谷歌抢发Gemini研究代理，与OpenAI上演AI卡位战

简报：

• 谷歌在OpenAI发布GPT-5.2前夕，抢先上线了基于最新Gemini 3 Pro模型的AI研究代理“Gemini Deep Research”。
• 该工具专为处理海量信息和复杂研究任务设计，并首次通过Interactions API向开发者开放，允许将研究能力嵌入第三方应用。
• 为应对AI“幻觉”，Gemini 3 Pro特别优化了事实准确性，并推出了新的基准测试DeepSearchQA来验证其能力。
• OpenAI随后正式发布GPT-5.2，其CEO萨姆·奥尔特曼表示，曾因Gemini的威胁在内部拉响“红色警报”，但其实际影响小于预期，警报将于明年1月结束。

相关链接：

• 谷歌紧急上线Gemini Deep Research迎战GPT-5.2，AI“研究代理”大战一触即发
• OpenAI 奥尔特曼：高看了谷歌 Gemini 3，明年 1 月结束“红色警报”

Google Rushes to Release Gemini Research Agent, Engaging in AI Positioning Battle with OpenAI 🚀

Brief:

• On the eve of OpenAI's GPT-5.2 release, Google preemptively launched "Gemini Deep Research," an AI research agent powered by its latest Gemini 3 Pro model.
• This tool is specifically designed to handle massive amounts of information and complex research tasks, and for the first time, it's open to developers via the Interactions API, allowing its research capabilities to be embedded into third-party applications.
• To counter AI "hallucinations," Gemini 3 Pro has been specifically optimized for factual accuracy, and a new benchmark, DeepSearchQA, has been introduced to validate its capabilities. 🧠
• OpenAI subsequently officially released GPT-5.2. Its CEO, Sam Altman, stated that an internal "red alert" was triggered due to the perceived threat from Gemini, but its actual impact was less than expected, and the alert will conclude in January next year. 😮‍💨

Related Links:

• Google urgently launched Gemini Deep Research to battle GPT-5.2, the AI "research agent" war is on the verge of breaking out
• OpenAI's Altman: Overestimated Google Gemini 3, "red alert" to end in January next year

奥尔特曼回应Gemini 3竞争：曾高估其影响，OpenAI将解除警报

简报：

• OpenAI CEO萨姆·奥尔特曼表示，谷歌发布的Gemini 3模型对公司业绩指标的影响，没有最初担心的那么大。
• 为应对来自Gemini 3的竞争威胁，奥尔特曼曾在公司内部启动“红色警报”，以集中资源改进ChatGPT。
• 在发布最新大模型GPT-5.2后，奥尔特曼称预计将在明年1月结束“红色警报”状态。
• GPT-5.2将提供即时版、思考版和专业版，分别针对快速文本生成、编码规划等结构化工作以及复杂难题提供优化解答。

相关链接：

• 谷歌紧急上线Gemini Deep Research迎战GPT-5.2，AI“研究代理”大战一触即发
• 谷歌深夜扔出“Disco”浏览器：标签页一键变App，Gemini3 亲自写代码
• 一张自拍搞定全身试衣！谷歌AI虚拟试穿功能迎来重大升级
• 谷歌NotebookLM升级在即：Ultra会员享50倍限额+无水印PPT，AI笔记革命加速
• OpenAI 奥尔特曼：高看了谷歌 Gemini 3，明年 1 月结束“红色警报”

Altman Responds to Gemini 3 Competition: Previously Overestimated Impact, OpenAI to Lift Alert 🚨

Briefing:

• OpenAI CEO Sam Altman stated that the impact of Google's Gemini 3 model on the company's performance metrics was not as significant as initially feared.
• In response to the competitive threat from Gemini 3, Altman had previously initiated a "red alert" internally at the company to focus resources on improving ChatGPT.
• Following the release of the latest large model, GPT-5.2, Altman expects to end the "red alert" status in January next year.
• GPT-5.2 will offer Instant, Think, and Pro versions, optimized for rapid text generation, structured tasks like coding planning, and providing improved solutions for complex problems, respectively. ✨🚀

Related Links:

• Google Urgently Launches Gemini Deep Research to Counter GPT-5.2, AI 'Research Agent' War Imminent
• Google Unveils 'Disco' Browser Late at Night: Tabs Become Apps with One Click, Gemini 3 Writes Code Personally
• Full Body Try-on with Just One Selfie! Google AI Virtual Try-on Feature Receives Major Upgrade
• Google NotebookLM Upgrade Imminent: Ultra Members Get 50x Limit + Watermark-Free PPT, AI Note Revolution Accelerates
• OpenAI's Altman: Overestimated Google Gemini 3, 'Red Alert' to End in January Next Year

英伟达开源自动驾驶AI，模拟人类推理提升决策能力

简报：

• 英伟达在NeurIPS大会上发布了其最新的自动驾驶AI模型Alpamayo-R1（AR1），这是全球首个用于自动驾驶的行业级开放推理视觉语言行动（VLA）模型。
• AR1模型能同时处理文本和图像，并通过结合推理链AI技术来模拟人类的思考方式，以提升车辆在复杂情境下的决策能力，目标是实现四级自动化驾驶。
• 该模型能将汽车传感器“看到”的信息转化为自然语言描述，帮助工程师理解AI的决策过程，从而提升车辆安全性。
• 目前，AR1模型已在GitHub和Hugging Face上开放，供研究人员进行非商业性定制、基准测试或开发。

相关链接：

• ​英伟达发布开放式推理 AI 助力自动驾驶技术迈向新台阶

NVIDIA Open-Sources Autonomous Driving AI, Simulating Human Reasoning to Enhance Decision-Making

Briefing:

• NVIDIA unveiled its latest autonomous driving AI model, Alpamayo-R1 (AR1), at the NeurIPS conference, marking it as the world's first industry-grade open reasoning Vision-Language-Action (VLA) model for autonomous driving. 🚀
• The AR1 model can process both text and images concurrently, simulating human thought processes through a chain-of-thought reasoning AI technique to enhance vehicle decision-making in complex scenarios, aiming for Level 4 autonomous driving. 🧠
• This model translates information "seen" by car sensors into natural language descriptions, helping engineers understand the AI's decision-making process and thereby improving vehicle safety.
• The AR1 model is now open-source on GitHub and Hugging Face, available for researchers for non-commercial customization, benchmarking, or development. 🛣️

Related Links:

• NVIDIA Releases Open Reasoning AI to Propel Autonomous Driving Technology to New Heights

特朗普允许售华H200芯片，但英伟达或为芯片加装“追踪器”

简报：尽管美国总统特朗普已表态允许向中国出口H200芯片，但由于英伟达证实其新芯片将搭载“定位验证”技术以防止走私，外交专家对中国公司是否愿意购买带有“追踪器”的芯片表示怀疑。

相关链接：

• 輝達首證實「定位驗證」技術　專盯晶片有無走私流入黑市

Trump Greenlights H200 Chip Sales to China, But Nvidia May Add 'Trackers' to Chips 🧠

Brief: Despite U.S. President Trump's statement allowing the export of H200 chips to China, diplomatic experts are skeptical about whether Chinese companies would be willing to purchase chips equipped with "trackers." This comes after Nvidia confirmed its new chips will feature "location verification" technology to prevent smuggling. 🕵️‍♂️⚠️

Related link:

• 輝達首證實「定位驗證」技術　專盯晶片有無走私流入黑市

Rivian发布自研智驾芯片，双芯算力达1600TOPS

简报：

• 美国电动汽车制造商Rivian发布其自研的智驾处理器RAP1，采用台积电5nm制程工艺。
• 单颗RAP1芯片的神经引擎具备800 TOPS的INT8稀疏算力，其双芯模块总算力可达1600 TOPS。
• 芯片集成14个Arm Cortex-A720AE核心作为应用处理部分，以及8个Arm Cortex-R52作为安全子系统。
• 包含该芯片和激光雷达的第三代自动驾驶硬件，预计将于2026年底开始在R2车型上交付。

相关链接：

• Rivian 发布自研 5nm 智驾处理器 RAP1：双芯模块 INT8 稀疏算力 1600 TOPS

Rivian Unveils In-House Developed ADAS Chip, Dual-Chip Compute Power Reaches 1600 TOPS

Briefing: 📝

• U.S. electric vehicle manufacturer Rivian has unveiled its in-house developed autonomous driving processor, RAP1, fabricated using TSMC's 5nm process technology.
• A single RAP1 chip's neural engine boasts 800 TOPS of INT8 sparse compute power, with its dual-chip module achieving a total compute power of up to 1600 TOPS. 🚀
• The chip integrates 14 Arm Cortex-A720AE cores for application processing and 8 Arm Cortex-R52 cores for the safety subsystem.
• This third-generation autonomous driving hardware, which includes the chip and lidar, is expected to begin delivery on R2 models by the end of 2026. 🚗

Related Link:

• Rivian Unveils In-House Developed 5nm Autonomous Driving Processor RAP1: Dual-Chip Module Achieves 1600 TOPS INT8 Sparse Compute Power

台积电熊本二厂或因AI需求升级4nm，施工暂停

简报：

• 据日经亚洲报道，为应对激增的人工智能（AI）芯片需求，台积电正考虑将其日本熊本二厂的生产工艺从原计划的6纳米和7纳米，升级至更先进的4纳米技术。
• 此举可能导致该厂重新设计并推迟原定于2027年的量产计划，报道引述知情人士称，10月底动工的厂房施工已暂时停止。
• 消息人士还透露，台积电同时在考虑将先进的芯片封装技术引入日本，但所有相关计划尚未最终确定。

相关链接：

• 台積電熊本二廠傳考慮產「更先進」4奈米晶片！　因應AI需求

TSMC Kumamoto Fab 2 Construction Halted as AI Demand Spurs Potential 4nm Upgrade 🚀

Brief:

• According to Nikkei Asia, TSMC is reportedly considering upgrading the production process at its Japan Kumamoto Plant 2 from the originally planned 6nm and 7nm to more advanced 4nm technology, in response to surging artificial intelligence (AI) chip demand.
• This potential move could lead to a redesign of the plant and a delay in its mass production plans, originally scheduled for 2027. The report, citing informed sources, indicates that construction on the facility, which began in late October, has been temporarily halted. ⏸️
• Sources also revealed that TSMC is simultaneously considering introducing advanced chip packaging technology to Japan 🇯🇵, though all related plans have not yet been finalized.

Related Link:

• 台積電熊本二廠傳考慮產「更先進」4奈米晶片！　因應AI需求

智谱AI开放四项视频生成核心技术

简报：

• 智谱AI在多模态开源周宣布，开源四项专注于视频生成的核心技术：SCAIL、RealVideo、Kaleido和SSVAE。
• SCAIL专注于影视级角色动画生成与姿态控制；RealVideo是实时流式视频生成系统，延迟仅2-3秒。
• Kaleido技术用于确保多主体视频生成的一致性；SSVAE则通过优化训练过程，将模型收敛速度提升三倍。
• 此次开源旨在激发视频生成技术社区的创新，为开发者提供更多工程方案和研究基础。

相关链接：

• 智谱多模态开源周圆满落幕：四项视频生成核心技术全面开放
• Medeo AI新版视频Agent震撼登场：自然语言一键改剧本，内容创作从此零门槛！

Zhipu AI Open-Sources Four Core Video Generation Technologies

Briefing:

• Zhipu AI announced during its Multimodal Open-Source Week the open-sourcing of four core technologies focused on video generation: SCAIL, RealVideo, Kaleido, and SSVAE. 🚀
• SCAIL specializes in cinematic-grade character animation generation and pose control; RealVideo is a real-time streaming video generation system with a latency of only 2-3 seconds. ⚡
• Kaleido technology ensures consistency in multi-subject video generation; SSVAE optimizes the training process, accelerating model convergence speed by threefold.
• This open-sourcing aims to inspire innovation within the video generation technology community, providing developers with more engineering solutions and foundational research.

Related Links:

• Zhipu Multimodal Open-Source Week Concludes Successfully: Four Core Video Generation Technologies Fully Open-Sourced
• Medeo AI's New Video Agent Debuts: One-Click Script Modification with Natural Language, Making Content Creation Effortless! ✨