单词学习：![[607f9a83-eace-46e2-9ad6-78347e3940f2.mp3]]

新闻播报：![[1ff3e1f9-432f-4688-9e04-095335fa69fc.mp3]]

MIT新技术让大模型按需“思考”，计算量减半

简报：

• 麻省理工学院（MIT）研究团队发布名为“实例自适应缩放”的新技术，旨在根据问题复杂性动态调整大型语言模型（LLM）的计算资源。
• 该方法通过重新设计推理过程奖励模型（PRMs），使模型在处理简单问题时使用更少资源，处理复杂问题时获得更多推理支持。
• 研究显示，新技术可将计算资源使用量减少一半，同时保持与现有模型相当的回答准确率。
• 该研究得到了 MIT-IBM 沃森人工智能实验室、MIT-Amazon 科学中心及 MIT-Google 计算创新项目等的支持。

相关链接：

• MIT 推出新方法，显著提升大型语言模型计算效率

MIT's New Tech Allows LLMs to "Think" On-Demand, Halving Computation ✨

Brief:

• A research team at the Massachusetts Institute of Technology (MIT) has unveiled a new technology called "Instance-Adaptive Scaling," designed to dynamically adjust Large Language Model (LLM) computational resources based on the complexity of the problem.
• This approach re-engineers the Process Reward Models (PRMs), enabling the model to utilize fewer resources for simpler questions and gain more reasoning support for complex tasks.
• Studies indicate that this new technology can reduce computational resource usage by half, all while maintaining answer accuracy comparable to existing models. 🚀
• The research received support from the MIT-IBM Watson AI Lab, the MIT-Amazon Science Hub, and the MIT-Google Computing Innovation Program, among others. 🧠

Related Links:

• MIT Introduces New Method to Significantly Boost Large Language Model Computational Efficiency

Claude Code集成Slack，开发者可在聊天中直接编程

简报：

• Anthropic宣布其编程代理Claude Code与Slack进行测试版集成，允许开发人员在Slack内直接委派编码任务、修复错误和生成拉取请求。
• 当用户在Slack中提及@Claude时，该系统会自动识别编码任务，并利用上下文信息开启一个新的编程会话，最终提供包含修改链接的拉取请求。
• Claude Code在推出后六个月内年收入已超过10亿美元，客户包括Netflix、Spotify和Salesforce等知名企业。
• 据Anthropic内部调查，其员工使用该工具的时间占工作总时间的60%，生产力提升了50%。

相关链接：

• Anthropic 推出 Claude Code 与 Slack 无缝集成，提升开发效率

Claude Code Integrates with Slack, Empowering In-Chat Coding for Developers

Brief:

• Anthropic has announced a beta integration of its coding agent, Claude Code, with Slack, allowing developers to directly delegate coding tasks, fix bugs, and generate pull requests within Slack. 🤖
• When users mention @Claude in Slack, the system automatically identifies coding tasks, initiates a new programming session using contextual information, and ultimately provides a pull request with links to the modifications.
• Within six months of its launch, Claude Code has generated over dJyH6kDkfC5RlM16.answerText billion in annualized revenue, boasting clients like Netflix, Spotify, and Salesforce. 💰
• According to Anthropic's internal survey, its employees use the tool for 60% of their total work time, leading to a 50% increase in productivity. 🚀

Related Links:

• Anthropic Launches Claude Code for Seamless Slack Integration, Boosting Dev Efficiency

美国称将允许对华出售H200芯片，中方回应望合作共赢

简报：

• 据报道，美国总统特朗普于12月8日表示，美国将允许英伟达公司向中国“经批准的客户”出售其H200人工智能芯片。
• 中国外交部发言人郭嘉昆在12月9日的例行记者会上对此回应称，中方已注意到有关报道，并一贯主张中美通过合作实现互利共赢。

相关链接：

• 美称将允许英伟达向中国出售H200人工智能芯片，外交部回应
• 英伟达黄仁勋称美国三年建一座超算太慢：中国一个周末就能建一座医院

US Says It Will Allow H200 Chip Sales to China, China Hopes for Win-Win Cooperation 🤝

Briefing:

• According to reports, U.S. President Trump stated on December 8 that the United States would allow Nvidia to sell its H200 artificial intelligence chips to "approved customers" in China. 🇺🇸
• Chinese Foreign Ministry spokesperson Guo Jiakun responded to this at a regular press conference on December 9, saying that China has noted the relevant reports and consistently advocates for mutual benefit and win-win outcomes through cooperation between China and the U.S. 🇨🇳

Related Links:

• US says it will allow Nvidia to sell H200 AI chips to China, Foreign Ministry responds
• Nvidia CEO Jensen Huang says it's too slow for the US to build a supercomputer in three years: China can build a hospital in a weekend

黄仁勋警示美国AI基建速度过慢

简报： 英伟达CEO黄仁勋于12月6日表示，美国建设AI超级计算机通常需要三年，速度太慢。他指出中国的优势在于建设速度和能源供给，并以“中国一个周末能建好一座医院”为例进行对比。黄仁勋还提到，中国的可用能源规模是美国的两倍且在持续增长，而美国的水平则长期停滞。

相关链接：

• 英伟达黄仁勋称美国三年建一座超算太慢：中国一个周末就能建一座医院

Jensen Huang Warns US AI Infrastructure Build-Out Is Too Slow 🚧

Brief: NVIDIA CEO Jensen Huang stated on December 6th that it typically takes three years for the United States to build an AI supercomputer, a pace he deems too slow. He highlighted China's advantage in construction speed and energy supply, drawing a comparison by noting, "China can build a hospital in a weekend." 🏥 Huang also mentioned that China's available energy scale is twice that of the U.S. and continues to grow, while the U.S. level has long been stagnant. 📉

Related Links:

• NVIDIA CEO Jensen Huang Says US Taking Three Years to Build Supercomputer Is Too Slow: China Can Build a Hospital in a Weekend

Sora功能命名侵权，OpenAI遭法院禁令

简报：

• 名人定制视频平台Cameo公司提起诉讼，指控OpenAI在Sora应用中推出的“cameo”数字肖像生成功能侵犯其注册商标。
• 美国地区法官已发布临时禁令，禁止OpenAI继续使用该名称；OpenAI随后将该功能更名为“角色”，但并不同意Cameo的指控。
• Cameo公司CEO史蒂文·加拉尼斯称，OpenAI的行为已损害其品牌并削弱其在谷歌搜索中的可见度，关乎品牌“生死存亡”。
• 此事并非孤例，OpenAI计划推出的硬件设备也因拟命名为“io”而遭到另一家公司“iyO”的商标侵权起诉。
• 在法院发布禁令后，OpenAI已将“cameo”标识从Sora应用中移除，下一场决定禁令是否延续的听证会将于12月19日举行。

相关链接：

• Sora功能改名风波
  怒斥OpenAI“损害品牌”，法官紧急下禁令

Sora Feature Naming Infringement: OpenAI Hit with Court Injunction 🚫

Briefing:

• Celebrity personalized video platform, Cameo, has filed a lawsuit alleging that OpenAI's "cameo" digital avatar generation feature within its Sora application infringes on Cameo's registered trademark.
• A U.S. District Judge has issued a temporary injunction, prohibiting OpenAI from continuing to use the name. OpenAI subsequently renamed the feature to "Characters" but disagrees with Cameo's allegations.
• Cameo CEO Steven Galanis stated that OpenAI's actions have damaged its brand and diminished its visibility in Google searches, calling it a matter of "brand survival." 💥
• This incident is not isolated; OpenAI's planned hardware device has also been sued for trademark infringement by another company, "iyO," over its proposed name "io."
• Following the court injunction, OpenAI has removed the "cameo" identifier from the Sora application. The next hearing to determine whether the injunction will continue is scheduled for December 19th. 📅

Related Link:

• Sora feature renaming controversy: Cameo blasts OpenAI for "damaging brand," judge issues urgent injunction

谷歌公布XR蓝图，AI眼镜预计2026年问世

简报：在《The Android Show: XR Edition》特别发表会上，Google宣布计划于2026年联合三星、Gentle Monster、Warby Parker及中国品牌XREAL等合作伙伴，共同推出包括无萤幕和带显示功能两种类型的AI眼镜，以及一款有线XR眼镜。

相关链接：

• Google輕盈AI眼鏡問世時間曝！Galaxy XR新增3大新功能　迎戰蘋果、Meta

Google Unveils XR Blueprint, AI Glasses Expected in 2026 👓

Brief: During 'The Android Show: XR Edition' special presentation, Google announced plans for 2026 to collaborate with partners including Samsung, Gentle Monster, Warby Parker, and Chinese brand XREAL. Together, they aim to launch two types of AI glasses – screenless and with display functionality – along with a wired XR headset. 🚀🗓️

Related Links:

• Google's Lightweight AI Glasses Release Date Revealed! Galaxy XR Adds 3 New Features to Compete with Apple, Meta

上海交大发布光子芯片大模型，研发效率提速7倍

简报：

• 上海交通大学无锡光子芯片研究院发布了全球首个面向光子芯片全链路的垂直大模型LightSeek。
• 该模型基于千亿级参数多模态架构，融合了研究院自建110nm中试线的数十万组真实工艺数据。
• LightSeek可将光子芯片从“设计-仿真-流片-测试”的全周期由传统的6-8个月压缩至1个月，使整体研发效率提升7倍。
• 研究院计划在2025年第一季度开放70B参数的轻量化模型LightSeek-Lite，并提供API接口供企业接入。

相关链接：

• 「设计→流片」1个月搞定！上海交大发布全球首个光子芯片垂直大模型 LightSeek，研发效率提升7倍

Shanghai Jiao Tong University Releases Photonics Chip Large Model, R&D Efficiency Boosted by 7x

Briefing:

• The Shanghai Jiao Tong University Wuxi Photonics Chip Research Institute has released LightSeek, the world's first vertical large model designed for the entire photonics chip development lifecycle.
• This model is built on a multi-modal architecture with hundreds of billions of parameters, integrating hundreds of thousands of real process data sets from the institute's self-built 110nm pilot line. ✨
• LightSeek can compress the full cycle of photonics chips, from "design-simulation-tape-out-testing," from the traditional 6-8 months to just 1 month, boosting overall R&D efficiency by 7 times. 🚀
• The institute plans to open access to a lightweight 70B-parameter model, LightSeek-Lite, in the first quarter of 2025, providing API interfaces for enterprises to integrate. 💡

Related Links:

• "Design → Tape-out" in 1 month! Shanghai Jiao Tong University Releases World's First Photonics Chip Vertical Large Model LightSeek, Boosting R&D Efficiency by 7 Times