转自：电脑报 近日，微信对朋友圈互动提醒进行了优化。 将微信更新至最新版本后，进入【我】—【朋友圈】—【我的朋友圈】— 右上角【···】— 【消息列表】— 再次点击右上角【···】进入【消息管理】，此时多了一个"仅提醒朋友与我的互动"设置页面。 如果右上角显示了清空，表示你没有被内测，可以等后续全面覆盖。 开启该选项后，朋友点赞或评论用户互动过的朋友圈时，发现页将不再提醒，但消息列表仍可查看。 之前，就有许多用户反馈，当点赞了一条朋友圈后，这条内容被其他好友点赞，发现页面会不断收到提示， 互动消息太多备受打扰，如果能选择只通知跟自己相关的点赞、评论就好了。 这样的更新正好解决了用户的烦恼，可以大大减少信息干扰，提升浏览效率。当你有大量好友在你互动过的 朋友圈下互动时，系统会只展示与自己相关的信息，清静多了。 部分网友发现，现在朋友圈消息列表已支持区分"与我的互动"和"朋友的互动"了。 打个比方，当你好友发了一条朋友圈消息，你和其他好友进行点赞或评论后，虽然发现页会有小红点提示， 但点击进入消息列表后，所有消息会被分类展示了，看起来更有条理了，同时也不会错过重要信息。 点击进入"全部互动消息"也可以看到所有的互 ...

Group 1 - Apple is considering utilizing AI models from OpenAI and Anthropic for an updated version of Siri instead of relying solely on in-house technology [1] - The company is developing an internal project named "LLM Siri" that incorporates its own AI models while also seeking external assistance for training [1] - Apple has requested OpenAI and Anthropic to create versions of their AI models compatible with Apple's cloud infrastructure for testing purposes [1] Group 2 - The launch of AI-enabled Siri has been postponed from 2025 to 2026 or later due to various technical challenges faced by the company [2] - Apple has been lagging behind competitors like Google, OpenAI, and Anthropic in the AI sector over the past few years [2] - Although Siri can currently utilize ChatGPT for complex inquiries, Apple is now looking into a more extensive integration with third-party AI technologies [2]

Core Viewpoint - Kunlun Wanwei is intensifying its investment in AI, particularly through strategic funding and partnerships, despite facing significant financial losses and challenges in commercializing its AI business [1][6][7]. Group 1: Financial Performance - In 2024, Kunlun Wanwei reported revenue of 5.66 billion yuan and a net loss of 1.95 billion yuan, marking its first loss in ten years [7]. - By Q1 2025, the company continued to experience losses, with revenue increasing by approximately 46% year-on-year to 1.67 billion yuan, while net losses quadrupled to 930 million yuan [7][8]. - The company’s sales expenses surged by 92% to 868 million yuan in Q1 2025, primarily due to increased marketing costs for AI software and short drama businesses [8][21]. Group 2: Investment and Funding - Kunlun Wanwei's affiliated fund plans to provide 200 million yuan in interest-free convertible bonds to its subsidiary Skywork AI Inc. to enhance capital efficiency [1]. - The company has applied for a total of 1.27 billion yuan in bank credit since the beginning of 2025, nearly doubling its credit requests compared to 2024 [1][5]. - In June 2023, a major shareholder provided a loan exceeding 50% of the after-tax proceeds from a share reduction to support the company's long-term development in AGI and AIGC businesses [5]. Group 3: Strategic Focus and Challenges - The company is heavily investing in AGI and AIGC sectors but is currently facing a funding gap and has not yet achieved AI commercialization [6][12]. - Kunlun Wanwei attributes its losses to geopolitical tensions affecting its investments in companies like Pony.ai and increased R&D expenses for AI technologies [8][12]. - The company anticipates entering a profitable phase post-2027, despite current heavy investments in marketing and R&D leading to substantial losses [12]. Group 4: Corporate Structure and Operations - Kunlun Wanwei's subsidiaries, including Skywork AI, have reported significant losses, with four out of five major subsidiaries experiencing substantial net losses due to R&D and marketing expenses [12][18]. - The company has engaged in strategic equity transfers among its subsidiaries, which has resulted in inflated goodwill values, raising concerns about the sustainability of its financial practices [18][32]. - The increase in goodwill from 960 million yuan in 2020 to 5.15 billion yuan in Q1 2025 indicates potential risks associated with overvaluation and future cash flow challenges [29][32].

Core Insights - Baidu officially open-sourced the Wenxin large model 4.5 series on June 30, which includes 10 models with mixed expert (MoE) architecture and dense models, featuring 47 billion and 0.3 billion parameters respectively [1] - The open-sourced models are available for download and deployment on platforms like PaddlePaddle and HuggingFace, with API services accessible through Baidu's intelligent cloud [1] - The Wenxin model 4.5 series introduces an innovative multi-modal heterogeneous model structure, enhancing multi-modal understanding while maintaining or improving text task performance [1] Group 1 - The Wenxin model 4.5 series is fully open-sourced under the Apache 2.0 license, supporting academic research and industrial applications [1] - Baidu has achieved a "dual-layer open-source" strategy with both framework and model layers, establishing a significant AI full-stack technology advantage [2] - PaddlePaddle, as China's first self-developed and open-source industrial-grade deep learning platform, has been enhanced with the release of the Wenxin model development suite ERNIEKit and the efficient deployment suite FastDeploy [2] Group 2 - The open-sourced series is designed to lower the barriers for post-training and deployment, being widely compatible with various chips [1][2] - The advancements in the Wenxin model 4.5 series are attributed to key technologies such as multi-modal mixed expert model pre-training and efficient training inference frameworks [1]

Core Insights - Tencent's Hunyuan-A13B is the first MoE model with 13B active parameters, featuring a total of 80B parameters and a native context length of 256K, achieving inference speeds over twice that of similar models, with daily calls exceeding 130 million [1][12][14] Model Performance - Hunyuan-A13B demonstrates strong general capabilities, achieving high scores on various authoritative industry benchmarks, particularly excelling in agent tool usage and long text processing [4][8][12] - In specific benchmark tests, Hunyuan-A13B outperformed other models in mathematics, reasoning, and instruction-following tasks, with notable scores such as 87.3 in AIME2024 and 91.8 in IF-Eval [7][10] Deployment and Accessibility - The model is available on open-source platforms like GitHub and Hugging Face, and its API is accessible via Tencent Cloud, allowing for easy integration and deployment [2][5][11] - Hunyuan-A13B is designed to be developer-friendly, requiring only a single mid-range GPU for deployment, and supports various quantization formats [11][12] Training Innovations - The model was pre-trained on 20 trillion tokens, covering multiple domains, which significantly enhances its general capabilities [12][13] - Tencent has developed a multi-stage training approach to improve the model's reasoning abilities while maintaining its creative and understanding capabilities [13][14] Future Developments - Tencent plans to continue its commitment to open-source by releasing more models of varying sizes and capabilities, including dense models ranging from 0.5B to 32B and MoE models with 13B active parameters [15]

Core Points - Apple announced adjustments to its App Store policies in the EU to avoid potential fines of up to €500 million (approximately $585 million) under the Digital Markets Act (DMA) [1] - The company is facing increasing digital regulatory scrutiny and is balancing rules against profitability [1] - Apple plans to implement a new fee structure, including a "Core Technology Fee" of 5% for developers even for purchases made outside the App Store [1][2] - The overall cost burden for developers may increase significantly, with some facing up to three fees during a single app installation [1] - The European Commission is evaluating whether Apple's new policies comply with the DMA and will seek input from market stakeholders before making a final decision [1] Industry Context - Apple's policy shift is not isolated, as it faces regulatory challenges in the U.S., including a court ruling that prevents it from charging commissions on apps that direct users to third-party platforms [2] - Companies like Amazon and Spotify have begun incorporating direct link features in their iOS apps to bypass Apple's commission structure of 15% to 30% [2] - Despite policy changes, Apple maintains its stance that it deserves fees for "core technology support" even if transactions occur outside the App Store, which has drawn criticism as "malicious compliance" [2] - The ongoing situation may serve as a landmark case for global platform companies navigating an era of heightened regulation [2]

Core Points - Apple announced updates to its developer policies to comply with the EU's Digital Markets Act (DMA) ahead of the June 26 deadline to avoid fines [1][2] - The updates include new "anti-steering" rules allowing EU app developers to link to alternative payment methods outside the App Store [3] - Apple introduced a new fee structure, replacing the Core Technology Fee (CTF) with an initial acquisition fee of 2% and a store services fee of either 13% or 5% based on developer tiers [4][5] Developer Fee Structure - The new fee structure includes a Core Technology Commission (CTC) that will replace the CTF, with a 5% commission for developers on standard terms in the EU [6][7] - Developers previously paying the CTF of €0.50 per app install after 1 million downloads will transition to the new rules by January 1, 2026 [7] - Tier 1 developers have limited access to App Store services, while Tier 2 developers gain access to more comprehensive tools and features [5] Industry Reaction - Epic Games CEO Tim Sweeney criticized Apple's compliance with the DMA, labeling it as "malicious compliance" and arguing it undermines fair competition in digital markets [8]

Core Insights - Huawei Developer Conference 2025 (HDC 2025) showcased advanced technologies and the future of the Harmony ecosystem, indicating significant ongoing effects for users and ecosystem partners [1] Group 1: AI Empowerment and Browser Upgrades - Huawei Browser introduced multiple innovative features and upgrades, leveraging AI technology for a comprehensive upgrade in information search and service [2] - The new browser information assistant utilizes authoritative content sources and cross-platform integration to become a reliable information partner for users [2] - Users can ask questions to the browser, which employs AI's multi-dimensional semantic understanding to provide logical and organized responses, enhancing information retrieval efficiency [2][4] Group 2: Advantages Over Traditional Browsers - Unlike traditional browsers that merely list search results, Huawei Browser's AI summarization extracts key points from lengthy articles, saving users time and effort [4] - AI recommendations tailor content based on user browsing history and preferences, broadening the scope of information available to users [4] - For sensitive users like investors and tech enthusiasts, the browser offers in-depth analysis and professional insights, aiding in informed decision-making [4] Group 3: Empowering Ecosystem Partners - The content strategy upgrade provides ecosystem partners with broader content distribution channels and more precise user targeting capabilities [5] - Huawei Browser optimizes algorithms to deliver high-quality visual content to target user groups, enhancing exposure and engagement [5] - Video content distribution has been improved, supporting high-definition playback and intelligent recommendations based on user habits and device types [5][7] Group 4: Enhancing Content Creation and Distribution - Video creators and content providers can present their work more attractively, reaching a larger audience within the Harmony ecosystem [7] - The browser offers dedicated spaces for audio content, expanding distribution channels for partners like audiobook platforms and radio stations [7] - AI capabilities enhance content distribution efficiency through deep user behavior analysis, leading to higher conversion rates for partners [7] Group 5: Localized Digital Economy Development - Huawei Browser collaborates with partners to build local services, injecting vitality into the localized digital economy [9] - The focus on user experience and ecosystem development is crucial for product success, with Huawei Browser's innovations providing better services for users and more opportunities for partners [9] - The browser's upgrades position it as a leading tool in the digital age, enhancing its role in the internet landscape [9]

RA团队 发自 凹非寺 量子位 | 公众号 QbitAI 总是"死记硬背""知其然不知其所以然"？ 奖励模型 训练也形成了学生选择标准答案的学习模式，陷入诸如"长回答=好回答""好格式=好答案"等错误规律之中。 北京大学知识计算实验室联合腾讯微信模式识别中心、William&Mary、西湖大学等机构提出的 RewardAnything 突破了这一瓶颈 ——通过让奖励模型直接理解自然语言描述的评判原则，实现了从"死记硬背"到"融会贯通"的范式跃迁。 RewardAnything降低了传统模式针对不同场景需要收集偏好数据训练奖励模型再进行RL的高昂成本，能够直接利用自然语言作为 RLHF的标准。 其作为奖励模型，仅需一句话描述的准则即可刷新传统Benchmark的SOTA，在RABench上展示出了与GPT-4.1等顶尖模型相媲美的 原则跟随能力与泛化能力。 | Model | Domains | | | | | Principle Categories | | | Overall | | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- ...

Core Viewpoint - The article discusses the innovative AI applications developed by Google Labs, emphasizing their fun, practical, and diverse nature, moving beyond traditional model updates and parameters [3][5][90]. Group 1: Google Labs Overview - Google Labs has over thirty AI products that are either open or will soon be available, showcasing significant innovation [5]. - The project aims to help users learn in more engaging ways, enhance productivity, and integrate AI into daily life [5][90]. - Google Labs was initially launched in 2002 to foster creativity among employees, allowing them to dedicate 20% of their time to innovative projects [92][94]. Group 2: Featured AI Products - **National Gallery Mixtape**: Generates music based on a given painting, enhancing the understanding of art through sound [10][11][28]. - **Learn About**: An AI-assisted learning tool that structures knowledge acquisition, providing a framework, interactive learning methods, and self-assessment [29][32][54]. - **Little Language Lessons**: A practical language learning tool that focuses on relevant vocabulary, local dialogues, and real-life applications [54][60][66]. - **Stitch**: A design tool that allows users to create UI interfaces using natural language or image references, streamlining the design process [67][72][78]. - **Portraits**: A workplace tool featuring a virtual expert that provides advice on professional challenges, simulating real-life interactions [80][82][86]. Group 3: Google Labs' Evolution - Google Labs was shut down in 2011 as the company shifted focus to core business needs, but innovation efforts continued through other projects [94][96]. - The resurgence of Google Labs is characterized by a focus on small-scale projects, rapid iteration, and future-oriented product development [100][104][106]. - The approach aims to maintain vitality in a rapidly changing AI landscape, emphasizing innovation as a key productivity driver [108][110].