整理｜冬梅 近日，据 TechCrunch 报道，谷歌悄然推出了一款名为 Google AI Edge Gallery 的应用（谷歌 称之为"实验性 Alpha 版本"），用户无需 Wi-Fi 连接即可在设备上运行 AI 模型。这款应用目前 仅适用于 Android 系统，允许用户使用 Hugging Face 的几款"公开可用的 AI 模型"，这标志着谷 歌在将先进的 AI 技术直接引入移动设备方面迈出了重要一步。 谷歌悄然发布一款新 AI 应用 这款全新 AI 应用允许用户搜索公开可用的 AI 模型，下载后无需连接 Wi-Fi 即可在本地运行。在 本地运行这些模型时，用户可以要求 AI 执行几乎所有连接 Wi-Fi 时通常会执行的操作，例如生 成图像、与 AI 聊天、浏览网页、编写代码以及提问。 总结起来，主要特点包括： 本地运行，完全离线：无需网络连接，所有处理都可直接在本地设备上进行。 智能模型选择：轻松在 Hugging Face 的不同模型之间切换并比较它们的性能。 图像问答：上传图片并提出相关问题。获取描述、解决问题或识别物体。 Prompt Lab：用户可以使用它来启动由模型驱动的"单轮"任务 ...

本播客由扣子空间（coze.cn）一键生成 整理｜冬梅、核子可乐 2024 年，全球录制音乐市场迎来里程碑时刻，总收入飙升至 296 亿美元。其中，流媒体经济贡献突 出，年收入首次突破 200 亿美元大关。然而，这一蓬勃发展的领域，却因 AI 技术的火爆成为了犯罪 分子眼中的 "肥肉"，利用 AI 技术进行的流媒体欺诈问题日益严峻。 利用 AI 诈骗流量，获利千万美元 近日，《连线》报道的一起涉嫌利用 AI 欺诈的案件引发热议。 在这起涉及 AI 生成音乐的开创性刑事案件中，一名 52 岁的北卡罗来纳州男子被控利用 AI 创作虚假 歌曲，并在 Amazon Music、苹果音乐、Spotify 及 YouTube Music 等音乐平台上进行欺诈性传播以 获取非法版税。 从起诉书来看，利用这些自动账户外加虚拟专用网络（VPN）模拟不同地理位置，被告人恶意操纵 流媒体数据，造成了全球用户都在收听他歌曲的假象。 Smith 承认， 他每天可以产生约 66 万 1440 次流媒体播放，每年版税收入超过 120 万美元 。为了 避免被平台管理员发现，他还将刻意将流量分摊到大量曲目当中。 为了进一步增加利润，Smith ...

本播客由扣子空间（coze.cn）一键生成 作者丨 Nick McGreivy 译者丨明知山、华卫 策划丨华卫 编者按：近日，一位专注于物理领域的科学家 Nick McGreivy 分享了其应用 AI 来做研究的真实经 历，主题是"我被 AI for Science 的炒作所愚弄"。McGreivy 的研究涵盖机器学习、科学计算、等离子 体物理学、核聚变能和核政策等领域，并在 Nature Machine Intelligence、ICLR 等多个知名期刊上发 表过成果论文。 曾经，他对 AI 可以加速物理学研究持乐观态度。但是，当他尝试将 AI 技术应用于实际物理问题时， 结果令人失望。McGreivy 表示，自己对 AI 即将"加速"甚至"革新"科学的观点产生了怀疑。而在其深入 调研后又发现，"即使 AI 在科学中取得了真正令人印象深刻的成果，这也不一定意味着 AI 对科学做出 了什么实质性的贡献，更多只是反映出 AI 在未来可能发挥重要作用的潜力......" 2018 年，当我在普林斯顿大学攻读等离子体物理博士二年级时，我做出了一个重要的决定——将我 的研究重点转向机器学习。当时我还没有一个具体的科研 ...

Group 1: Investment Opportunities and Industry Trends - Recently, Industrial Bank launched the "2025 Elite Internship Program" aimed at high-net-worth clients' children, requiring a minimum deposit of 10 million yuan for new clients and 5 million yuan for existing clients to participate [2][3] - The program faced backlash and was later suspended due to misunderstandings regarding its promotion and the lack of partnerships with major firms like ByteDance [3] - Baidu reported a total revenue of 32.5 billion yuan for Q1 2025, a 3% year-on-year increase, with a net profit of 7.7 billion yuan, up approximately 42% [6] Group 2: Corporate Developments - YuTree Technology has changed its name to Hangzhou YuTree Technology Co., Ltd., indicating potential plans for a Hong Kong listing as discussed with local government officials [5] - Meituan's CEO Wang Xing stated that the company would spare no expense to win competition against JD's recent 10 billion yuan subsidy for its food delivery service [10] - UU Run's management was reported to be required to participate in delivery tasks during lunch breaks, which has raised concerns about employee treatment [11][12] Group 3: Regulatory and Market Changes - The U.S. government is reportedly moving to ban the export of semiconductor design software to China, which could significantly impact the Chinese semiconductor industry [9] - ByteDance announced the internal ban of third-party AI development tools due to data security concerns, leading to employee dissatisfaction [7][8] Group 4: Technological Advancements - Kuaishou's Keling AI achieved a revenue of 150 million yuan in Q1 2025, with nearly 70% of this coming from paid subscriptions by professional users [18] - The DeepSeek R1-0528 model was integrated into Baidu's AI search, enhancing its ability to understand user intent and provide personalized search results [20]

本播客由扣子空间（coze.cn）一键生成 整理 | 褚杏娟 核子可乐 Redis 之父 Salvatore Sanfilippo 近日分享了自己的一次研发经历并直接表达了自己的观点：人 类程序员仍比大模型更出色。"因为我们能够真正打破常规、设想出一些奇特且并不精确、但就是 更有成效的解法，而这对大模型来说则极其困难。" 在社区中，大家更习惯称呼 Sanfilippo 为 Antirez。Antirez 在 2009 年启动了 Redis 项目，并在 2020 年卸任维护者职位，转而担任 Redis Labs 的技术顾问，继续为 Redis 的未来发展提供指 导。Antirez 的分享迅速引发广大开发者的激烈讨论。 Antirez：AI 水平不错， 但远落后人类智能 "今天我要分享一个人类为何仍比大语言模型更有优势的小故事。道德澄清，我并不反对 AI 或者 类似的技术成果，持续关注我的朋友都知道。我经常用大模型，现在也一样。之所以会有这段故 事，是因为我想测试自己的想法、进行代码审查、看看 AI 会不会有比我更好的灵感、探索点专 业范围内的更多可能性之类。"Antirez 在开篇写道，并直接抛出了结论： 总之 ...

Core Viewpoint - Black Forest Labs (BFL) has launched a new image generation model called FLUX.1 Kontext, which allows for both image generation and editing based on contextual inputs, marking a significant shift from traditional methods [1][3]. Group 1: Model Features - FLUX.1 Kontext can generate and edit images based on context, allowing users to modify content without starting from scratch [4]. - The model operates with a flow matching architecture, achieving top character consistency across multiple edits while maintaining interactive inference speeds of 3-5 seconds at 1MP resolution [3][19]. - BFL has released two versions of the model: FLUX.1 Kontext [pro] for rapid iterative editing and FLUX.1 Kontext [max] for enhanced performance and adherence to prompts [16][17]. Group 2: Company Background - BFL was founded in August 2022 by Robin Rombach, a key engineer behind Stable Diffusion, and has quickly gained attention in Europe [6][15]. - The company has received investments from notable venture capital firms such as General Catalyst and Andreessen Horowitz, and its AI models are among the most downloaded [6][15]. - BFL currently employs around 30 staff, with a significant number coming from Stability AI, indicating a strong foundation in AI expertise [14]. Group 3: Competitive Landscape - FLUX.1 Kontext is positioned to compete with established models like MidJourney and Adobe's Firefly, which also offer image generation and editing capabilities [17][30]. - The model's unique flow-based approach differentiates it from diffusion models used by competitors, potentially offering more flexibility in image generation tasks [19][20]. - Early user feedback on FLUX.1 Kontext has been positive, highlighting its impressive performance in generating and editing images quickly [23][28].

Core Insights - Meta has restructured its AI teams into two main groups: an AI product team led by Connor Hayes and an AGI Foundations team co-led by Ahmad Al-Dahle and Amir Frenkel, aiming to enhance product development speed and flexibility [2][3] - The restructuring is a response to increasing competition in the AI space from companies like OpenAI and Google, as Meta seeks to maintain its relevance in the rapidly evolving landscape [3][4] - Despite the restructuring, Meta faces significant challenges, including a talent exodus from its foundational AI research team, FAIR, which has seen 11 out of 14 core members leave [4][8] Team Structure and Focus - The AI product team will focus on consumer-facing applications across platforms like Facebook, Instagram, and WhatsApp, while the AGI Foundations team will work on broader technologies, including improvements to the Llama model [2][3] - FAIR remains independent but has lost key personnel, raising concerns about its future role within Meta's AI strategy [3][4] Talent and Competition - The departure of key researchers from FAIR has led to the emergence of competitors like Mistral, founded by former Meta researchers, which poses a direct challenge to Meta's AI initiatives [8][9] - Meta's recent AI model, Llama 4, has not received a warm reception, leading developers to explore faster-growing alternatives from competitors [9][11] Internal Dynamics and Leadership Changes - Joelle Pineau, who led FAIR for eight years, recently resigned, and her departure has highlighted internal concerns regarding Meta's AI leadership and performance [9][11] - The integration of FAIR into product-focused teams has diminished its role in exploratory research, leading to a shift in priorities towards generating AI-driven products rather than foundational research [18][19] Financial Commitment and Future Outlook - Meta plans to invest approximately $65 billion in AI projects by 2025, indicating a strong commitment to regaining leadership in the AI sector [24] - Despite significant investments, Meta lacks a dedicated reasoning model, which is becoming increasingly important as competitors prioritize such capabilities [27]

作者｜冬梅 采访嘉宾｜谭宇，枫清科技 Fabarta 合伙人、智能引擎事业部总经理 2024 年 11 月 25 日，Anthropic 公司发布的 MCP 协议在推动 AI 技术发展方面具有里程碑式的意 义。 在电子设备领域，USB-C 接口的普及彻底解决了不同设备间的连接难题——无论是充电、数据传输 还是外设扩展，一个接口即可满足所有需求。事实上，从大模型还未如此火爆之前，数据孤岛和工具 碎片化问题就制约 AI 生产力的发展。Anthropic 推出的 MCP 为什么会受到如此高的评价，归根结底 它就是在彻底改变上述局面。 模型上下文协议 (MCP) 是一项开放标准，使开发者能够在其数据源和 AI 驱动的工具之间建立安全的 双向连接。其架构简单易懂：开发者可以通过 MCP 服务器公开数据，也可以构建连接到这些服务器 的 AI 应用程序（MCP 客户端）。 为什么需要 MCP？ 在没有 MCP 之前，开发者需要为每个工具或平台单独定制连接方式，这既耗时又低效。MCP 通过 统一的系统接口解决了这个问题，简化了 AI 与外部服务（如 Slack、Gmail 等）的交互。 在 MCP 发布后，Anthropic ...

昨日，英伟达发布了截至 4 月 28 日的 2026 财年第一季度财报。截至 2025 年 4 月 27 日，英伟达 的第一季度收入为 441 亿美元，环比增长 12%，同比增长 69%。其中，数据中心收入为 391 亿美 元，环比增长 10%，同比增长 73%。 该公司的销售额超过了华尔街预期的 432.8 亿美元，但利润低于预期的 194.9 亿美元。财报公布 后，英伟达股价在盘后几个小时内上涨了约 4%。 同时，英伟达公开披露了特朗普政府最新芯片出口限制对其业务的影响数据，主要围绕 H20。该芯 片是根据美国现行及此前出口规则，英伟达能够出口至中国的最先进人工智能芯片。 H20 被迫"滞销"， 二季度继续翻倍亏 早在美国于 4 月初宣布许可证要求时，英伟达就曾预计其在截至 4 月 27 日的 2026 财年第一季度或 产生 55 亿美元相关费用。 据悉，在新出口许可要求生效前，英伟达第一季度的 H20 产品销售额为 46 亿美元。 整理 | 华卫 由于其受到向中国企业出售 H20 芯片的许可证要求影响，英伟达在第一季度因 H20 库存过剩及采购 义务产生 45 亿美元（约合人民币 323.6 亿元）费用 ...

节前更新似乎已经是 DeepSeek 的惯例了。刚刚，DeepSeek 在 Huggingface 平台开源了 R1 的新 版本 DeepSeek-R1-0528。 项目地址： https://huggingface.co/deepseek-ai/DeepSeek-R1-0528 据悉，新版本主要是在推理精度和代码生成速度的升级。在 Live CodeBench 基准测试中， DeepSeek-R1-0528 的性能可以媲美 OpenAI 的 o3（High）版本。 | | | 8/1/2024 | | 5/1/2025 | | --- | --- | --- | --- | --- | | Rank Model | | Pass@1 ↓ | Easy-Pass@1 | Medium-P | | 1 | 04-Mini (High) | 80.2 | 99.1 | 8 | | 2 | 03 (High) | 75.8 | 99.1 | 8 | | 3 | 04-Mini (Medium) | 74.2 | 98.2 | 8 | | 4 | DeepSeek-R1-0528 | 73.1 | 98.7 | 8 ...