单词学习：![[23f50bad-d9ca-42ce-899e-772ccb0e080a.mp3]]

新闻播报：![[0b863fe3-00b4-40b2-8d83-ced389a1a441.mp3]]

DeepSeek计划年底发布高级AI智能体

简报：

• 据彭博社报道，DeepSeek正在开发一款具备更先进AI智能体功能的人工智能模型，目标是在今年最后一个季度发布该新产品。
• 该智能体模型能够在用户提供最少指令的情况下，完成规划旅行、编写和调试代码等多步复杂任务。
• 新模型还能根据以往操作不断学习和改进，旨在与OpenAI等美国竞争对手在前沿技术上展开竞争。

相关链接：

• 消息称 DeepSeek 正开发更先进模型：具备 AI 智能体能力，剑指年底发布
• DeepSeek目标在今年年底前发布AI智能体

DeepSeek Plans to Release Advanced AI Agents by Year-End

Brief:

• According to Bloomberg, DeepSeek is developing an AI model with more advanced AI agent capabilities, aiming to release the new product in the last quarter of this year. 🚀
• This agent model will be capable of completing complex multi-step tasks such as planning trips, writing, and debugging code with minimal user instructions. ✨
• The new model will also continuously learn and improve based on past operations, aiming to compete with U.S. rivals like OpenAI in cutting-edge technology. 🎯

Related Links:

• Report: DeepSeek Developing More Advanced Model with AI Agent Capabilities, Targeting Year-End Release
• DeepSeek Aims to Release AI Agent Before End of This Year

上海AI实验室开源书生・万象3.5，宣称多模态感知能力超越GPT-5

简报： 上海人工智能实验室宣布开源通用多模态大模型“书生・万象 3.5”，该模型提供从10亿至2410亿共9种参数尺寸，其旗舰模型据称在多模态通用感知能力上已超越GPT-5，并在多学科推理基准MMMU中获得开源模型最高分。

相关链接：

• 技术报告
• GitHub 开源地址
• 在线体验

Shanghai AI Lab Open-sources InternVL 3.5, Claims Multimodal Perception Capabilities Surpass GPT-5

Brief: Shanghai AI Lab has announced the open-sourcing of its general-purpose multimodal large model, 'InternVL 3.5'. The model offers 9 parameter sizes, ranging from 1 billion to 241 billion. Its flagship model reportedly surpasses GPT-5 in general multimodal perception capabilities and achieved the highest score among open-source models on the MMMU multi-discipline reasoning benchmark. 🚀🧠✨

Related Links:

• Technical Report
• GitHub Open-source Link
• Online Demo

OpenAI完成百亿美元股票出售，估值达5000亿

简报：据CNBC报道，人工智能公司OpenAI已成功向符合条件的员工出售价值103亿美元的二级股票，使其估值在二级交易中从2025年初的3000亿美元飙升至5000亿美元。

相关链接：

• OpenAI 估值飙升至 5000 亿美元，成功出售103亿美元的二级股票

OpenAI Completes Multi-Billion Dollar Stock Sale, Valued at $500 Billion

Brief: According to CNBC, AI company OpenAI has successfully sold $10.3 billion worth of secondary shares to eligible employees, 🚀 propelling its valuation in secondary market transactions from$300 billion in early 2025 to $500 billion. 💰 That's a huge leap!📈

Related Links:

• OpenAI's valuation soars to $500 billion after successful$10.3 billion secondary stock sale

ChatGPT Projects功能向免费用户开放，提供工作空间管理

简报：

• OpenAI宣布，此前为付费用户专享的ChatGPT Projects功能现已向所有免费用户开放。
• 该功能类似一个智能“工作空间”，允许用户为特定任务创建专属对话文件夹，并通过自定义指令调整AI的回答。
• 免费用户现在每个Project可上传最多5个文件，而Plus和Pro用户分别可上传25个和40个。
• 新增“Project-only memory”功能，可限制AI仅引用当前Project内的对话和文件，增强了数据隐私和任务专注度。

相关链接：

• OpenAI放开ChatGPT Projects功能，可以免费用

ChatGPT Projects Feature Now Available to Free Users, Offering Workspace Management

Brief:

• OpenAI announced that the ChatGPT Projects feature, previously exclusive to paid users, is now available to all free users. 🎉
• This feature functions like an intelligent "workspace," allowing users to create dedicated conversation folders for specific tasks and tailor AI responses through custom instructions. 🤖
• Free users can now upload up to 5 files per Project, while Plus and Pro users can upload 25 and 40 files respectively.
• A new "Project-only memory" feature has been added, limiting the AI to referencing only conversations and files within the current Project, thereby enhancing data privacy and task focus. 🔒

Related Links:

• OpenAI unlocks ChatGPT Projects feature, now free to use

微软AI助手Copilot视觉形象功能将全球推广

简报：

• 微软计划在未来数月内，将其AI助手Copilot的视觉形象功能“Copilot Appearance”从当前的美国、英国和加拿大测试市场扩展至全球更多地区。
• Copilot Appearance为AI助手提供了一个拥有动态表情和情感反应的虚拟角色，能够根据对话内容展现微笑、惊讶等实时表情，以提升交互体验。
• 该功能目前仅在网页版上提供支持，微软已明确表示后续将逐步拓展至PC和移动端平台。
• 微软强调该功能在设计上遵循严格的隐私标准，其技术依托于OpenAI的GPT-4及微软自有的Prometheus模型。

相关链接：

• 微软Copilot Appearance即将面向更广泛用户推广

Microsoft Copilot AI Assistant's Visual Persona Feature Set for Global Rollout

Brief:

• Microsoft plans to expand "Copilot Appearance," the visual persona feature for its AI assistant Copilot, from its current test markets in the U.S., U.K., and Canada to more regions globally in the coming months. 🌐
• Copilot Appearance provides the AI assistant with a virtual character that boasts dynamic expressions and emotional reactions, capable of displaying real-time emotions like smiles and surprise based on conversation content, thus enhancing the interactive experience. 😊
• Currently, this feature is exclusively available on the web version, but Microsoft has explicitly stated its intention to gradually expand it to PC and mobile platforms in the future.
• Microsoft emphasizes that the feature adheres to strict privacy standards in its design, powered by OpenAI's GPT-4 and Microsoft's proprietary Prometheus model. 🔒

Related Links:

• Microsoft Copilot Appearance Rolling Out to Broader Audience

谷歌Gemini for Home将取代谷歌助手，并推出新Nest硬件

简报：谷歌预告将公布"Gemini for Home"计划的更多详情，该计划旨在全面取代Google Assistant，并为其开发全新的Nest硬件设备，新系统将支持更自然的语言交互和处理复杂任务。

相关链接：

• 谷歌预告Gemini for Home新进展 全新Nest设备即将亮相

Google Gemini for Home to Replace Google Assistant, New Nest Hardware Incoming 🚀

Brief: Google has teased more details about its "Gemini for Home" initiative, a plan aimed at completely replacing Google Assistant 🤖 and developing entirely new Nest hardware devices for it. The new system will support more natural language interaction and handle complex tasks. ✨

Related Link:

• Google Teases Gemini for Home Progress, New Nest Devices Coming Soon

苹果发力AI搜索，计划2026年春季上线“世界知识问答”

简报：据彭博社消息，苹果公司正全力研发一款名为“世界知识问答”的自研AI搜索引擎，计划于2026年春季上线，旨在集成于Siri以提升其智能搜索能力，并可能与谷歌合作利用其AI模型来支持部分功能。

相关链接：

• 消息称苹果自研 AI 搜索引擎 世界知识问答，计划 2026 年春季上线

Apple Ramps Up AI Search Efforts, Plans Spring 2026 Launch for "World Knowledge Q&A"

Brief: According to Bloomberg, Apple Inc. is reportedly pouring significant resources into developing an in-house AI search engine called "World Knowledge Q&A." 🔍 This ambitious project is set for a Spring 2026 launch and aims to be integrated with Siri, dramatically enhancing its intelligent search capabilities. Furthermore, there's speculation about a potential collaboration with Google, leveraging its AI models to power some of its functionalities. 💡🤖

Related Link:

• Report: Apple Developing In-House AI Search Engine "World Knowledge Q&A," Scheduled for Spring 2026 Launch

Meta发布DeepConf技术，用置信度信号大幅降低大模型推理成本

简报：Meta联合加州大学圣地亚哥分校推出DeepConf新技术，通过评估模型内部的置信度信号来筛选或加权推理路径，从而在大型语言模型的复杂推理任务中，实现在大幅减少（最高达84.7%）计算成本的同时，保持极高（可达99.9%）的准确率。

相关链接：

• Meta 推出 DeepConf 技术，智能平衡大型语言模型的推理成本与准确性

Meta Unveils DeepConf Technology: Drastically Reducing LLM Inference Costs with Confidence Signals 🧠

Brief: Meta, in collaboration with the University of California San Diego (UCSD), has introduced new DeepConf technology. This innovation drastically reduces computational costs (by up to 84.7%) in complex reasoning tasks for Large Language Models (LLMs) while maintaining extremely high accuracy (up to 99.9%). It achieves this by evaluating the model's internal confidence signals to intelligently filter or weight inference paths. 🚀💡📉

Related Links:

• Meta Introduces DeepConf Technology to Intelligently Balance LLM Inference Costs and Accuracy