目录
单词学习：![[7eea1fbf-62ba-49d2-83fb-32379b04cfa3.mp3]]
新闻播报：![[51c0657f-b8b5-42c8-9ca0-3976cb2ca20b.mp3]]
中国AI核心产业规模破万亿
简报：根据全国工业和信息化工作会议发布的数据，2023年我国人工智能核心产业规模首次突破万亿元大关。
相关链接：
China's AI Core Industry Exceeds One Trillion Yuan 🚀
Brief: According to data released at the National Industry and Information Technology Work Conference, China's artificial intelligence core industry scale surpassed the one trillion yuan mark for the first time in 2023. 🥳
Related Link:
AI大模型国标正式实施，产业进入规范化新阶段
简报：
- 据市场监管总局消息，我国首部聚焦通用人工智能大模型的系列国家标准已于近日正式实施，标志着大模型产业进入科学、统一的规范化发展阶段。
- 该系列标准明确了对大模型在性能、安全和服务能力方面的具体要求，填补了国内相关技术评价体系的空白。
- 标准评测工具已完成上千项评测任务，调用大模型超95万次，有效识别了幻觉控制、内容安全等共性问题，助力近30家厂商完成技术迭代。
相关链接：
National Standards for AI Large Models Officially Implemented, Industry Enters New Stage of Standardization 🚀
Brief:
- According to the State Administration for Market Regulation (SAMR), China's first series of national standards focusing on general artificial intelligence large models has recently been officially implemented, marking the large model industry's entry into a scientific, unified, and standardized development phase. ✅
- This series of standards specifies detailed requirements for large models in terms of performance, security, and service capabilities, filling a gap in the domestic technical evaluation system. 💡
- The standard evaluation tools have completed thousands of evaluation tasks, invoking large models over 950,000 times, effectively identifying common issues such as hallucination control and content security, and assisting nearly 30 manufacturers in completing technical iterations.
Related Links:
英伟达与斯坦福开源通用游戏AI，可泛化千款游戏
简报：
- NVIDIA与斯坦福大学联合发布了名为NitroGen的通用游戏智能体，旨在突破“一个模型玩一款游戏”的局限。
- 该AI在覆盖超过1000款不同类型游戏的4万小时高质量数据上训练而成，展现出强大的跨游戏泛化能力。
- NitroGen采用纯视觉输入，不依赖游戏内部API，能像人类一样通过观察原始像素玩游戏，并快速适应新游戏。
- 研究团队已将NitroGen模型权重和包含4万小时游戏记录的GameVerse-1K数据集全面开源，以推动AI与游戏研究。
相关链接：
NVIDIA and Stanford Open-Source General Game AI, Playable Across Thousands of Games
Brief:
- NVIDIA and Stanford University have jointly released a general game agent named NitroGen, aiming to break the limitation of "one model for one game." 🎮
- This AI was trained on 40,000 hours of high-quality data covering over 1000 different types of games, demonstrating strong cross-game generalization capabilities.
- NitroGen uses pure visual input, without relying on in-game APIs, enabling it to play games by observing raw pixels like humans do, and quickly adapt to new games. 👁️
- The research team has fully open-sourced the NitroGen model weights and the GameVerse-1K dataset, which contains 40,000 hours of game recordings, to advance AI and game research. 🚀
Related Links:
OpenAI与博通合作定制AI芯片，计划2026年部署10GW算力
简报：
- OpenAI与博通（Broadcom）达成合作，已完成一款定制AI推理引擎的设计，计划在未来五年内部署高达10吉瓦（GW）的计算能力。
- 该芯片专为OpenAI的“o1”系列推理模型和未来的GPT版本量身定制，采用“系统阵列”设计以优化Transformer架构。
- 芯片将由台积电（TSMC）使用其3纳米工艺制造，并集成博通的以太网架构，预计于2026年下半年首次投入数据中心使用。
- 此次合作的目标是通过硬件与软件的共同设计，将每个生成token的能耗降低30%，并增强对AI基础设施的控制力。
相关链接：
OpenAI Partners with Broadcom for Custom AI Chip, Planning 10GW Compute Deployment by 2026
Brief:
- OpenAI and Broadcom have collaborated to finalize the design of a custom AI inference engine, with plans to deploy up to 10 gigawatts (GW) of computing power over the next five years. 🚀
- This chip is custom-tailored for OpenAI's "o1" series inference models and future GPT versions, utilizing a "system-on-array" design to optimize the Transformer architecture.
- Manufactured by TSMC using its 3-nanometer process and integrated with Broadcom's Ethernet architecture, the chips are expected to be first deployed in data centers in the second half of 2026. 💻
- The goal of this partnership is to reduce energy consumption per generated token by 30% through co-design of hardware and software, while also enhancing control over AI infrastructure. 💡
Related Links:
ChatGPT新增富文本编辑功能，可直接排版无需复制
简报：OpenAI正为ChatGPT分批上线名为“格式化模块”的新功能，当用户撰写邮件或博客等结构化内容时，系统会自动生成一个支持加粗、斜体、列表等内置排版工具的富文本编辑区域，从而无需再将内容复制到Word等外部应用中进行二次排版。
相关链接：
ChatGPT Adds Rich Text Editing, Enabling Direct Formatting Without Copy-Pasting
Brief: OpenAI is progressively rolling out a new feature for ChatGPT called "Formatting Module." 🚀 When users are composing structured content like emails or blog posts, the system will automatically generate a rich text editing area with built-in formatting tools such as bold, italics, and lists. This eliminates the need to copy content to external applications like Word for secondary formatting.
Related Links: 👇
北京发布开源生态三年方案，聚焦AI与百个大模型落地
简报：北京市近日正式印发《北京市开源生态体系建设实施方案（2026— 2028 年）》，明确提出到 2028 年，将培育 10 个具有国际影响力的开源项目，打造 30 个国内明星开源项目，并推动 100 个行业大模型实现规模化落地应用。
相关链接：
Beijing Releases Three-Year Open-Source Ecosystem Plan, Focusing on AI and 100 Large Model Implementations 🚀
Brief: Beijing recently officially issued the "Beijing Open-Source Ecosystem Construction Implementation Plan (2026-2028)," clearly stating that by 2028, it aims to cultivate 10 internationally influential open-source projects, establish 30 domestic star open-source projects, and promote the large-scale implementation and application of 100 industry-specific large models. 💡💻
Related Link:
阿里云开源新版通义千问图像模型，专注修复编辑漂移问题
简报：阿里云开源了其最新的图像编辑模型Qwen-Image-Edit-2511，该模型通过引入改进的参考注意力机制等技术优化，重点解决了前代版本中存在的编辑后图像人物或物体位置发生“轻微漂移”的问题，从而显著提升了编辑前后的一致性与视觉稳定性。
相关链接：
Alibaba Cloud Open-Sources New Qwen-Image Model, Focusing on Fixing Edit Drift Issues
Brief: Alibaba Cloud has open-sourced its latest image editing model, Qwen-Image-Edit-2511 🚀. This model, through technical optimizations including an improved reference attention mechanism, primarily addresses the "slight drift" issue where people or objects in edited images would subtly shift from their original positions—a problem found in previous versions 🛠️. This significantly enhances consistency and visual stability before and after edits ✨.
Related Links:
本文作者:topwind
本文链接:
版权声明:本博客所有文章除特别声明外，均采用 BY-NC-SA 许可协议。转载请注明出处！