🔧国行 Apple 智能启动测试,Siri、写作工具与视觉智能接入千问
🌦️苹果天气浮现数据异常,多地下周五「跌至」零下 32 度
🧑💻苹果向所有开发者开放 Siri AI,可将多个 App 动作串成任务
🧠GPT 爆改 Qwen 权重,团队测试数学成绩提高约 6 倍
🤝OpenAI 用 Astra 一周外观语音 Codex 设备,并与马自达合作研究
💬曝 DeepSeek V5 拟用昇腾预训练
💼DeepSeek 又崩了
🏢红果使用者规模超越豆包,千问领跑亿级 App 增速榜
🖥️盗用 AI 平台算力,上海一男子获刑三年
📱中兴否认出售手机业务传闻
🧠Claude Code 机型负责人 Cat Wu 离开 Anthropic
💬姚顺宇否认创业:仍是 Google DeepMind 员工
🎓多所高校免费发 Token,同济每周给出 1 亿 Token
🚗蔚来累计换电突破 1.3 亿次,最近 1000 万次用时 88 天
🏢Seedance 海外收入占比达一半,腾讯与 B 站成为国内大客户
📈AMD 数据中心收入接近翻倍,第四季度营收指引达 170 亿美元
💡阿里董事会主席:只有开源才能普及 AI
🔓英伟达与 Pika 基于 MiniMax H3 开源视频增强和生成模型
⚡商汤发布 SenseNova 6.8 Flash,兼容动态外观与可编辑 3D 工程
🚗法拉利首款电动车 Luce,2027 年产量已全部被预订
🖥️美团 LongCat-2.5 正式发布,国产算力训练原生多模态模型
📉智能投影销量连跌四季,第三季度中国市面下滑 26.2%
🛍️佳能计划明年起调价,应对零部件成本上涨
✨全国放心消费平台上线,扫码可查周边商家承诺
🌐《好莱坞往事》衍生片再请《社交网络》配乐搭档
🎬《蓝眼武士》第二季定档明年 1 月,复仇之旅转向伦敦
🧠Netflix 测试 AI Discover,按心情找片还能屏蔽不想看的建议
国行 Apple 智能启动测试,Siri、写作工具与视觉智能接入千问

今天凌晨,苹果推送 iOS 27.3 首个开发者测试版,国行「Apple 智能与 Siri」开始开放感受。
苹果中文官网列出的首批能力涵盖写作工具、实时翻译、照片消除、智绘表情及通知摘要。
千问被整合进 Siri、视觉智能和写作工具,使用者可在平台内提问、生成文本或处理图像。
据爱范儿实测,视觉智能的以图搜物还接入了百度搜索。Siri 换上环绕显示屏边缘的光晕,并兼容双击显示屏底部切换为文字输入。
在 Siri 中,基础问答可直接完成。以「最近总是睡不好」为例,Siri 会调用千问给出通用睡眠建议。
如需更个性化的思路,使用者可点击千问 Logo 跳转至千问 App,对话信息会自动同步。据了解,千问可结合使用者此前的睡眠记录、运动情况和生活习惯制定改善思路。
实时翻译覆盖信息、电话和 FaceTime,其中电话与 FaceTime 限一对一通话。AirPods 翻译需适配耳机更新至最新固件,并与启用 Apple 智能的 iPhone 配对。
首批能力适用于 iPhone 15 Pro 系列、iPhone 16 全系列及后续机型,以及 iPhone Air、iPhone Duo,Siri 和设备语言须设为简体中文或英语。
值得注意的是,此次国行更新给出的是首批 Apple 智能能力,WWDC 2026 预告的新版 Siri AI 未包含在内。
🔗 相关阅读:连夜实测 iPhone 国行 AI,亮点和槽点都在这了
苹果天气浮现数据异常,多地下周五「跌至」零下 32 度



向左滑动查看更多内容
今天凌晨,苹果自带天气 App 在国内多个城市浮现预报数据异常,将下周五(11 月 13 日)的最低气温错误表明为 -32℃ 左右。
受影响的区域覆盖北京、上海、南京、武汉、合肥、福州、西安、昆明等多个城市,页面表明的极端低温普遍集中在 -29℃ 至 -33℃ 区间。
例如昆明地区页面表明当日最低气温将在 19:00 至 20:00 骤降至 -32℃,而白天最高气温仍表明为 17℃ 至 24℃,单日昼夜温差接近 50℃。与此同时,广州、深圳、南宁、海口等两广及海南地区的天气预报表明正常,并未浮现类似数据。
中央气象台及各地气象部门近期并未发布寒潮预警,国内外主流数值预报模式亦未预测 11 月 13 日前后会有剧烈降温天气。
有气象博主分析指出,异常表明的时间节点恰好处于北京时间 20:00(世界协调时 12:00)的数值模式预报节点,推测系该时次的预测数据浮现偏差后经算法插值放大所致。
苹果向所有开发者开放 Siri AI,可将多个 App 动作串成任务

苹果宣布,随 iOS、iPadOS、macOS 和 visionOS 27.3 测试版开放新的 Siri AI 开发者预览,所有开发者均可借助 App Intents,将应用中的内容与操作接入新版 Siri。
接入后,Siri AI 能结合个人上下文、显示屏内容和自然语言请求执行应用内动作,例如查配送、订位或下单。苹果还引入「技能」,允许开发者把多个操作组合为一次完整感受,并让 Siri AI 检索 Spotlight 语义索引中的实体。
日历、文件、访达、邮件、地图、信息、照片等 15 款苹果应用已开始给出相关能力,例如寻找上周去过的咖啡店、归档超过一年未打开的文件。苹果称,这批接口将先供开发者探索,几个月后开始面向使用者交付,明年春季开放 App Store 提交。
当前 Siri AI 测试版要求兼容 Apple 智能的设备及受兼容语言;简体中文不在当前语言列表中,繁体中文、挪威语、土耳其语和越南语将随后加入。
GPT 爆改 Qwen 权重,团队测试数学成绩提高约 6 倍

MIT CSAIL 博士生 Rikiya Takehi 及合作者公布「权重空间编程」实验:让 GPT-6 Astra 直接读取、修改 Qwen3.5-0.8B 的权重,不使用梯度下降或训练循环。在团队的 MATH-500 测试中,修改后模型的平均准确率由 7.9% 升至 48.1%,约为原来的 6 倍。
实验向 Astra 给出权重操作用的 Python 环境,以及不含测试集题目的数学练习材料。成效最好的这一轮允许最多 50 次前向调用,供模型检查生成结果和内部状态;最终编辑涉及约 0.016% 的权重。团队称,Astra 找到并削弱了使小模型陷入重复推理的方向,集中修改第 15 层和第 19 层的部分 MLP 输出列。
同一轮测试中,GSM8K 准确率由 13.8% 升至 67.1%,IFEval 严格指令遵循准确率由 21.0% 升至 62.3%。团队已公开实验提示、代码、实测和权重差分;作者随后还分享了重建后的模型及在线演示。
🔗 相关阅读:GPT-6 Astra 爆改 Qwen 大脑:只动 0.016% 参数,数学成绩疯涨
OpenAI 用 Astra 一周外观语音 Codex 设备,并与马自达合作研究

OpenAI 公布一项硬件外观实验:两名团队成员利用 GPT-6 Astra,在不到一周的业余项目时间内完成一台可与 Codex 语音交互的设备外观,代工商随后用不到一个月制作出小批原型。
模型参与了外壳和内部结构外观、元件挑选、PCB 布局与连线,以及界面、固件和 Codex 软件集成。进入制造阶段后,团队仍开展了多轮审查与迭代,应对模型未充分考虑的生产限制、公差与装配细节。
公司同时宣布与马自达开展研究合作,让模型辅助空气动力学和碰撞安全工程师外观、模拟及评估汽车思路。与 Chip Ganassi Racing 的合作则在探索将现有零件的 3D 扫描转换为可编辑 CAD 模型和二维制造图纸,输出由车队工程师检查准确性与可制造性。
在测试三维零件重建、装配和电路板连接识别的 BenchCAD 2.0 上,OpenAI 披露一款较新的内部模型平均得分为 0.875,GPT-6 Astra 最高推理档为 0.594。
曝 DeepSeek V5 拟用昇腾预训练
《科创板日报》援引多名产业链人士消息称,DeepSeek 下一代大模型 V5 计划基于华为昇腾芯片开展预训练。此前,昇腾芯片主要承担 DeepSeek 模型的推理任务。
报道梳理,DeepSeek 于 9 月 30 日开源了面向昇腾的基础设施组件,涵盖 TileLang 编译工具、DeepGEMM 和 FlashMLA 等计算库,以及 DeepEP 分布式通信库,与此前面向英伟达平台的组件对应。
华为则给出联合定义的超节点及组网思路,并开放底层编程接口和指令兼容。
在训练侧,深圳河套学院等团队此前已依托昇腾 910C 集群完成 DeepSeek-V4-Pro 的全参数后训练;V5 传出的计划则涉及下一代模型的基础预训练。
DeepSeek 又崩了

11 月 3 日下午,DeepSeek 网页与 API 服务浮现部分中断。第一财经记者测试发现,对话能力无法使用,聊天界面提示「服务器繁忙,请稍后重试」。
官方状态信息表明,故障从 16:58 开始,受影响的涵盖对话服务、DeepSeek V4 Pro API 和 V4.1 Flash API。服务随后进入修复与监控阶段。
截至当晚 21 时许,DeepSeek 官方状态页表明「一切运行正常」,各项平台均按预期运行。
红果使用者规模超越豆包,千问领跑亿级 App 增速榜

QuestMobile 于 11 月 3 日发布秋季移动互联网报告及实力价值榜。9 月,抖音集团内部按去重总使用者量计算,前三大应用变为抖音、红果免费短剧和豆包,分别约为 10.29 亿、4.3 亿和 4.19 亿,红果已超过豆包。
千问则以约 1.58 亿月活、同比增长 5077%,位居月活过亿 App 的使用者规模增长榜首。该榜前五还涵盖抖音商城、豆包、红果免费短剧和汽水音乐。在 AI 原生 App 月活榜中,豆包、千问和 DeepSeek 仍分列前三,DeepSeek 月活约为 1.23 亿。
短剧同时吸走更多使用时长。报告表明,短剧视频总使用时长净增 133.29 亿小时,短视频净增 57.17 亿小时,在线视频则下滑 13.8 亿小时。两条短视频赛道合计贡献全网使用时长增量的 111.5%。
盗用 AI 平台算力,上海一男子获刑三年

上海首例利用 AI 伪造签名、盗用大模型算力案近日一审宣判。澎湃新闻 11 月 2 日从徐汇区检察院获悉,被告人宗某某因犯非法获取计算机信息平台数据罪,被判有期徒刑三年、缓刑三年,并处罚金 4 万元。
据检察机关介绍,宗某某利用 AI 编写自动化注册脚本、伪造电子签名,绕过平台的注册和调用限制,自建中转网站低价出售图片、视频生成服务。至案发,他非法调用平台接口 29 万余次,违法所得 3.4 万余元。
涉案平台未向外部给出 API 付费调用服务,使用者协议禁止自动化抓取和绕过验证,其账号也曾因违规被批量封禁。承办检察官表示,调用接口需要取得合法授权。
中兴否认出售手机业务传闻

针对中兴通讯拟与互联网大厂等合作成立合资公司、承接其手机业务的消息,中兴通讯 11 月 3 日向《科创板日报》回应称:「不属实。」
此前,雷峰网援引相关公司员工消息称,中兴正在与某互联网大厂商谈手机业务合作,思路涉及由中兴、该互联网大厂及上海国资共同成立合资公司。
报道所述资产范围为面向消费者的国内外手机业务,红魔手机业务不在拟打包范围内。
Claude Code 机型负责人 Cat Wu 离开 Anthropic

Claude Code 和 Cowork 的机型负责人 Cat Wu 今天在 X 发文称已离开 Anthropic,上周五是她在 Anthropic 的最后一天。
据她在文中回忆,两年前 Claude Code 还只是内部小工具时,她让它搭建一个 RL 环境。她花了几十个小时引导,但全程没有手动写过一行代码,最终任务完成,她也因而加入了团队。
她提到,机型刚推出时,多数使用者会手动批准每一条权限提示,主要用 Claude 回答难题和编写测试。如今的默认做法是信任 Claude 的判断,让它自动批准,Claude 已能完成大型迁移,也能构建全新机型。她在帖文中感谢了 Claude Code、Cowork 和 Tag 背后的团队以及使用者。
Cat Wu 早年在 Scale AI 和 Dagster 担任工程师,之后在风投机构 Index Ventures 工作,2024 年 8 月加入 Anthropic,起初负责研究方向的机型经理工作。
姚顺宇否认创业:仍是 Google DeepMind 员工

「我没有创办公司,当下仍是谷歌 DeepMind 的员工。」针对网传筹备递归自我改进(RSI)公司的消息,AI 研究员姚顺宇 11 月 2 日晚间向澎湃科技作出否认。这篇由记者喻琰采写的采访于 11 月 3 日发布。
此前传闻称,相关创业项目的联合创始团队已基本成形,并在吸引 Google DeepMind、Anthropic 和 OpenAI 的人才。姚顺宇回应:「我完全不知道这个假消息是从哪里来的。」
这已是他近期第二次回应创业传言。10 月 12 日,他曾在社交平台表示,当下没有在中国融资的计划,并称自己「只是个普通的研究员」。
多所高校免费发 Token,同济每周给出 1 亿 Token
高校开始把 AI 使用额度纳入校园公共资源。据第一财经报道,中国科学工艺大学、北京邮电大学、同济大学已先后向师生给出免费 Token 或算力账户。
中科大年中推出「词元计划」,面向全体在校师生给出每人每天 1 亿 Token,次日清零重置,可调用 DeepSeek、GLM 和千问等模型。
同济大学 10 月 26 日上线人工智能服务平台,为每位师生给出 AI 账号,每周发放 1000 积分,折合 1 亿 Token;平台整合模型调用、智能应用开发、实践教学和多模态生成。
北邮使用另一种方式:9 月上线的「人人有算力」平台为近万名本科和研究生新生开通免费账户,集成 10 余种大模型与智能体。
学生取得可自主规格的虚拟服务器环境,可直接调用已部署模型、开发智能体。
蔚来累计换电突破 1.3 亿次,最近 1000 万次用时 88 天

蔚来宣布,11 月 3 日 18:53:21,累计换电次数突破 1.3 亿次,日均换电单量超过 11 万次。从第 1 次到第 1000 万次换电,蔚来用了 1506 天;从 1.2 亿次到 1.3 亿次,用时缩短至 88 天。
截至当天,蔚来在全国布局 4236 座换电站,其中 1077 座位于高速公路,连接 550 座城市,形成「9 纵 11 横 16 大城市群」高速换电网络。公司称,84.5% 的使用者在居住地或工作地周边 3 公里内可找到至少一座换电站。
充电网络维度,蔚来另有 5556 座充电站、32261 根充电桩,累计建成充换电站 9792 座;已贯通 92 条加电风景线,计划年内达到 100 条。
Seedance 海外收入占比达一半,腾讯与 B 站成为国内大客户
据《智能涌现》报道,字节跳动视频生成模型 Seedance 的海外收入占比已达 50%。接近火山引擎的人士称,过去半年,其海外收入增长近 5 倍。
国内市面的主力客户仍是互联网企业。报道援引知情人士消息称,按 Token 消耗和贡献收入衡量,腾讯位列 Seedance 中国区客户前三,B 站进入前十。
实体产业客户数量占比则已超过四分之一,覆盖制造、新零售、商业服务、教育、医疗健康、通信、汽车和能源等领域。
火山引擎今年成立了 FDE(前沿部署工程师)团队,深入汽车、医疗、教育等重点行业。总裁谭待曾在 6 月 FORCE 大会上表示,希望 Seedance 进入各行业的生产环境,特别是高端生产力制造。
AMD 数据中心收入接近翻倍,第四季度营收指引达 170 亿美元

AMD 公布 2026 年第三季度业绩,营收创纪录达到 134 亿美元,同比增长 45%,超过此前指引上限。GAAP 净利润为 24 亿美元,摊薄每股收益 1.46 美元;non-GAAP 净利润为 35 亿美元,摊薄每股收益 2.09 美元。
- 数据中心:收入 85 亿美元,同比增长 97%,主要由 EPYC 处理器和 Instinct MI450 系列 GPU 的需要推动。
- 客户端与游戏:合计收入 37 亿美元,同比下降 8%。其中客户端收入 29 亿美元、增长 7%;游戏收入 7.73 亿美元、下降 40%。
- 嵌入式:收入 11 亿美元,同比增长 29%。本季 GAAP 与 non-GAAP 毛利率分别为 54% 和 56%。
苏姿丰表示,随着 Helios 机架级思路和「Venice」机型扩大生产,预计第四季度增长将进一步加速。公司给出的第四季度营收指引为 170 亿美元、上下浮动 3%,中值对应同比增长约 66%,non-GAAP 毛利率预计约 55%。
💡 阿里董事会主席:只有开源才能普及 AI

在 11 月 3 日世界互联网大会乌镇峰会开幕式上,阿里巴巴董事会主席蔡崇信表示,开源能降低中小企业研发成本,对小语种和发展中国家尤其有价值。他认为,开源让更多地区可以参与 AI 发展,并可将模型部署在自己的数据中心。
蔡崇信以乌干达的一个非营利组织为例:该组织基于千问开发本地语言模型,已覆盖 68 种非洲语言,协助当地使用者跨越语言门槛。他表示,千问当下已开源超过 470 款模型,全球累计下载量突破 40 亿次,衍生模型超过 33 万个。
他还强调,开放应涵盖人才、想法、数据与工具的共享。阿里运营的魔搭社区已吸引全球 239 个国家和地区的 3300 万使用者,参与者来自中国、美国、印度、东南亚、非洲和拉丁美洲等地区。
蔡崇信表示,AI 应用正在工业制造、医疗健康、公共服务、企业办公和科技研发等领域规模化,需要持续增加。中国企业挑选开放模型与生态,是希望借助共享让更多人参与进来。对于下一阶段,他提出的议题涵盖如何更公平分享智能知识,以及如何共同应对 AI 伦理与治理难题。
英伟达与 Pika 基于 MiniMax H3 开源视频增强和生成模型

MiniMax H3 迎来两项外部开源项目:英伟达与 MIT 团队的 Open-H3-Regenerate-2K,以及 Pika 于 10 月 29 日发布的 H3 Flex。前者补上高分辨率视频重生成环节,后者面向更快的视频、声音联合生成。
Open-H3-Regenerate-2K 冻结 H3 主干与视频 VAE,借助细化 LoRA 和三步扩散重绘低分辨率草稿的细节。项目兼容将 H3 或其他模型生成的草稿处理为 2K、原生 4K 等分辨率,单次可处理 5 至 15 秒视频。
Pika H3 Flex 将去噪从 50 步压至 6 步,开放生成适配权重和推理代码,覆盖文本生成视频、首尾帧生成及多模态参考生成,并同步生成音频。开放模型原生输出为 480p 或 768p;1080p、2K、4K 放大和音频分轨由 Pika 托管服务给出。
商汤发布 SenseNova 6.8 Flash,兼容动态外观与可编辑 3D 工程

商汤于 11 月 2 日推出 SenseNova 6.8 Flash 正式版,围绕多步骤任务协作、长期记忆及外观交付开展更新。模型已上线 SenseNova Studio 和 Token Plan,商汤小浣熊也给出限时免费感受。
外观能力覆盖网页、SVG、Three.js、Blender 和 PPT,并可衔接 Figma 工作流。商汤称,模型可在生成 PPTX 时统筹转场、元素入场和图表动效;与 Blender 协作时,可依据实际渲染结果及使用者反馈逐轮修改情形,输出可继续编辑的工程。这项动态 3D 演示依赖本地环境搭建和使用者多轮交互。
商汤还展示了长期记忆与主动交互的办公案例:将新机型兼容私有化部署的信息,与此前因不兼容该能力而流失的客户需要关联,提醒销售重新跟进。该主动交互演示基于小浣熊内部测试版。
法拉利首款电动车 Luce,2027 年产量已全部被预订

法拉利 CEO Benedetto Vigna 在第三季度财报电话会上表示,公司首款电动车 Luce 的 2027 年生产配额已全部分配,新客户需要等待约一年才能提车。首批交付预计于第四季度开始。
据 TechCrunch 报道,这款五座电动车由前苹果外观主管 Jony Ive 操刀,5 月亮相时曾引发外观争议。Vigna 称,预订客户在类型和地域分布上都有良好组合,没有特定的人群模式。
法拉利尚未公开 Luce 的计划产量。公司在本次财报中还提到,个性化定制需要及限量超跑 F80 交付推动了业绩,定制业务已占汽车及零部件收入的逾两成。
美团 LongCat-2.5 正式发布,国产算力训练原生多模态模型

美团于 11 月 2 日正式发布 LongCat-2.5,面向需要持续执行的长程智能体任务。模型总参数量为 1.6 万亿,每个 Token 激活约 480 亿参数,强化了原生视觉理解和智能体编程能力。
新模型把图像转为多层级离散 Token,与文字交给同一主干模型处理,兼容看图编程、创意外观和 3D 任务。训练任务来自 GitHub 公开仓库及内部真实代码情形,团队还将 Agent 的失败轨迹还原为可复现、带测试的任务,补强复杂代码库开发和完整项目构建能力。
训练和推理围绕国产 AI 加速卡优化。在美团测试的长序列监督微调任务中,单卡吞吐较 DDP 基线提高 86%;线上压测中,W4A4 量化相较 W8A8 基线,使预填充和解码吞吐分别提高 43.3% 和 42.77%。模型已给出在线感受和 API 接入。

智能投影销量连跌四季,第三季度中国市面下滑 26.2%

据界面新闻援引洛图科技 11 月 3 日发布的数据,2026 年第三季度中国智能投影市面销量为 82.9 万台,同比下降 26.2%;销售额为 12.3 亿元,同比下滑 27.1%。统计不含激光电视,市面已连续四个季度同比下滑,各季跌幅均超过 20%。
市面收缩的同时,线上品牌集中度上升。第三季度销量前十品牌的合计份额达 74%,销额前十品牌合计占 81.7%。极米位居线上销量、销额双榜第一,大眼橙销量第二、康佳第三;坚果则位居销额第二。
洛图科技还将全年零售量预测下调至 385.7 万台,同比下降 25.9%;预计全年销额为 62.8 亿元,同比下降 24.8%。
佳能计划明年起调价,应对零部件成本上涨

佳能在 2026 年第三季度分析师问答中表示,公司正设法确保明年所需的紧缺零部件供应,预计采购成本将上升,计划从 2027 年起借助提高售价抵消成本增长。公司会综合考虑销量与盈利,作出长期售价决策。
原始问答还区分了各业务的供应状况:打印和影像业务也面临零部件短缺,但因提前应对,尚未受到重大影响;医疗业务则因生产平台更新后的运营难题,未能及时下单采购足量零部件,预计到明年第二季度才能取得满足当前生产计划的供应。
佳能预计,明年的内存和原材料售价不会下降,相关成本将影响全年经营。
全国放心消费平台上线,扫码可查周边商家承诺

市面监管总局 11 月 3 日正式上线全国放心消费平台,消费者搜索「消费查」小程序或扫描线下放心消费码即可进入。平台会按定位展示周边商家,首批已有超 1 万个放心消费单元、100 个放心消费集聚区入驻。
商家注册、在线承诺并经核验公示后,可取得「一店一码」的放心消费码和全国统一标识。平台首批已对接百度地图、美团外卖、淘宝闪购和京东外卖,让同一商家的标识在多个入口展示。
消费者可在商家主页查看证照、门店实景、售后方式、消费承诺和评价,也能一键导航、评价或电话联系商家。对发生违法失信行为、拒不履行承诺或推诿合理诉求的商家,监管部门查实后将责令其退出平台,并收回线下消费码与标识。
《好莱坞往事》衍生片再请《社交网络》配乐搭档

Trent Reznor 与 Atticus Ross 将为大卫·芬奇的新片《The Further Mis-Adventures of Cliff Booth》配乐,这是两人与芬奇在电影上的第六次合作。新海报公布了配乐阵容,影片将于 11 月 25 日在 IMAX 独家上映两周,12 月 23 日上线 Netflix。
影片由昆汀·塔伦蒂诺编剧,是《好莱坞往事》的衍生作品,故事发生在前作八年后,围绕布拉德·皮特饰演的 Cliff Booth 展开。演员还涵盖伊丽莎白·德比茨基、卡拉·古奇诺和蒂莫西·奥利芬特等。
Reznor 与 Ross 从 2010 年《社交网络》开始为芬奇电影配乐,并凭该片取得奥斯卡最佳原创配乐奖;此后合作延续至《龙纹身的女孩》《消失的爱人》《曼克》和《杀手》。
《蓝眼武士》第二季定档明年 1 月,复仇之旅转向伦敦

Netflix 宣布,《蓝眼武士》(Blue Eye Samurai)第二季将于 2027 年 1 月 19 日上线,共六集。混血剑客 Mizu 的旅程将从江户时代的日本延伸至伦敦,继续寻找另两名可能是她父亲的男子。
主创 Amber Noizumi 和 Michael Green 表示,新一季设定在 1657 年的伦敦,Mizu 将进入当地的地下世界,并面对身为亚洲女性在当时英国社会的处境。两人称,团队花了三年时间研究并绘制这些历史情形。
Maya Erskine 继续为 Mizu 配音,新增配音演员涵盖 Jonathan Pryce、Joel Kim Booster 和 Sherry Cola。Netflix 同时确认,这段故事将在 2028 年迎来终章。
Netflix 测试 AI Discover,按心情找片还能屏蔽不想看的建议

Netflix 宣布在电视 App 中测试新栏目 AI Discover,先向美国使用英语、设备受兼容的少量会员开放,并将在将来几周逐步扩大范围。
首批感受围绕主题找片:使用者可以从「从容一点」「吓吓我」「家庭电影夜」或「短时间观看」等主题进入,查看匹配片单,再挑选更细的主题继续缩小范围,或用「换些不一样的」取得新建议。
每部作品旁会新增调整建议按钮,与播放、加入片单并列。使用者可以保留该建议,或要求以后不再建议这部作品;AI Discover 会记住反馈,减少重复浮现。
Netflix 表示,这一栏目由新的智能层驱动,会依据会员输入实时调整。公司将据使用情况和反馈决定保留、改进哪些感受,部分能力也可能进入主界面。
三周年倒计时|人人都爱贴贴线

距离制糖工厂贴贴线三周年,还有 1 天。
三年前,为了应对数据线容易缠绕、收纳麻烦的小难题,第一根会自己磁吸的线——MagTie 贴贴线诞生了。三年里,从贴贴线到云朵线、花线、细雳线、细雳线 Pro 和贴贴卷,这个家族逐步有了不同的模样:更柔软、更有个性、更高速,也有更省空间的挑选。不变的,是让日常连接更顺手的想法。
三周年之际,笔者也想听听:你最喜欢哪一款,又和它一起度过了哪些日常?欢迎分享你的贴贴故事,一起迎接三周年,蹲蹲重磅福利!
本文来源:https://www.ifanr.com/1687477
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
🔧 Mainland China Apple Intelligence begins testing, Siri, Writing Tools and Visual Intelligence integrated with Qwen
🌦️ Apple Weather shows data anomaly, multiple cities to "drop to" -32 degrees next Friday
🧑💻 Apple opens Siri AI to all developers, can chain multiple App actions into tasks
🧠 GPT rewrites Qwen weights, team tests show math scores improved by approximately 6 times
🤝 OpenAI uses Astra to unveil voice Codex device within one week, partners with Mazda for research
💬 DeepSeek V5 reportedly plans to use Ascend for pre-training
💼 DeepSeek crashes again
🏢 Hongguo user scale surpasses Doubao, Qwen leads the 100-million-level App growth rankings
🖥️ Stealing AI platform computing power, a Shanghai man sentenced to three years
📱 ZTE denies rumors of selling its mobile phone business
🧠 Claude Code device lead Cat Wu leaves Anthropic
💬 Yao Shunyu denies entrepreneurship: still an employee of Google DeepMind
🎓 Multiple universities distribute free Tokens, Tongji provides 100 million Tokens weekly
🚗 NIO's cumulative battery swaps surpass 130 million, the latest 10 million took 88 days
🏢 Seedance's overseas revenue reaches half, Tencent and Bilibili become major domestic clients
📈 AMD's data center revenue nearly doubles, Q4 revenue guidance reaches $17 billion
💡 Alibaba Board Chairman: only open source can popularize AI
🔓 NVIDIA and Pika open-source video enhancement and generation model based on MiniMax H3
⚡ SenseTime releases SenseNova 6.8 Flash, compatible with dynamic appearance and editable 3D workflows
🚗 Ferrari's first electric car Luce, 2027 production fully booked
🖥️ Meituan officially releases LongCat-2.5, native multimodal model trained on domestic computing power
📉 Smart projector sales decline for four consecutive quarters, China market down 26.2% in Q3
🛍️ Canon plans price adjustments starting next year to cope with rising component costs
✨ National trusted consumption platform launches, scan QR code to check nearby merchants' commitments
🌐 Once Upon a Time in Hollywood spinoff again recruits The Social Network scoring duo
🎬 Blue Eye Samurai Season 2 set for next January, revenge journey turns to London
🧠 Netflix tests AI Discover, find movies by mood and block unwanted recommendations
Mainland China Apple Intelligence begins testing, Siri, Writing Tools and Visual Intelligence integrated with Qwen
Early this morning, Apple pushed the first developer beta of iOS 27.3, and mainland China "Apple Intelligence and Siri" begins to open for experience.
The first batch of capabilities listed on Apple’s Chinese official website covers Writing Tools, Real-Time Translation, Clean Up, Genmoji, and Notification Summaries.
Qwen has been integrated into Siri, Visual Intelligence, and Writing Tools, allowing users to ask questions, generate text, or process images within the platform.
According to Ifanr’s hands-on testing, Visual Intelligence’s image-based search has also connected to Baidu Search. Siri now features a glowing halo around the edge of the display and supports double-tapping the bottom of the screen to switch to text input.
In Siri, basic Q&A can be completed directly. Taking “I’ve been having trouble sleeping lately” as an example, Siri will invoke Qwen to provide general sleep advice.
For more personalized guidance, users can tap the Qwen logo to jump to the Qwen app, and conversation information will sync automatically. According to reports, Qwen can draw on a user’s prior sleep records, exercise activity, and lifestyle habits to devise improvement plans.
Real-Time Translation covers Messages, Phone, and FaceTime, with Phone and FaceTime limited to one-on-one calls. AirPods translation requires updating the earbuds to the latest firmware and pairing them with an iPhone that has Apple Intelligence enabled.
The first batch of capabilities applies to the iPhone 15 Pro lineup, the full iPhone 16 lineup, and later models, as well as iPhone Air and iPhone Duo; Siri and the device language must be set to Simplified Chinese or English.
Notably, this mainland China update delivers only the first batch of Apple Intelligence capabilities; the new Siri AI previewed at WWDC 2026 is not included.
🔗 Related reading: We tested the mainland China iPhone AI overnight—here are the highlights and the gripes
Apple Weather shows data anomalies; many places “drop to” minus 32 degrees next Friday
Swipe left to see more content
Early today, Apple’s built-in Weather app showed forecast data anomalies in many Chinese cities, incorrectly displaying next Friday (November 13)’s minimum temperature as around -32°C.
Affected areas include multiple cities such as Beijing, Shanghai, Nanjing, Wuhan, Hefei, Fuzhou, Xi’an, and Kunming, with the displayed extreme lows generally concentrated between -29°C and -33°C.
For example, the Kunming page showed that the day’s minimum temperature would plunge to -32°C between 19:00 and 20:00, while the daytime high remained at 17°C to 24°C, making the single-day day-night temperature swing close to 50°C. Meanwhile, weather forecasts for Guangzhou, Shenzhen, Nanning, Haikou, and other parts of Guangdong, Guangxi, and Hainan appeared normal, with no similar data showing up.
The National Meteorological Center and local meteorological authorities have not issued cold-wave warnings recently, and mainstream domestic and international numerical forecast models have not predicted a sharp temperature drop around November 13.
A weather blogger analyzed that the time when the anomalies appeared coincided with the numerical model forecast node at 20:00 Beijing Time (12:00 UTC), and speculated that the forecast data at that run contained errors that were then amplified by algorithmic interpolation.
Apple opens Siri AI to all developers, enabling multiple app actions to be chained into tasks
Apple announced a new Siri AI developer preview rolling out with the iOS, iPadOS, macOS, and visionOS 27.3 beta, allowing all developers to use App Intents to bring app content and actions into the new Siri.
Once connected, Siri AI can perform in-app actions based on personal context, what’s on screen, and natural language requests—such as checking deliveries, making reservations, or placing orders. Apple also introduced “skills,” allowing developers to combine multiple operations into a single end-to-end experience and letting Siri AI search for entities in Spotlight’s semantic index.
Fifteen Apple apps—including Calendar, Files, Finder, Mail, Maps, Messages, and Photos—have begun to offer related capabilities, such as finding the coffee shop you visited last week or archiving files that haven’t been opened for over a year. Apple said these interfaces will first be available for developers to explore, then roll out to users after a few months, with App Store submissions opening in spring next year.
The current Siri AI beta requires an Apple Intelligence–compatible device and supported languages; Simplified Chinese is not in the current language list, while Traditional Chinese, Norwegian, Turkish, and Vietnamese will be added later.
GPT reworks Qwen weights, team’s math test score up about 6x
MIT CSAIL PhD student Rikiya Takehi and collaborators released a “weight-space programming” experiment: they let GPT-6 Astra directly read and modify the weights of Qwen3.5-0.8B without using gradient descent or a training loop. In the team’s MATH-500 tests, the modified model’s average accuracy rose from 7.9% to 48.1%, about 6 times the original.
The experiment provided Astra with a Python environment for weight operations, as well as math practice materials that did not include MATH-500 questions. The most effective run allowed up to 50 forward passes for the model to inspect generated outputs and internal states; the final edit involved about 0.016% of the weights. The team said Astra identified and weakened directions that caused the small model to fall into repetitive reasoning, focusing on modifying certain MLP output columns in layers 15 and 19.
In the same round of tests, GSM8K accuracy increased from 13.8% to 67.1%, and IFEval strict instruction-following accuracy rose from 21.0% to 62.3%. The team has published the experiment prompts, code, test results, and weight diffs; the authors later also shared the reconstructed model and an online demo.
🔗 Related Reading: GPT-6 Astra Rewires Qwen's Brain: Touching Only 0.016% of Parameters, Math Scores Skyrocket
OpenAI Uses Astra to Design Voice Codex Device in One Week, Partners with Mazda on Research
OpenAI revealed a hardware design experiment: two team members used GPT-6 Astra to complete the design of a device capable of voice interaction with Codex in less than a week of spare-time project time, and the manufacturer then produced a small batch of prototypes in under a month.
The model participated in the exterior design of the shell and internal structure, component selection, PCB layout and wiring, as well as interface, firmware, and Codex software integration. After entering the manufacturing phase, the team still carried out multiple rounds of review and iteration to address production constraints, tolerances, and assembly details that the model had not fully considered.
The company also announced a research collaboration with Mazda, letting the model assist aerodynamics and crash safety engineers in designing, simulating, and evaluating vehicle concepts. The collaboration with Chip Ganassi Racing is exploring converting 3D scans of existing parts into editable CAD models and 2D manufacturing drawings, with outputs checked by team engineers for accuracy and manufacturability.
On BenchCAD 2.0, which tests 3D part reconstruction, assembly, and circuit board connection recognition, OpenAI disclosed that a newer internal model averaged a score of 0.875, while GPT-6 Astra's highest reasoning tier scored 0.594.
DeepSeek V5 Reportedly Plans to Use Ascend for Pre-training
According to multiple industry chain sources cited by STAR Market Daily, DeepSeek's next-generation large model V5 plans to carry out pre-training based on Huawei Ascend chips. Previously, Ascend chips primarily handled inference tasks for DeepSeek models.
The report notes that DeepSeek open-sourced infrastructure components for Ascend on September 30, covering the TileLang compiler tool, compute libraries such as DeepGEMM and FlashMLA, and the DeepEP distributed communication library, corresponding to components previously available for the NVIDIA platform.
Huawei, for its part, presented a jointly defined supernode and networking approach, and opened up underlying programming interfaces and instruction compatibility.
On the training side, teams such as the Shenzhen Hetao Academy have previously completed full-parameter post-training of DeepSeek-V4-Pro relying on Ascend 910C clusters; the reported plan for V5 involves basic pre-training of the next-generation model.
DeepSeek Crashed Again
On the afternoon of November 3, partial outages emerged in DeepSeek's web and API services. A Yicai reporter's testing found that conversation capabilities were unusable, with the chat interface showing "Server busy, please try again later."
Official status information indicates the outage began at 16:58, affecting conversation services, DeepSeek V4 Pro API, and V4.1 Flash API. The service then entered a repair and monitoring phase.
By around 21:00 that evening, DeepSeek's official status page showed "Everything is running normally," with all platforms operating as expected.
Hongguo's User Base Surpasses Doubao; Qianwen Leads the Billion-User App Growth Rankings
QuestMobile released its Autumn Mobile Internet Report and Strength Value Ranking on November 3. In September, ByteDance's internal tally based on deduplicated total users saw the top three apps become Douyin, Hongguo Free Short Dramas, and Doubao, at approximately 1.029 billion, 430 million, and 419 million respectively, with Hongguo surpassing Doubao.
Qianwen, with approximately 158 million monthly active users and year-on-year growth of 5077%, ranked first in user-base growth among apps with over 100 million monthly active users. The top five on this list also included Douyin Mall, Doubao, Hongguo Free Short Dramas, and Qishui Music. In the AI-native app monthly active rankings, Doubao, Qianwen, and DeepSeek remained the top three, with DeepSeek's monthly active users at approximately 123 million.
Short dramas also drew away more usage time. The report indicates that total usage time for short drama videos saw a net increase of 13.329 billion hours, short-form videos a net increase of 5.717 billion hours, while online videos declined by 1.38 billion hours. The two short-video categories together contributed 111.5% of the total increase in usage time across the entire network.
Shanghai Man Sentenced to Three Years for Stealing AI Platform Compute
Shanghai's first case of using AI to forge signatures and steal large-model compute was recently decided at the first-instance trial. The Paper learned on November 2 from the Xuhui District People's Procuratorate that the defendant, surnamed Zong, was sentenced to three years' imprisonment with a three-year reprieve and fined 40,000 yuan for the crime of illegally obtaining computer information platform data.
According to the procuratorial authorities, Zong used AI to write automated registration scripts, forge electronic signatures, bypass the platform's registration and call limits, and built his own relay website to sell image and video generation services at low prices. By the time of the incident, he had illegally called platform interfaces more than 290,000 times, with illicit gains of over 34,000 yuan.
The platform involved did not offer paid API access to external parties, and its user agreement prohibited automated scraping and bypassing verification; his account had also been mass-banned for violations. The handling prosecutor stated that calling interfaces requires obtaining legitimate authorization.
ZTE Denies Rumors of Selling Its Phone Business
In response to reports that ZTE planned to partner with internet giants and others to establish a joint venture to take over its phone business, ZTE replied to STAR Market Daily on November 3: "Not true."
Earlier, Leiphone reported, citing an employee at a related company, that ZTE is discussing a smartphone business partnership with a major internet company, with the plan involving a joint venture to be established by ZTE, the internet company, and Shanghai state-owned capital.
The assets covered by the report refer to consumer-facing domestic and overseas smartphone businesses, while the RedMagic smartphone business is not part of the planned package.
Claude Code model lead Cat Wu leaves Anthropic
Cat Wu, the model lead for Claude Code and Cowork, posted on X today that she has left Anthropic, and last Friday was her last day at the company.
She recalled in the post that two years ago, when Claude Code was still just an internal tool, she had it set up an RL environment. She spent dozens of hours guiding it, but did not manually write a single line of code throughout the process; the task was ultimately completed, and she then joined the team.
She noted that when the model first launched, most users would manually approve every permission prompt, mainly using Claude to answer hard questions and write tests. Today’s default approach is to trust Claude’s judgment and let it auto-approve, and Claude has been able to complete large-scale migrations and build entirely new models. In the post, she thanked the teams behind Claude Code, Cowork, and Tag, as well as the users.
Cat Wu previously worked as an engineer at Scale AI and Dagster, then worked at venture capital firm Index Ventures. She joined Anthropic in August 2024, initially serving as a model manager working on research direction.
Yao Shunyu denies starting a company: still an employee at Google DeepMind
“I have not founded a company, and I am still an employee at Google DeepMind.” Responding to online rumors that he was preparing a recursive self-improvement (RSI) company, AI researcher Yao Shunyu denied it to The Paper on the evening of November 2. The interview, conducted by reporter Yu Yan, was published on November 3.
Rumors had earlier claimed that the co-founding team for the startup project was largely in place and was attracting talent from Google DeepMind, Anthropic, and OpenAI. Yao Shunyu responded: “I have no idea where this fake news came from.”
This is the second time recently that he has responded to startup rumors. On October 12, he said on a social platform that he currently has no plans to raise funding in China, adding that he is “just an ordinary researcher.”
Multiple universities are giving out free Tokens; Tongji is providing 100 million Tokens per week
Universities have begun incorporating AI usage quotas into campus public resources. According to a report by Yicai, the University of Science and Technology of China (USTC), Beijing University of Posts and Telecommunications (BUPT), and Tongji University have successively provided free Tokens or compute accounts to faculty and students.
In mid-year, USTC launched the “Token Plan,” offering every faculty member and student 100 million Tokens per person per day, with the quota reset to zero the next day. The plan supports models including DeepSeek, GLM, and Qwen.
On October 26, Tongji University launched an artificial intelligence service platform, providing every faculty member and student with an AI account and issuing 1,000 points per week, equivalent to 100 million Tokens. The platform integrates model invocation, intelligent application development, practical teaching, and multimodal generation.
BUPT is taking a different approach: its “Compute for Everyone” platform, launched in September, has opened free accounts for nearly 10,000 freshmen majoring in undergraduate and graduate programs, integrating more than 10 large models and agents.
Students gain access to a virtual server environment with configurable specifications, and can directly call deployed models and develop agents.
NIO’s cumulative battery swaps surpass 130 million; the most recent 10 million took 88 days
NIO announced that at 18:53:21 on November 3, its cumulative number of battery swaps exceeded 130 million, with average daily swaps surpassing 110,000. From the 1st to the 10 millionth swap, NIO took 1,506 days; from 120 million to 130 million, the time was shortened to 88 days.
As of that day, NIO had deployed 4,236 battery swap stations nationwide, including 1,077 along highways, connecting 550 cities and forming a highway swap network of “9 vertical corridors, 11 horizontal corridors, and 16 major city clusters.” The company said that 84.5% of users can find at least one battery swap station within 3 kilometers of where they live or work.
On the charging side, NIO also operates 5,556 charging stations with 32,261 charging piles, bringing the total number of charging and battery swap stations to 9,792. It has connected 92 scenic charging routes and plans to reach 100 by the end of the year.
Seedance derives half of its revenue from overseas; Tencent and Bilibili become major domestic customers
According to Zhi Neng Yong Xian, ByteDance’s video generation model Seedance has reached a 50% share of revenue from overseas markets. A person close to Volcano Engine said that over the past six months, its overseas revenue has grown nearly fivefold.
The main customers in the domestic market are still internet companies. The report, citing people familiar with the matter, said that by Token consumption and revenue contribution, Tencent ranks among the top three Seedance customers in China, while Bilibili ranks in the top ten.
The share of customers from physical industries has exceeded one quarter, covering areas such as manufacturing, new retail, business services, education, healthcare, communications, automotive, and energy.
This year, Volcano Engine established an FDE (Forward Deployed Engineer) team to go deep into key industries such as automotive, healthcare, and education. President Tan Dai said at the FORCE conference in June that he hopes Seedance will enter production environments across industries, especially high-end productivity manufacturing.
AMD’s data center revenue nearly doubled, with fourth-quarter revenue guidance reaching $17 billion
AMD announced its third-quarter 2026 results, with record revenue reaching $13.4 billion, a 45% year-over-year increase, exceeding the upper limit of prior guidance. GAAP net income was $2.4 billion, with diluted EPS of $1.46; non-GAAP net income was $3.5 billion, with diluted EPS of $2.09.
Data Center: Revenue was $8.5 billion, up 97% year-over-year, driven primarily by demand for EPYC processors and Instinct MI450 series GPUs.
Client and Gaming: Combined revenue was $3.7 billion, down 8% year-over-year. Specifically, client revenue was $2.9 billion, up 7%; gaming revenue was $773 million, down 40%.
Embedded: Revenue was $1.1 billion, up 29% year-over-year. GAAP and non-GAAP gross margins for the quarter were 54% and 56%, respectively.
Su Zifeng stated that as production scales for the Helios rack-scale platform and the "Venice" model, growth is expected to accelerate further in the fourth quarter. The company's fourth-quarter revenue guidance is $17.0 billion, plus or minus 3%, with the midpoint representing an approximate 66% year-over-year increase; non-GAAP gross margin is expected to be approximately 55%.
💡 Alibaba Chairman: Only open source can popularize AI
At the opening ceremony of the Wuzhen Summit of the World Internet Conference on November 3, Alibaba Chairman Joe Tsai said that open source reduces R&D costs for SMEs and is especially valuable for low-resource languages and developing nations. He believes open source allows more regions to participate in AI development and enables the deployment of models in their own data centers.
Joe Tsai cited a non-profit organization in Uganda as an example: the organization developed a local language model based on Qwen that has already covered 68 African languages, helping local users cross language barriers. He stated that Qwen has currently open-sourced over 470 models, with global cumulative downloads exceeding 4 billion and derivative models surpassing 330,000.
He also emphasized that openness should encompass the sharing of talent, ideas, data, and tools. The ModelScope community operated by Alibaba has attracted 33 million users from 239 countries and regions globally, with participants from regions including China, the United States, India, Southeast Asia, Africa, and Latin America.
Joe Tsai stated that AI applications are scaling in fields such as industrial manufacturing, healthcare, public services, enterprise operations, and tech R&D, requiring sustained growth. Chinese companies choose open models and ecosystems hoping to leverage sharing to let more people participate. For the next stage, he proposed topics covering how to share intelligence knowledge more fairly and how to jointly address AI ethics and governance challenges.
Nvidia and Pika Open Source Video Enhancement and Generation Models Based on MiniMax H3
MiniMax H3 has welcomed two external open-source projects: Open-H3-Regenerate-2K from Nvidia and an MIT team, and H3 Flex released by Pika on October 29. The former fills the gap in high-resolution video regeneration, while the latter targets faster joint video and audio generation.
Open-H3-Regenerate-2K freezes the H3 backbone and video VAE, utilizing refinement LoRAs and a three-step diffusion to repaint details of low-resolution drafts. The project is compatible with drafts generated by H3 or other models, processing them into resolutions such as 2K and native 4K, and can process 5 to 15 seconds of video in a single instance.
Pika H3 Flex reduces denoising steps from 50 to 6, open-sources the adapted weights and inference code, covering text-to-video generation, first-last frame generation, and multimodal reference generation, while simultaneously generating audio. The open model's native output is 480p or 768p; upscaling to 1080p, 2K, 4K, and separate audio tracks are provided via Pika's hosted service.
SenseTime Releases SenseNova 6.8 Flash, Supporting Dynamic Appearance and Editable 3D Engineering
SenseTime launched the official version of SenseNova 6.8 Flash on November 2, featuring updates focused on multi-step task collaboration, long-term memory, and appearance delivery. The model is now live on SenseNova Studio and Token Plan, and SenseTime's Raccoon also offers a limited-time free trial.
Appearance capabilities cover webpages, SVG, Three.js, Blender, and PPT, and can connect with Figma workflows. SenseTime claims the model can orchestrate transitions, element entrances, and chart animations when generating PPTX; when working with Blender, it can modify scenarios iteratively based on actual rendering results and user feedback, outputting a project that can continue to be edited. This dynamic 3D demonstration relies on local environment setup and multi-turn user interaction.
SenseTime also demonstrated a long-term memory and proactive interaction office case: correlating information about new model compatibility for private deployment with the needs of customers previously lost due to incompatibility with this capability, reminding sales to follow up again. This proactive interaction demo is based on the internal beta version of Raccoon.
Ferrari's first electric car Luce, 2027 production completely booked
Ferrari CEO Benedetto Vigna stated at the third-quarter earnings call that the 2027 production allocation for the company's first electric car, Luce, has been fully allocated, and new customers will need to wait approximately one year for delivery. Initial deliveries are expected to begin in the fourth quarter.
According to a TechCrunch report, this five-seater electric car was designed by former Apple design chief Jony Ive and sparked controversy over its appearance when unveiled in May. Vigna said that reservation customers have a good mix in terms of type and geographic distribution, with no specific demographic pattern.
Ferrari has not yet disclosed the planned production volume for Luce. In the current earnings report, the company also mentioned that personalized customization demand and deliveries of the limited-edition supercar F80 drove performance, with the customization business accounting for over 20% of automotive and parts revenue.
Meituan LongCat-2.5 Officially Released, Native Multimodal Model Trained on Domestic Computing Power
Meituan officially released LongCat-2.5 on November 2, targeting long-horizon agent tasks that require continuous execution. The model has a total parameter count of 1.6 trillion, activating approximately 48 billion parameters per Token, strengthening native visual understanding and agent programming capabilities.
The new model converts images into multi-level discrete tokens and feeds them, together with text, into the same backbone model, supporting image-to-code programming, creative appearance generation, and 3D tasks. Training tasks come from public GitHub repositories as well as internal real-world code scenarios; the team also turns Agent failure trajectories into reproducible, test-equipped tasks to strengthen capabilities in complex codebase development and full-project construction.
Training and inference are optimized around domestic AI accelerator cards. In a long-sequence supervised fine-tuning task tested at Meituan, single-card throughput increased by 86% compared with the DDP baseline; in online stress tests, W4A4 quantization raised prefill and decode throughput by 43.3% and 42.77%, respectively, versus the W8A8 baseline. The model has been made available for online experience and API integration.
Smart projector sales have fallen for four consecutive quarters, with China’s market down 26.2% in the third quarter.
According to data released by RUNTO on November 3 and cited by Jiemian News, China’s smart projector market recorded sales of 829,000 units in the third quarter of 2026, down 26.2% year over year; sales revenue was RMB 1.23 billion, down 27.1% year over year. The statistics exclude laser TVs, and the market has declined year over year for four consecutive quarters, with declines exceeding 20% in each quarter.
While the market contracts, online brand concentration has risen. In the third quarter, the top ten brands accounted for 74% of unit sales and 81.7% of revenue. XGIMI ranked first on both the online unit-sales and revenue charts; Dangbei ranked second in unit sales and Konka third; JMGO ranked second in revenue.
RUNTO also lowered its full-year retail volume forecast to 3.857 million units, down 25.9% year over year; it expects full-year revenue to be RMB 6.28 billion, down 24.8% year over year.
Canon plans price adjustments starting next year to counter rising component costs.
In a Q&A with analysts in the third quarter of 2026, Canon said it is working to secure supplies of scarce components needed next year, expects purchasing costs to rise, and plans to offset cost increases by raising prices from 2027. The company said it would make long-term pricing decisions by comprehensively balancing sales volume and profitability.
The original Q&A also distinguished supply conditions across businesses: the printing and imaging businesses are also facing component shortages, but due to early responses they have not yet been materially affected; the medical business, however, faced operational challenges after updating its production platform and failed to place orders in time to procure sufficient components. It is expected that supply meeting the current production plan will only be secured by the second quarter of next year.
Canon expects that next year’s prices for memory and raw materials will not decline, and that related costs will affect operations throughout the year.
National Trusted Consumption Platform launches; scan a QR code to check nearby merchants’ commitments.
The State Administration for Market Regulation officially launched the National Trusted Consumption Platform on November 3. Consumers can access it by searching for the “Xiaofei Cha” mini program or by scanning an offline Trusted Consumption QR code. The platform displays nearby merchants based on location; in the first batch, over 10,000 Trusted Consumption units and 100 Trusted Consumption clusters have joined.
After registering, making online commitments, and passing verification and public listing, merchants can obtain a “one store, one QR code” Trusted Consumption QR code and a unified national logo. In its initial batch, the platform has integrated with Baidu Maps, Meituan Waimai, Taobao Flash Purchase, and JD Waimai, so that the same merchant’s logo can be displayed across multiple entry points.
On a merchant’s page, consumers can view licenses and permits, real-store scenes, after-sales contact methods, consumption commitments, and reviews, and can also navigate with one click, leave a review, or call the merchant. For merchants that engage in illegal or dishonest conduct, refuse to fulfill commitments, or deflect reasonable requests, regulators will, after verification, order them to exit the platform and revoke their offline consumption QR codes and logos.
“Once Upon a Time in Hollywood” spin-off again enlists “The Social Network” scoring duo.
Trent Reznor and Atticus Ross will score David Fincher’s new film “The Further Mis-Adventures of Cliff Booth,” marking the pair’s sixth film collaboration with Fincher. The new poster announced the scoring lineup; the film will be released exclusively in IMAX for two weeks on November 25, and will debut on Netflix on December 23.
The film is written by Quentin Tarantino and is a spin-off of “Once Upon a Time in Hollywood.” Set eight years after the original, it centers on Cliff Booth, played by Brad Pitt. The cast also includes Elizabeth Debicki, Cara Gugino, and Timothy Olyphant, among others.
Reznor and Ross began scoring Fincher’s films with “The Social Network” in 2010 and won the Academy Award for Best Original Score for that film; subsequent collaborations continued with “The Girl with the Dragon Tattoo,” “Gone Girl,” “Mank,” and “The Killer.”
“Blue Eye Samurai” Season 2 set for next January; the revenge journey turns toward London.
Netflix announced that Season 2 of “Blue Eye Samurai” will premiere on January 19, 2027, with six episodes. The mixed-race swordswoman Mizu’s journey will extend from Edo-period Japan to London, continuing her search for two other men who may be her father.
Creators Amber Noizumi and Michael Green said the new season is set in London in 1657, where Mizu will enter the city’s underworld and confront the realities faced by an Asian woman in that era’s English society. They added that the team spent three years researching and designing these historical circumstances.
Maya Erskine will continue to voice Mizu; new voice cast members include Jonathan Pryce, Joel Kim Booster, and Sherry Cola. Netflix also confirmed that this story will reach its final chapter in 2028.
Netflix tests AI Discover: find films by mood and also block suggestions you don’t want to see.
Netflix announces testing a new section, AI Discover, in its TV App, initially opening to a small number of members in the US who use English on compatible devices, with plans to gradually expand availability in the coming weeks.
The first experience centers on finding films by theme: users can enter from themes like "Take It Easy," "Scare Me," "Family Movie Night," or "Quick Watch" to view matching lists, then select more specific themes to further narrow results, or use "Something Different" to get new suggestions.
Each title will feature a new "Adjust Recommendations" button alongside play and add-to-list. Users can keep the suggestion or request that this title not be recommended in the future; AI Discover will remember the feedback and reduce repeated appearances.
Netflix stated that this section is powered by a new intelligence layer that adjusts in real time based on member input. The company will decide which experiences to retain or improve based on usage and feedback, and some features may also be integrated into the main interface.
Third Anniversary Countdown | Everyone Loves the MagTie Cable
One day left until the third anniversary of the Sugar Factory's MagTie Cable.
Three years ago, to address the small challenge of data cables easily tangling and being hard to store, the first self-magnetic cable—MagTie Cable—was born. Over the past three years, from MagTie Cable to Cloud Cable, Flower Cable, Thunder Cable, Thunder Cable Pro, and MagTie Roll, this product family has gradually taken on different forms: softer, more distinctive, faster, and with more space-saving options. What remains unchanged is the vision of making everyday connectivity more effortless.
On this third anniversary, I'd also love to hear: which model is your favorite, and what everyday moments have you shared with it? Feel free to share your MagTie story, let's celebrate the third anniversary together, and stay tuned for big perks!