整理｜冬梅 2025 年 10 月 23 日，Microsoft 在其官方博客发布了 "Copilot 秋季发布版"，标志着其人工智能助手 Copilot 平台迈入一个新的 阶段。 微软 CEO 人工智能业务主管 Mustafa Suleyman 在博客中指出，这次发布体现其"以人为本的 AI"（human-centred AI）理 念，强调"科技应服务于人，而不是让人服务于科技"。 简单来看，此次发布有三个关键词：协作（Together）、个性化（Personalised）、连接（Connected）。微软借此将 Copilot 从过去主要作为一个"生产力工具插件"（例如 Office 助手）升级为跨设备、跨场景、深入日常生活和工作流程的"情境 AI 基础 设施"。 此举被一些外媒评论为："这是微软将其生产力体验与生成式 AI 能力更深整合的一次重大举措"。 微软更新 Copilot， 发布 12 项关键功能 那么，围绕着 Copilot 发布的 12 项关键功能都有什么？ Office 大眼夹（Clippy）退役 24 年后，Mico 重新上岗 Mico 这个名字是" Microsoft "和"Copil ...

Core Insights - OpenAI released gpt-oss in August 2025, providing two reasoning models: 120b and 20b, which gained support from major cloud providers and inference engines [3] - The model architecture follows mainstream designs, utilizing tiktoken for tokenization, MoE architecture, and various optimizations for efficiency [5][9] - The Java port of gpt-oss achieved a high-performance CPU inference engine with approximately 1000 lines of code, demonstrating the feasibility of running LLMs on CPU [3][37] Model Architecture Overview - gpt-oss retains a conventional model architecture, employing techniques like Grouped Query Attention and MoE to balance model capability and inference efficiency [5] - The 20b model is structured with 24 layers, each containing 32 experts, activating only 4 experts per forward pass to reduce computational load [5] - The model file size for the 20b version is approximately 13GB due to mxfp4 quantization [5] Implementation Process - The Java porting process involved replicating the original PyTorch model structure, focusing on key implementations and performance optimizations [9][10] - The model's MLP layer parameters are quantized using mxfp4, optimizing memory requirements during inference [12] Performance Optimization - Initial performance on AWS EC2 was 0.04 tokens/sec, but optimizations improved this to approximately 7 tokens/sec for decoding and 10 tokens/sec for prefill [23][34] - Matrix multiplication optimizations included cache optimization, vectorization, and parallel processing, achieving significant performance gains [24][28] - The final implementation on AWS EC2 reached 61.4 GFLOPS, representing 42% of the machine's peak performance [27] Memory Management - The project utilized Java Foreign Memory API for memory mapping, allowing the model to run with only 16GB of memory [29] - Memory copy reductions were achieved by pre-allocating intermediate data and using mmap for MLP weights [30] Conclusion - The project demonstrated the potential of Java for high-performance LLM inference, with ongoing improvements in Java's performance capabilities [38] - The experience highlighted the importance of engineering optimizations in LLM inference, distinguishing it from pre-training and post-training processes [37]

Core Insights - Meta is laying off approximately 600 positions from its AI department, specifically within the "Superintelligence Lab" [2][3] - The layoffs aim to streamline operations and reduce bureaucracy, allowing for more efficient decision-making and greater individual responsibility [3][18] - The restructuring follows a series of internal changes and dissatisfaction with the performance of the AI teams, particularly regarding the development of large models like Llama 4 [11][12] Summary by Sections Layoffs and Restructuring - Meta is cutting around 600 jobs in its AI division, affecting teams such as FAIR and product-related AI groups, while the newly established TBD Lab continues to hire [2][3] - The layoffs are part of a broader strategy to create a more agile and talent-dense team structure, as stated by Meta's Chief AI Officer Alexandr Wang [3][16] Internal Dynamics and Reactions - Employees affected by the layoffs have been encouraged to apply for other positions within Meta, with expectations that many will find new roles [3][18] - The layoffs have sparked discussions about internal politics and the effectiveness of the current leadership, with some employees expressing skepticism about the reasons given for the cuts [19][21] Development Projects and Future Plans - The Superintelligence Lab has been divided into four sub-departments, focusing on various aspects of AI research and product development, including the management of the latest large model projects [10][17] - Meta's CEO Mark Zuckerberg has been actively involved in AI recruitment and project oversight, aiming to build a team focused on achieving artificial general intelligence (AGI) [11][13] Financial and Strategic Moves - Meta plans to invest significantly in AI, including a $14.3 billion investment in Scale AI, while also exploring various acquisition opportunities to bolster its AI capabilities [12][13] - The company has been aggressively recruiting talent from competitors, leading to a rapid expansion of its AI teams, which has also resulted in internal friction and turnover [15]

Core Insights - The GTLC Global Technology Leadership Conference is a premier event organized by TGO Kunpeng Club, focusing on technology leadership and innovation since its inception in 2016 [2] - The upcoming conference in Chengdu on October 25, 2025, will center around the theme "AI New 'Shu' Light," emphasizing the AI application ecosystem and featuring over 20 top observers and practitioners from various fields [3][4] Event Details - The conference will take place at Chengdu · Jingronghui, with multiple high-quality keynote speeches, closed-door lunch meetings, and themed discussions to facilitate deep exchanges among technology leaders [4][16] - The agenda includes a variety of sessions, such as discussions on the future of intelligent driving, AI applications in different sectors, and the transformation of traditional enterprises through AI [7][10][11] Participation Information - The ticket price for the conference is ¥2999 per person, while TGO Kunpeng Club members can attend for free and invite three eligible friends [20][21] - Non-members interested in attending can apply for free tickets, subject to approval [22] Additional Activities - The conference will feature a Programmer's Day celebration on October 24, including a welcome dinner and a friendly football match, along with various engaging activities post-conference [17][18]

AI 时代，技术人的棋盘正在重构。 10 月 24 日上午 9:00-10:30，InfoQ《C 位面对面》1024 特别策划，极客邦科技 CEO 霍太稳对话阿里云 CIO 蒋林泉，围绕 AI 时代的人才进化、组织 协同与工具实践进行讨论。从开发者到 CIO，蒋林泉将在直播间与大家一起复盘 AI 落地的真实路径以及技术人如何构建"可迁移"的核心竞争力，打开技 术人下一个十年的成长密码。 直播介绍 直播时间 10 月 24 日 9:00-10:30 如何看直播？ 扫描下图海报 【二维码】 ，预约 AI 前线视频号直播。 直播主题 AI 如何重塑开发者未来十年 直播嘉宾 直播亮点 阿里云智能集团副总裁、CIO 蒋林泉 极客邦科技创始人兼 CEO 霍太稳 从开发者到阿里云 CIO 的心路历程； AI 时代，技术人才与角色的重构； 技术人如何构建"可迁移"的核心竞争力； 推荐使用的 AI 提效工具； AI 开发范式下的组织落地实践； 组织拥抱 AI 的文化与心理建设； 如何向讲师提问？ 文末留言写下问题，讲师会在直播中为你解答。 ...

模力工场 新鲜事： 以"年青人因科技而团聚"为愿景的全球科技社区活动——2050 恳谈会，于 10 月迎来杭州主会场和北京分会场特别活动。 本次北京分会场活动由模力工场携手极客时间八周年共同发起，在极客邦科技总部举办，以"AI Builder"为主题，汇聚开发者、产品人与创作者，共同探 讨 AI 应用时代的创造力与工具变革。 杭州主会场分享者阵容包括杭州久痕科技创始人、网易研究院原院长汪源，SegmentFault 创始人左志鹏等行业代表；北京站则邀请了 DeepPath 时踪 创始人王泰、智能投标助手团队等开发者进行实战分享。 模力工场秋季赛 进行中，本周榜单如下，恭喜各位上榜的同学！完整榜单见模力工场网页端或小程序端秋季赛区域。 | | △ 应用榜单 | | | | 图 用户榜单 | | | ■ 城市榜单 | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | 秋零总榜 | 应用点赞榜 应用传播榜 | 应用热评榜 | | 模力开发者榜 | 傾力推荐人榜 | 模力体验官榜 | 被市总榜 | 一线&新一线城市榜 | 二线及以下城市 ...

Core Viewpoint - OpenAI's launch of the ChatGPT Atlas browser represents a significant shift in the browser landscape, aiming to redefine how users interact with the web by integrating AI capabilities directly into the browsing experience, posing a direct threat to Google's dominance in the browser market [2][10][11]. Group 1: Product Features and Innovations - ChatGPT Atlas is designed to be an intelligent assistant embedded within the browsing experience, utilizing the open-source Chromium engine and deeply integrating with ChatGPT to enhance user interaction [6][10]. - Key features include a sidebar assistant that can answer questions related to the current webpage, a default search engine powered by ChatGPT, and a recommendation engine that suggests content based on user behavior [6][9]. - The "Agent Mode" allows paid subscribers to execute tasks autonomously, such as booking hotels or editing documents, while maintaining user control over the process [7][9]. Group 2: Market Impact and Competitive Landscape - The launch of Atlas has already impacted Google's stock, with a notable drop in share price, indicating investor concern over the potential loss of market share [2][10]. - OpenAI's approach to redefining search through a conversational model contrasts sharply with Google's traditional search methods, which rely heavily on advertising [11][12]. - The competition is intensifying, with other companies like Perplexity and The Browser Company also entering the AI browser space, suggesting a new era of browser wars [15][16]. Group 3: User Experience and Feedback - Initial user feedback on Atlas has been mixed, with some praising its clean interface while others criticize its functionality and integration with existing tools [16]. - Users express concerns about data privacy and the willingness to entrust their online activities to OpenAI, reflecting the strong user habits tied to existing browsers like Chrome [16][17]. Group 4: Future Considerations and Challenges - The emergence of AI-driven browsers raises new security concerns, particularly regarding the potential for AI to execute harmful commands without user awareness [17]. - OpenAI's future success with Atlas will depend on its ability to attract users and generate revenue, as well as its strategy for addressing privacy and security issues [13][17].

作者 | 傅宇琪、Tina 在 AI 写码逐渐成为"新常态"的当下，编程语言的选择反而更重要。 Flask 作者、创业者 Armin Ronacher 指出：语言背后是复杂的权衡，在 AI 时代必须 被重新审视；更关键的是，语言会直接影响 Agent 生成代码的质量。 Armin 认为，Go 在 AI 场景下更"对味"，它抽象层薄、结构规整，便于模型读懂与改写。同一类小程序，他让 AI 分别用多种语言各写十次再比通过率， Go 明显优于 Python，也好于 Rust。与此同时，他也强调一个现实：无论你创办什么公司，最终都绕不开 Python。也许不会用它写核心服务，但只要 涉及机器学习或数据处理，Python 一定会出现。同理，JavaScript 也无法回避；既有 JavaScript，通常就会有 TypeScript。 更长远地看，"为人类与 Agent 共编而设计的下一代语言"正成为行业趋势。Armin 认为，现在正是创造"更完美语言"的窗口期：我们短期内不会摆脱 AI 生成代码的范式，而现有语言也未必是人机协作的最优解。那么，如今的创业公司在 Python / Go / Rust / TypeScr ...

Core Insights - DeepSeek has released a new model, DeepSeek-OCR, which is a 6.6GB model specifically fine-tuned for OCR, achieving a 10× near-lossless compression and a 20× compression while retaining 60% accuracy [2] - The model introduces DeepEncoder to address the trade-offs between high resolution, low memory, and fewer tokens, achieving state-of-the-art performance in practical scenarios with minimal token consumption [2][4] - The model's architecture is lightweight, consisting of only 12 layers, which is suitable for the pattern recognition nature of OCR tasks [5] Model Innovations - DeepSeek-OCR allows for rendering original content as images before input, leading to more efficient information compression and richer information flow [6] - The model eliminates the need for tokenizers, which have been criticized for their inefficiencies and historical baggage, thus enabling a more seamless end-to-end process [6] - It employs a "Mixture of Experts" paradigm, activating only 500 million parameters during inference, allowing for efficient processing of large datasets [7] Market Position and Future Implications - Alexander Doria, co-founder of Pleiasfr, views DeepSeek-OCR as a milestone achievement, suggesting it sets a foundation for future OCR systems [4][8] - The model's training pipeline includes a significant amount of synthetic and simulated data, indicating that while it has established a balance between inference efficiency and model performance, further customization for specific domains is necessary for large-scale real-world applications [8] Developer Engagement - The release has attracted many developers, with Simon Willison successfully running the model on NVIDIA Spark in about 40 minutes, showcasing the model's accessibility and ease of use [9][21] - Willison emphasized the importance of providing a clear environment and task definition for successful implementation, highlighting the model's practical utility [24]

最近和几位企业管理者交流，发现大家在 AI 投入上普遍陷入两种困境： 有的企业盲目跟风"全员 AI "，斥资数百万购买系统、组织培训，结果员工只学会了用 AI 聊天、做 PPT，业务场景依然原地踏步；有的企业则因" AI 替代焦 虑"仓促调整组织架构，反而导致团队士气低落、业务衔接不畅。 这些现象背后，反映了一个共同问题：大多数企业的 AI 投入，都走错了方向。 真正的 AI 能力建设，从来不是靠堆砌预算或盲目调整团队，而是要找到那个能同时实现"技术普及"与"业务价值"的精准切入点。 一次零成本的 AI 能力提升机会 正是看到企业在 AI 落地中的这些痛点，在极客时间企业版 8 周年之际，我们推出了「 AI 应用全员加速中 」特别活动——旨在让企业完全零成 本验证 AI 人才培养的可行性。 从现在到 10 月 31 日，企业可免费申领 30 天 SVIP 权益，不限账号数量，让全体员工无障碍体验平台上的 AI 课程资源。 这不是又一次"蜻蜓点水"的体验，而是一次完整的 AI 能力建设验证： 过去，企业要启动同等规模的 AI 培训，至少需要数十万的预算投入和数月的筹备期。现在，这个门槛被彻底打破了。 为什么这次 ...