿元

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

美东时间 11 月 9 日,英伟达股价收跌 4.27%,报 178.67 美元,单日市值缩水约 1921 亿美元,折合人民币约 1.29 万亿元。当下,英伟达总市值约为 4.31 万亿美元。

当天美股三大股指集体上涨,其中纳斯达克综合指数涨幅达 2.74%,英伟达却逆势下跌。盘中其股价一度跌至 170.24 美元,较前一交易日收盘价跌幅接近 9%。

此次股价下跌正值英伟达管理层面临调查。据路透社报道,英伟达董事会于 11 月 8 日宣布启动独立调查,以核实 CEO 黄仁勋是否曾成为外国情报活动的目标。调查由独立董事委员会及外部法律顾问负责。

英伟达在官方声明中强调,当下没有理由认为美国国家安全或公司商业机密受到损害。董事会仍对黄仁勋及其管理团队保持信心,黄仁勋本人也承诺配合调查。

路透社指出,调查可能加剧投资者对英伟达管理层稳定性的担忧。与此同时,市面对 AI 基础设施投资可持续性的疑虑也持续影响科技股表现。

此前,英伟达股价已于 11 月 5 日大跌 9.26%,虽然次日反弹 2.78%,但最新一轮下跌使其股价再度承压。

曝苹果调整 iOS 28 开发节奏,首阶段留出六周探索能力

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

据彭博社 Mark Gurman 报道,苹果已调整下一代操作平台 iOS 28 的开发流程,首个开发阶段(M1)留出六周用于编码、探索和能力开发,随后仅有两周用于整合。

报道称,苹果当下已深入推进明年的新平台开发,其中 iOS 28 和 iPadOS 28 代号为 Bell,macOS 28 代号为 Poppy。

按照惯例,苹果每年的新平台开发分为 M1、M2、M3 三个主要阶段,每个阶段持续八周,包含四周编码和四周整合,整合指的是将各项能力集成进操作平台。

此次 M1 阶段,苹果没有沿用两个四周的常规安排,而是让工程师先用六周开展编码、探索和能力开发,再用两周完成整合。报道提到,有员工表示喜欢这一变化,认为这样能在能力锁定前留出更多时间构思和尝试。

报道同时提到,苹果新任 CEO John Ternus 对公司 AI 进展并不满意,近几周已与软件负责人 Craig Federighi 举行闭门会议,希望其加快 AI 路线图。

曝玛莎拉蒂拟与尊界合作,新车将由江淮生产

据第一财经 11 月 9 日报道,玛莎拉蒂母公司 Stellantis 正与华为、江淮推进一项名为「玛莎拉蒂 Design」的新车项目。报道称,新车外观已经确定,计划在 2027 年第四季度迎来进一步的量产进展。

知情人士称,项目将由江淮负责制造,华为参与给出兼容。新车计划在国内由尊界销售,在海外沿用玛莎拉蒂标识。

Anthropic 发布 Claude Fable 和 Mythos 5.2,两款模型 API 降价

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

Anthropic 发布 Claude Fable 5.2 和 Mythos 5.2。两款机型共用同一个底层模型,开放范围由安全设置区分:Fable 会拦截渗透测试、漏洞利用代码生成及多数生物研究类请求,符合条件的网络安全或生命科学研究机构可申请使用 Mythos。

Anthropic 称,新模型生成速度较 5.1 版最高快 30%。标准 API 售价降至每百万输入 token 8 美元、每百万输出 token 40 美元。Fable 5.2 已登陆 Claude 平台及 AWS、Google Cloud 和 Microsoft Azure。

在发布前测试中,Anthropic 用模型把费马大定理的既有证明写成了可由程序检查的形式化证明。在公司的 FrontierSWE v2 编程实测中,Fable 5.2 得分为 69.9%,Fable 5.1 为 56.3%。

谷歌发布两套第八代 TPU 平台,分别面向长上下文和保密推理

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

谷歌 11 月 9 日公布 TPU 8i Max 和 TPU 8tx,扩充今年 4 月亮相的第八代 TPU 阵容。两套平台沿用现有芯片,分别调整了芯片封装和配套硬件。

TPU 8i Max 把两枚 TPU 8i 芯片封装在一起。谷歌称,相比 TPU 8i,新平台的算力、片上 SRAM 和高带宽内存带宽均翻倍,用于处理整套代码库或长视频等长上下文任务。

TPU 8tx 使用与 TPU 8t 相同的芯片,搭配定制 AMD 主机处理器。硬件在推理时隔离敏感数据和模型权重,企业也可以在同一套平台上用私有数据微调模型。

迪士尼与 Adobe 达成多年合作,将为旗下 IP 定制生成模型

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

迪士尼与 Adobe 宣布达成多年合作。Adobe 将向迪士尼提前开放最新的创作和媒体工具,并给出工程与研究资源。双方将用 Firefly Foundry 开发迪士尼专用的生成模型,把角色特征、情形和视觉风格纳入模型训练。

迪士尼的幻想工程研发团队已在主题乐园开发中使用这套工具,例如将手绘草图转成完整的二维概念图。合作还覆盖华特迪士尼影业的 StudioLAB 和 ESPN Edge,后者将探索用 AI 整理和利用体育档案,并尝试新的内容展现方式。

詹姆斯·古恩将卸任 DC 影业联席 CEO,前漫威高管接掌创意

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

Skydance 宣布,曾参与搭建漫威电影宇宙的杰里米·拉查姆将于 12 月 1 日出任 DC 影业总裁兼首席创意官,统筹电影、电视和动画业务。詹姆斯·古恩与彼得·萨夫兰将卸下管理职务,回归创作和制片。

拉查姆曾参与制作《钢铁侠》《复仇者联盟》和《银河护卫队》。古恩编剧并执导的超人续集《Man of Tomorrow》当下处于后期制作,计划于 2027 年 7 月 9 日上映,古恩和萨夫兰仍共同担任制片人。

瑞银 CEO:AI 股回调前,市面一直按「完美情景」定价

据路透社报道,瑞银首席执行官塞尔吉奥·埃尔莫蒂 11 月 9 日在悉尼出席瑞银会议时表示,过去数周 AI 股票的显著回调,反映出投资者此前对行业的预期过高。他仍看好 AI 带来的经济收益,但认为市面高估了这些收益兑现的速度。

直到不久前,市面一直按完美情景定价。当然,过去几周笔者看到了一些回调。

埃尔莫蒂提到,美国中期选举结果,以及全球围绕 AI 基础设施建设影响的争论,都可能拖慢行业取得成果的进度。

他也以瑞银为例说明企业落地的速度。瑞银在 2023 年同意收购瑞信,埃尔莫蒂表示,随后的整合工作拖慢了瑞银内部部分 AI 应用的落地。

比亚迪第 1900 万辆新能源汽车下线,方程豹钛 9 同日开订

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

11 月 9 日,比亚迪第 1900 万辆新能源汽车在西安工厂下线,下线车型为方程豹钛 9。同一天,方程豹启动钛 9 的全国预订,并公布顶配版规格。

钛 9 是一款使用方正车身的大六座插混 SUV,搭载比亚迪 DMS 混动平台和第二代刀片电池,配备云辇-A 车身控制平台。顶配版增加了 1.1 米全景抬头表明和 30 英寸座舱屏。新浪汽车报道称,新车将于 11 月内正式上市。

千问办公 2.0 发布,团队可共享资料并让 AI 按消息触发任务

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

千问办公 11 月 9 日发布 2.0 版。团队可以在协作空间里关联钉钉群聊、听记、表格和文档,管理并指派任务。成员借助即时通讯交办后,数字员工「小 Q」可以结合空间里的资料梳理项目进展、设置待办。成员也可以查看、纠正或停用空间记忆。

新版兼容按消息触发自动任务,首批规则涵盖收到指定联系人发来的消息,以及在指定群里被 @。纯云端任务在电脑关机后仍可运行,使用者可以在网页端或桌面端修改、停用规则,并查看执行结果和失败原因。

修改 AI 生成的内容时,使用者可以选中具体的段落、单元格、幻灯片或应用区域,一次提交多处修改意见,由 AI 在原有成果上直接修改。

英伟达推出 GR00T 2,让机器人从人类示范中学习精细操作

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

英伟达在 11 月 9 日的机器人学习大会 CoRL 上公布 GR00T 2、训练框架 Isaac Lab 3.0,以及两套面向工业操作的开放实测。新模型基于 NVIDIA Cosmos,能从人类示范中学习,也能把技能迁移到不同的机器人本体上。

英伟达称,在其评估的 50 项任务中,GR00T 2 没看过对应任务示范时,有 40 项至少成功完成过一次。每项给出一次示范后,平均成功率为 71%,增加到五次示范后为 95%。

Isaac Lab 3.0 可用 AI 智能体搭建训练环境,并在 PhysX 和 Newton 两套物理求解器之间切换。同期公布的 DexBench 覆盖零售、物流、制造和酒店服务四类情形,VABAR 则为六项工业任务给出标准化的条件变化、难度和成功指标。

字节跳动发布 Seedream 5.5,生成图片可直接带透明背景


字节跳动 Seed 团队 11 月 9 日发布图像模型 Seedream 5.5。新模型可直接生成带 Alpha 通道的图片,透明和半透明的部分可以保留下来,生成的元素可直接用作独立外观素材。

团队也展示了同一张人物图连续编辑 100 轮的结果。每轮以前一张图为底稿,先后修改服装、配饰和桌面物件,到后段仍能单独调整戒指、镜框等细节。Seed 称,这一版针对局部编辑做了专项训练,以减少未修改区域在反复生成中的偏色和细节流失。

新版也改善了光照、材质、皮肤纹理和密集小字的渲染,正陆续在即梦、豆包、豆包工作和火山方舟上线。

LG 免费开放 K-EXAONE 2.5 模型 API

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

LG AI 研究院 11 月 9 日在 EXAONE API Platform 开放 K-EXAONE 2.5。开发者、初创公司、企业及研究机构可以申请 API 密钥,免费调用模型,用于开发服务和研究。

K-EXAONE 2.5 是在 LG 今年 8 月公布的版本上追加训练的工艺验证模型,面向代码编写、网页检索和工作流自动化等任务。LG 公布了基于公开数据集的自测结果,其中评估真实软件开发任务的 SWE-Bench Pro 得分为 59.2,评估网页信息检索的 BrowseComp 得分为 78.0。

百度披露 ERNIE Evolve,已用于优化文心 5.5 Flash 推理

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

百度文心 11 月 9 日公开了自动研究平台 ERNIE Evolve,并披露已用它优化文心 5.5 Flash 的推理效率。平台会依据历史实验结果提出研究方向,再由多个智能体分工挑选代码起点、修改代码。每轮改动须借助独立实测和复核,实验代码和结果会留给后续研究使用。

百度在工艺博客中列出了三类实际研发任务,涵盖昆仑芯 P800 上的低时延推理、GPU 上的确定性推理吞吐,以及单并发解码速度。在 P800 的低时延任务中,团队限定改动范围为芯片适配插件,保持模型和部署方式不变。平台先后尝试融合算子、投机解码以及调度和通信优化,最优版本的生成速度达到初始基线的 623.6%,并借助了团队的正确性检查。

百度称,接下来计划将这套方法用于训练平台优化、模型能力研究和实测体系迭代,并继续限制平台的权限和资源范围。

Meta 推出小模型 Muse Ember,专门处理 AI 助手的重复任务

Meta 发布多模态模型 Muse Ember,面向需要反复执行的任务,例如编程中的小步骤、工具调用、文档摘要和字段提取。它是 Meta 首个使用 Muse Wonder Full 训练的模型,已在 Muse Code 和 Meta Model API 上线,兼容 Responses API。

Muse Ember 兼容 100 万 token 的上下文。提示不超过 20 万 token 时,API 每百万输入 token 收费 0.10 美元,每百万输出 token 收费 0.50 美元。

门店贴出「买烟优先结账」,零食有鸣客服称会核实

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

近日,一家零食有鸣门店贴出「购买香烟者优先结账」的告示,引发消费者对排队规则的质疑。11 月 9 日,零食有鸣客服向界面新闻表示,公司没有就此出台文件,客服会把线索转交相应监督部门核实。

客服还介绍,门店售卖卷烟须先取得烟草专卖零售许可证,货源由门店向当地烟草公司采购,品牌总部不参与进货。对于加盟门店的经营难题,品牌方有权介入管理。

12 月起,进口小客车可在「交管 12123」网上办结上牌

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

公安部交通管理局宣布,自 12 月 1 日起,购买进口小客车的车主可借助「交管 12123」App 在线办结新车选号、登记上牌业务,无需前往窗口。交管部门将在共享车辆销售发票、购置税和保险信息的基础上,接入进口小客车预查验信息。

同批措施还允许汽车生产企业网上申领试验车临时号牌。企业出示电子营业执照,即可办理新车上牌、二手车过户等业务,并可借助 App 网上注册,查询企业名下车辆状态及违法事故信息。

直辖市、省会城市、自治区首府及计划单列市将试点 App 智能客服「交管小安」,兼容边咨询边办理业务。

Netflix 推进《卡坦岛》真人电影,据报将由尼兄弟编剧执导

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

据 Deadline 报道,曾执导《迷失之城》的亚伦·尼和亚当·尼将为 Netflix 编剧并执导《卡坦岛》真人电影。这是 Netflix 去年拿下这款桌游的全球改编权后,首个披露主创阵容的真人电影项目。

《卡坦岛》由德国桌游外观师克劳斯·托伊伯创作,玩家在岛上争取有限资源、结盟交易并建设定居点。Netflix 此前宣布与桌游发行商 Asmodee 合作,开发以《卡坦岛》为基础的真人、动画和非虚构内容。

郑雨盛主演武打片《杀生簿》国际销售权由 Finecut 拿下

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

据 Deadline 和 Variety 报道,韩国电影销售公司 Finecut 拿下武打片《杀生簿》(The Hit List)的国际销售权。影片由执导过《狼狩猎》和《伦敦黑帮》第三季的金泓善执导,已于今年 10 月开拍。

影片改编自 Kakao 平台一部浏览量超过 1.2 亿次的网漫。故事发生在朝鲜王朝,《首尔之春》演员郑雨盛饰演权臣韩明浍,《鱿鱼游戏》演员魏嘏隽饰演首阳大君,后者借韩明浍之力争夺年幼侄子的王位。

《伦敦黑帮》第三季的摄影指导拉斯·考克斯和灯光师基兰·韦茨也加入了本片。

笔者正在招募伙伴

简历投递邮箱hr@ifanr.com

✉️ 邮件标题「姓名+岗位名称」(请随简历附上项目/作品或相关链接)

英伟达市值蒸发1.29万亿元;千问办公2.0、GR00T 2、Seedream 5.5 与谷歌 TPU 发布

本文来源:https://www.163.com/dy/article/LBC0N31Q0511CSAO.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

0

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

Apple Intelligence filing expanded to include iPad and notebooks

Nvidia plunges 4.27% against the trend, wiping out 1.29 trillion yuan overnight

Apple reportedly adjusts iOS 28 development pace, carving out a six-week exploration phase

Maserati reportedly plans to partner with Zunjie, with new models to be produced by JAC

Anthropic releases Claude Fable and Mythos 5.2, cutting API prices for both models

Google unveils two 8th-gen TPU platforms, targeting long-context and confidential inference respectively

Disney and Adobe announce a multi-year partnership to create custom generative models for its IPs

James Gunn to step down as co-CEO of DC Films; former Marvel executive to head creative

UBS CEO: Before the AI stock correction, the market was pricing in a "perfect scenario"

BYD produces its 19 millionth NEV; orders for the Fang Cheng Bao Tai 9 open on the same day

Qianwen Office 2.0 released, allowing teams to share data and have AI trigger tasks from messages

Nvidia introduces GR00T 2, enabling robots to learn fine-grained operations from human demonstrations

ByteDance releases Seedream 5.5, generating images with transparent backgrounds directly

LG opens K-EXAONE 2.5 model API for free

Baidu unveils ERNIE Evolve, used to optimize ERNIE 5.5 Flash inference

Meta launches Muse Ember, a small model tailored for repetitive tasks of AI assistants

Store posts "Priority checkout for cigarette buyers"; Keshangfeng customer service says it will verify

Starting in December, registration for imported passenger cars can be completed online via "Traffic Management 12123"

Netflix moves forward with live-action "Catan" movie, reportedly to be written and directed by the Nee brothers

Finecut secures international sales rights for martial arts film "Kill Manual," starring Jung Woo-sung

Apple Intelligence filing expanded to include iPad and notebooks

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

On November 9, the Cyberspace Administration of Shanghai announced that the filing for Apple's generative artificial intelligence service, Apple Intelligence, has added two new applicable categories: Apple tablets and notebooks, and the filing has been completed. The updated scope of application now covers Apple mobile phones, tablets, and notebooks.

The filing entity listed in the announcement is Apple Process Development (Shanghai) Co., Ltd., with the original filing date being July 8, and the date of this situation change being November 9.

Nvidia plunges 4.27% against the trend, wiping out 1.29 trillion yuan overnight

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

On November 9, Eastern Time, Nvidia's stock closed down 4.27% at $178.67, with its market capitalization shrinking by approximately $192.1 billion in a single day, equivalent to about 1.29 trillion RMB. Currently, Nvidia's total market capitalization is approximately $4.31 trillion.

On the same day, all three major U.S. stock indices rose collectively, with the Nasdaq Composite Index gaining 2.74%, while Nvidia fell against the trend. During the session, its share price once dropped to $170.24, a decline of nearly 9% from the previous trading day's closing price.

Shares of Nvidia fell as management faced an investigation. Reuters reported that on November 8 Nvidia's board announced an independent investigation to determine whether CEO Jensen Huang has been the target of foreign intelligence operations. The investigation is conducted by an independent director committee with outside legal counsel.

Nvidia's official statement stressed there was no reason to believe U.S. national security or company trade secrets shall be compromised. The board retains confidence in Jensen Huang and his management team, Huang said he vowed to turn up everything to the investigation.

Reuters reported that the probe could intensify investor concern regarding stability of Nvidia 's management. Meanwhile, market concerns about sustainability of AI infrastructure spending continue to weigh on tech equities.

Prior to this, on November 5 Nvidia saw its share price tumble 9.26 percent, only rebound 2.78 percent the next second. The latest decline has this again put pressure on value.

Reports suggest Apple changes iOS 28 development cycle, leaving a six-week window to explore capabilities

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

According to Bloomberg reporter Mark Gurman, Apple changed the way it develops its next-gen software iOS 28, leaving six weeks for coding, exploration, and capability development, and only two weeks to integration.

The report added that Apple has already progressed with next year's software builds: iOS 28 and iPadOS 28 are codenamed "Bell," while "macOS 28" is reportedly called "Poppy."

In typical years Apple's annual software development cycle is split into three major phases —M1, M2, and M3—each lasts 8 weeks, consisting of four weeks of coding and four weeks of consolidation, or platforms with different pieces integrated together.

This year's M1 phase, Apple has moved from two four-week periods, letting engineers spend six weeks coding and exploring capabilities then use two weeks to finish integration. Some employees like the new arrangement because it gives them more time to brainstorm and experiment before features are locked.

The report also said Apple's new CEO John Ternus is not satisfied with the company's AI progress, and in recent weeks has held closed-door meetings with software chief Craig Federighi to urge acceleration of the AI roadmap.

Reports: Maserati may team up with Zaojie; JAC Motors to make the car

According to a story by China Business Today on November 9, Maserati parentStellantis is in advanced talks with Huawei and JAC Motors about a new car project named "Maserati Design." The shape of the new car has already been finalized, with further mass production planned for Q4 2027.

People familiar with the matter say JAC Motors will handle manufacturing. While Huawei has agreed to carry Maserati nameplate overseas, the new car could be sold domestically under the Zaojieator brand.

Anthropic launches Claude Fable and Mythos 5.2, cuts API prices

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

Anthropic announced Claude Fable 5.2 and Mythos 5.2. Both are built on the same underlying base model, with safety settings determining access: Fable will block prompts related to penetration testing, exploit code generation and most bio-research requests, whereas vetted cybersecurity and life-science research organizations will be able to request the more permissive Mythos variant.

Anthropic claims the new models deliver generation speeds up to 30% faster than version 5.1. Standard API price will be set at $8 per million input tokens and $40 per million output tokens. Fable 5.2 is available on Claude and AWS, Google Cloud and Microsoft Azure.

In pre-release testing, Anthropic used the model to produce a proof of Fermat's Last Theorem in a machine-checkable format. Anthropic also said Fable 5.2 earned 69.9% on its FrontierSWE v2 coding test, where 5.1 scored 56.3%.

Google debuts second gen-8 TPU platform, with variants for long-context and secure inference

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

On November 9, Google unveiled TPU 8i Max and TPU 8tx, expanding the Gen-8 TPU line that debuted in April. Both are built on Google's existing silicon, with changes to package and supporting hardware.

TPU 8i Max pairs two TPU 8i chips in package. According to Google, the new system doubles compute, on-chip SRAM and HBM memory bandwidth, making it suitable for handling long-context jobs including full codebases or long-form videos.

TPU 8tx uses the same chip as TPU 8t but adds an AMD-designed host processor. During inference, system isolates sensitive data and model weights, while enterprises can also fine-tune models on private data within same platform.

Disney inks multi-year partnership with Adobe, ready to custom generative models for its properties

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

Disney and Adobe have entered multi-year agreement, Adobe will give Disney early access to latest creative tools along with engineering and research resources. Using Firefly Foundry, the two companies will develop custom generative models with Disney's characters, situations and visual style included during training.

Disney's Imagineering R&D team has used the tools at its theme parks to turn hand-drawn rough sketches into finished 2D concept paintings. The deal also covers Walt Disney Studios' StudioLAB and ESPN Edge, where ESPN plans to explore AI tools for organizing and reusing the company's sports archive, along with new ways of presenting content.

James Gunn to vacate co-CEO of DC Studios; Marvel Studios executive takes over creative leadership

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

Skydance announced that Jeremy Latcham, who was involved in building the Marvel Cinematic Universe, will become the President and Chief Creative Officer of DC Studios on December 1, overseeing film, television, and animation operations. James Gunn and Peter Safran will step down from their administrative roles to return to creation and production.

Latcham previously participated in the production of "Iron Man," "The Avengers," and "Guardians of the Galaxy." The Superman sequel "Man of Tomorrow," written and directed by Gunn, is currently in post-production and is scheduled to be released on July 9, 2027, with Gunn and Safran still serving as producers together.

UBS CEO: Before the AI stock correction, the market was pricing for a "perfect scenario"

According to Reuters, UBS CEO Sergio Ermotti stated while attending the UBS conference in Sydney on November 9 that the significant correction in AI stocks over the past few weeks reflects investors' previously overly high expectations for the industry. He remains optimistic about the economic benefits brought by AI, but believes the market has overestimated the speed at which these benefits will materialize.

Until recently, the market had been pricing for a perfect scenario. Of course, the author has seen some corrections in the past few weeks.

Ermotti mentioned that the outcome of the US midterm elections, along with global debates surrounding the impact of AI infrastructure construction, could potentially slow down the progress of the industry in achieving results.

He also cited UBS as an example to demonstrate the speed of enterprise implementation. UBS agreed to acquire Credit Suisse in 2023, and Ermotti stated that the subsequent integration work delayed the implementation of some internal AI applications at UBS.

BYD’s 19 millionth new energy vehicle rolls off the production line; Fang Cheng Bao Tai 9 bookings start the same day

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

On November 9, BYD's 19 millionth new energy vehicle rolled off the production line at the Xi'an factory, marking the completion of the Fang Cheng Bao Tai 9. On the same day, Fang Cheng Bao started nationwide reservations for the Tai 9 and released the specifications for the top-tier version.

The Tai 9 is a large six-seater plug-in hybrid SUV with a boxy body. It is equipped with BYD's DMS hybrid platform and second-generation Blade Battery, as well as the Cloud Chassis-A body control platform. The top-tier version adds a 1.1-meter panoramic head-up display and a 30-inch in-cabin screen. Sina Auto reports that the new vehicle will officially launch within November.

Qianwen Office 2.0 released; teams can share materials and let AI trigger tasks based on messages

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

Qianwen Office released version 2.0 on November 9. Teams can associate DingTalk group chats, meeting notes, spreadsheets, and documents in a collaboration space to manage and assign tasks. Once members assign tasks via instant messaging, the digital employee "Xiao Q" can use the materials in the space to organize project progress and set up to-do lists. Members can also view, correct, or disable space memory.

The new version supports automatic tasks triggered by messages. The first batch of rules includes receiving messages from specified contacts and being @-mentioned in designated groups. Pure cloud-based tasks can continue to run even after the computer is shut down, and users can modify or disable rules and view execution results and failure reasons via the web or desktop client.

When modifying AI-generated content, users can select specific paragraphs, cells, slides, or application areas and submit multiple revision suggestions at once, allowing the AI to directly modify the original output.

NVIDIA launches GR00T 2, enabling robots to learn fine motor skills from human demonstrations

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

NVIDIA announced GR00T 2, the training framework Isaac Lab 3.0, and two sets of open real-world tests for industrial manipulation at CoRL, the Conference on Robot Learning, on November 9. The new model is based on the NVIDIA Cosmos platform; it can learn from human demonstrations and transfer skills to different robot embodiments.

NVIDIA stated that across 50 tasks evaluated, GR00T 2 successfully completed 40 tasks at least once without seeing demonstrations for those specific tasks. With one demonstration provided per task, the average success rate was 71%, and this increased to 95% when demonstrations were increased to five.

Isaac Lab 3.0 allows AI agents to build training environments and switch between two physics solvers: PhysX and Newton. The DexBench released at the same time covers four scenarios: retail, logistics, manufacturing, and hospitality services; VABAR provides standardized condition variations, difficulty, and success metrics for six industrial tasks.

ByteDance releases Seedream 5.5; generated images can be directly output with transparent backgrounds

ByteDance's Seed team released the image model Seedream 5.5 on November 9. The new model can directly generate images with an Alpha channel, preserving transparent and semi-transparent parts, so that generated elements can be directly used as independent appearance assets.

The team also demonstrated the results of 100 consecutive editing rounds on the same portrait. Using the previous image as a base, each round modified clothing, accessories, and desktop items, and later stages could still individually adjust details such as rings and eyeglass frames. Seed stated that this version has undergone specialized training for local editing to reduce color shifts and detail loss in unmodified areas during repeated generation.

The new version also improves the rendering of lighting, materials, skin textures, and dense small text. It is currently being rolled out on platforms including Jimeng, Doubao, Doubao Office, and Volcano Engine Ark.

LG opens K-EXAONE 2.5 model API for free

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

On November 9, LG AI Research Institute released K-EXAONE 2.5 on the EXAONE API Platform. Developers, startups, enterprises, and research institutions can apply for API keys to use the model free of charge for service development and research.

K-EXAONE 2.5 is a process validation model with additional training based on the version announced by LG this August, targeting tasks such as coding, web search, and workflow automation. LG has released self-test results based on public datasets, including a score of 59.2 on SWE-Bench Pro, which evaluates real software development tasks, and a score of 78.0 on BrowseComp, which evaluates web information retrieval.

Baidu discloses ERNIE Evolve, which has already been used to optimize the inference of ERNIE 5.5 Flash.

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

Baidu Wenxin publicly launched the automated research platform ERNIE Evolve on November 9, and disclosed that it had used it to optimize the inference efficiency of Wenxin 5.5 Flash. The platform proposes research directions based on historical experimental results, then multiple agents divide tasks to select code starting points and modify code. Each round of changes must be verified through independent testing and review; experimental code and results are retained for subsequent research use.

In a process blog, Baidu listed three categories of actual R&D tasks, covering low-latency inference on Kunlun P800, deterministic inference throughput on GPUs, and single-concurrency decoding speed. For the low-latency task on P800, the team limited the scope of changes to the chip adapter plugin, keeping the model and deployment method unchanged. The platform attempted operator fusion, speculative decoding, and scheduling and communication optimization in succession; the optimal version achieved 623.6% of the initial baseline generation speed, and passed the team's correctness checks.

Baidu stated that it plans to apply this methodology to training platform optimization, model capability research, and iteration of the testing system, and will continue to restrict the platform's permissions and resource scope.

Meta introduces small model Muse Ember, specifically designed for handling repetitive tasks of AI assistants

Meta released the multimodal model Muse Ember, targeting tasks that need repeated execution, such as small steps in programming, tool calls, document summarization, and field extraction. It is Meta's first model trained using Muse Wonder Full, and is already available on Muse Code and Meta Model API, compatible with the Responses API.

Muse Ember supports a context window of 1 million tokens. When prompts do not exceed 200,000 tokens, the API costs $0.10 per million input tokens and $0.50 per million output tokens.

Store posts "cigarette buyers get priority checkout"; Lingshouyouming customer service says it will verify

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

Recently, a Lingshouyouming store posted a notice saying "cigarette buyers get priority checkout," sparking consumer questioning of the queuing rules. On November 9, Lingshouyouming customer service told Jiemian News that the company has not issued any documents regarding this, and that customer service will forward the information to the relevant supervisory department for verification.

Customer service also explained that stores selling cigarettes must first obtain a tobacco monopoly retail license; shipments are purchased by the store from the local tobacco company, and the brand headquarters does not participate in procurement. For operational difficulties in franchise stores, the brand has the right to intervene in management.

Starting in December, owners of imported small passenger cars can complete online registration via "Jiaoguan 12123"

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

The Traffic Management Bureau of the Ministry of Public Security announced that starting December 1, owners purchasing imported small passenger cars can use the "Jiaoguan 12123" app to complete online new-car number selection and registration, without needing to visit a service window. Traffic management authorities will add pre-inspection data for imported small passenger cars, building on the shared information of vehicle sales invoices, purchase tax, and insurance.

The same batch of measures also allows automobile manufacturers to apply online for temporary license plates for test vehicles. Enterprises can complete procedures such as new-car registration and used-car ownership transfer by presenting electronic business licenses, and can use the app for online registration to check the status and violation/accident information of vehicles registered under the enterprise's name.

Municipalities, provincial capitals, autonomous region capitals, and cities with separate planning status will pilot the app's intelligent customer service "Jiaoguan Xiao'an," which supports consulting while simultaneously handling business transactions.

Netflix advances "Catan" live-action film, reportedly to be written and directed by the Nee brothers

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

According to Deadline, Aaron Nee and Adam Nee, who previously directed "The Lost City," will write and direct the "Catan" live-action film for Netflix. This is the first live-action film project to disclose its creative team since Netflix acquired the global adaptation rights to the board game last year.

"Catan" was created by German board game designer Klaus Teuber; players compete for limited resources, form alliances, trade, and build settlements on the island. Netflix previously announced a partnership with board game publisher Asmodee to develop live-action, animated, and non-fiction content based on "Catan."

Jung Woo-sung stars in martial arts film "The Hit List"; Finecut acquires international sales rights

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

According to Deadline and Variety, Korean film sales company Finecut has acquired the international sales rights to the martial arts film "The Hit List." The film is directed by Kim Hong-sun, who previously directed "Wolf Hunt" and Season 3 of "Gangs of London," and began shooting in October this year.

The film is adapted from a webtoon on the Kakao platform with over 120 million views. Set in the Joseon Dynasty, "Seoul Spring" actor Jung Woo-sung plays the powerful official Han Myeong-hoe, and "Squid Game" actor Wi Ha-joon plays Prince Shoyang. The latter, with the help of Han Myeong-hoe, competes for the throne of his young nephew.

Season 3 of "Gangs of London" director of photography Ross Cox and gaffer Kieran Wetz have also joined the film.

The author is looking for partners

Send resumes to hr@ifanr.com

Email subject: "Name + Position Title" (please include project/work or relevant links with the resume)

NVIDIA Sheds 1.29T Yuan; Qwen Office 2.0, GR00T 2, Seedream 5.5 and Google TPU Unveiled

/ Read the next article /

Return to NetEase Home Download NetEase News App

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry