转自：新智元 编辑：Aeneas 好困 【导读】 48小时前，互联网女皇的一份340页《人工智能趋势报告》一经发布，立刻引起圈内地 震。时隔6年，Mary Meeker依然宝刀未老。硅谷大佬们已经不眠不休，连夜开始研读了！ 互联网女皇、传奇投资者Mary Meeker，再度出山！ 曾经，女皇的《互联网趋势报告》一出，整个科技圈都要抖三抖。硅谷大佬觉都不睡了，都要连夜研 读这份刷屏圈内头条的重磅报道。 蛰伏几年后，她带着一份340页重磅报告，又回来了。 这一次，她瞄准了AI界的当红炸子鸡OpenAI。 在各个创始人和CEO的圈子，这份报告已经全面爆火 在这份340页报告中，51次出现「前所未有」这个词，核心要点就是——AI驱动的这场变革已经全 面且不可逆转，既是机遇遍地的黄金时代，也是奇点的「关键时刻」！ 女皇本皇，五年后回归 Mary Meeker，大名鼎鼎的互联网女皇。 曾经，她是曾是摩根士丹利TMT团队的一员。这个团队分量举足轻重，曾经领导了Netscape的 IPO，这直接开启了1996年的互联网繁荣！ 在1996年，她发布了《互联网趋势报告》的第1版——一份长达322页的深度分析，探讨了网络的潜 力。 ...

Core Viewpoint - The introduction of new wireless charging stone pillars in Shenzhen provides a free charging option for citizens, combining urban intelligence with convenience, although opinions on its practicality vary widely [7][9]. Group 1: Wireless Charging Stone Pillars - The new stone pillars feature video screens and offer free wireless charging, which has sparked discussions among netizens [3][5]. - The operational company plans to expand this project to locations like Wanda Plaza and nationwide, indicating a potential growth opportunity in urban infrastructure [7]. - Supporters view this as a valuable service for those reliant on electronic payments, while critics argue it may create congestion and safety hazards [9]. Group 2: Existing Wireless Charging Solutions - Wireless charging is not a new concept, with implementations seen in various public spaces such as bus stops and park benches [10][11]. - Previous examples include Wuhan's charging streetlights and Chongqing's smart bus stops, which have been positively received by the public [11][13]. Group 3: Challenges of Wireless Charging - Despite advancements, wireless charging still faces issues such as slower charging speeds compared to wired options, with current wireless power reaching up to 50W, while wired charging exceeds 200W [15]. - The need for precise alignment of devices for effective charging remains a significant challenge, often leading to user frustration [16]. - Wireless charging lacks flexibility, as users cannot easily use their devices while charging, which limits its appeal [17]. - Compatibility issues arise from proprietary protocols developed by manufacturers, increasing consumer costs and limiting device interoperability [20]. - Heat generation during charging can lead to performance degradation, particularly in high-temperature environments [21].

Core Viewpoint - The article discusses the paradox of a successful Rust project leading to its eventual abandonment within a company due to management changes and a lack of understanding of its value [2][12][15]. Group 1: Project Background and Development - The project took place in a rapidly growing unicorn startup that primarily used Ruby on Rails and Node.js, with no initial use of high-performance languages like Rust or Go [3]. - The team decided to develop a real-time service to support 100,000 concurrent users, leading to a comparison of different programming languages including Rust, Elixir, Ruby, and Node.js [4][10]. - Rust was ultimately chosen for its flexibility and performance, despite initial bugs, and the team completed the MVP in just two weeks [8][9][10]. Group 2: Performance and Stability - The Rust version outperformed others in speed and memory usage, successfully supporting 100,000 concurrent users without issues [10][11]. - After a month of stable operation, the service was put into maintenance mode as company strategy shifted, leading to the disbanding of the Rust development team [12][13]. Group 3: Management Changes and Consequences - A new management team, unsatisfied with the perceived lack of activity from the Rust service, decided to phase out Rust in favor of more familiar technologies like Ruby and Node.js [13][14]. - The decision resulted in the departure of three experienced Rust engineers, which was seen as a missed opportunity for the company to leverage Rust's capabilities further [15]. - Ultimately, the management's decision to rewrite the Rust service in Node.js led to complications and performance issues, highlighting the challenges of maintaining high concurrency with Node.js [16][17]. Group 4: Industry Insights and Reactions - The article sparked discussions among developers who shared similar experiences of successful projects being undervalued or abandoned due to management decisions [17][18]. - Many industry professionals noted that such occurrences are common in companies as they grow and management changes, often stifling innovation [18].

Core Insights - The article discusses a new research paper by Tri Dao and his team from Princeton University, which introduces two attention mechanisms specifically designed for inference, significantly improving decoding speed and throughput while maintaining model performance [1][2][5]. Group 1: Research Contributions - The paper presents two main contributions: Grouped-Tied Attention (GTA) and Grouped Latent Attention (GLA). GTA reduces KV cache usage by approximately 50% compared to the GQA mechanism integrated into LLaMA 3, while GLA offers faster decoding speeds than the MLA mechanism used by DeepSeek, achieving up to 2x speed improvements in certain scenarios [2][11][22]. - GTA is described as an effective alternative to GQA, and GLA serves as a practical substitute for MLA, maintaining comparable model quality while optimizing memory usage and computational efficiency [3][12]. Group 2: Mechanism Design - GTA combines and reuses the key and value states of different query heads, reducing memory transfer frequency. It groups multiple heads to share the same KV parameters, contrasting with traditional multi-head attention mechanisms that require independent storage for each head [15][16]. - GLA enhances hardware efficiency by increasing the computational load per byte of memory, thereby reducing reliance on memory bandwidth while maintaining parallel scalability for decoding speed [18][19]. Group 3: Experimental Results - The team conducted experiments on models of various sizes (small, medium, large, and XL) trained on the FineWeb-Edu-100B dataset, demonstrating that GTA outperforms GQA in medium to large models, indicating its suitability for further model expansion [22][23]. - The results show that both GTA and GLA maintain or improve downstream task performance as model size increases, validating their effectiveness as alternatives to existing mechanisms [25][37]. Group 4: Performance Metrics - The evaluation metrics included perplexity and accuracy on downstream tasks, as well as efficiency indicators such as decoding latency, throughput, and KV cache usage. GTA reduced KV cache usage by about 50% compared to GQA without sacrificing model quality [27][28]. - GLA demonstrated superior throughput in real-time server performance tests, especially under concurrent requests, indicating its efficiency in handling long contexts and varying request lengths [31][34].

Group 1 - The article discusses the negative perceptions associated with women who enjoy playing Mahjong, suggesting that they may lack life goals and responsibilities [3][5][6] - It highlights that women who frequently play Mahjong may be more prone to infidelity due to their focus on the game and neglect of family duties [5][6] - The article implies that the gambling environment can lead to women being more susceptible to temptations from men, especially when they experience financial losses [5][6] Group 2 - The article presents a narrative about a man's experience with a woman who plays Mahjong, indicating that it caused issues in their relationship [7][9] - It reflects on societal views, suggesting that men generally do not favor women who play Mahjong, which may affect their dating prospects [2][4] - The article includes comments from various individuals expressing their opinions on the topic, showcasing a range of perspectives on women who play Mahjong [10][12][18]

转自：吐槽星君 同事失恋就把工作推给我？ （vi a .@momo ） 我有很重要的事情 ( •) :D % 公众号 · 吐槽星君 我今天和我朋友在外面 ... ( 抱歉抱歉 我 分手了 要是昨天的话可以 -11 找住家 我有好多事情 ·)) 2 公众号 · 吐槽星君 但是我今天和我朋友在外面 我心态要崩溃了 抱歉 晚上可以吗,有两个,或者帮 十十么都可以,我 我做一下 夜里回来做 UB可 我已经一夜没睡了 2 公众号 · 吐槽星君 個景 怎么了怎么了 没事如果你没空我就回 了 夜里做 我也不想麻烦你 但我 实在是没办法了 9 公众号 · 吐槽星君 你做了吗? 如果没做的话 不用做了,我已经回家了 已经解决好了 5月18日 17:08 还没呢 我想回家 6.多去做 好 我自己做。 麻烦你了。 好的 没事众。 5月18日 18:57 t 在不在 我现在又有事情要出去了，你 不用帮我做(要是能做格式也 行呜呜呜呜) 帮我下一下好不 好 放在桌面上 6 公众号 · 吐槽星君 让我觉得无语以及过了这么久还是想发出来 1902 是因为用其他任何事来说都会让人更好接受 但偏偏是分手... 分手是你自己的私事 成年人 分手 ...

Core Viewpoint - Apple is restructuring its six major operating systems (iOS, iPadOS, macOS, watchOS, tvOS, and visionOS) by adopting a new naming convention that uses the year as a suffix instead of traditional version numbers, aiming for better consistency and clarity for users and developers [1][3]. Group 1: Naming Convention Changes - The new naming strategy will see the next iOS version announced at the 2025 Worldwide Developers Conference (WWDC) referred to as iOS 26 instead of iOS 19, aligning with a model similar to the automotive industry [1]. - This adjustment is intended to unify the version numbers across different systems, which currently vary due to their initial release dates, causing confusion among users and developers [3]. Group 2: User Interface Overhaul - A significant user interface redesign, codenamed "Solarium," is set to be the largest in Apple's history, focusing on a glass-like aesthetic for visionOS and extending to tvOS and watchOS, aiming for a cohesive visual language across devices [3]. - The redesign is expected to provide a smoother and more modern user experience, enhancing the transition between different Apple devices [3]. Group 3: New Features in iOS 26 - iOS 26 will introduce a new feature on the lock screen that displays an estimated time to full charge for the device, addressing a common user pain point of guessing charging times [5][6]. - Although this is a minor feature, it is considered practical for users, as iPhones have not previously provided accurate charging time information [6].

Core Viewpoint - The article discusses the challenges faced by job seekers, particularly in the context of campus recruitment, highlighting that the difficulties in securing a job are not solely the fault of the candidates but often due to the hiring practices of companies [1]. Group 1: Job Market Insights - Many candidates feel discouraged and anxious when they cannot find a job, leading to negative mental health outcomes such as depression [1]. - The article reveals that companies may continue to advertise job openings even after filling positions, creating a misleading impression for job seekers [1]. Group 2: Algorithm Problem - The article presents a coding problem from LeetCode, specifically problem number 209, which involves finding the minimum length of a contiguous subarray whose sum is greater than or equal to a given target [4][5]. - The problem is categorized as medium difficulty and requires the use of a sliding window technique to efficiently find the solution [6]. - The provided examples illustrate how to determine the minimum length of the subarray based on the given inputs [6].

Core Viewpoint - The article emphasizes that while AI has made significant advancements, human programmers still possess superior creativity and problem-solving abilities compared to large language models [2][9]. Group 1: Antirez's Experience - Antirez, the creator of Redis, shares his experience in developing Vector Sets and fixing a complex bug, highlighting the limitations of AI in providing innovative solutions [3][9]. - He encountered a performance issue when loading a large vector set, which led him to consult the AI model Gemini for faster solutions, but found its suggestions lacking [5][6]. - Ultimately, Antirez developed a more effective method for checking link interchangeability, demonstrating human ingenuity in problem-solving [8][9]. Group 2: Developer Perspectives on AI - Some developers view AI as a valuable tool for brainstorming and refining ideas, likening it to a "rubber duck" that aids in debugging [10][11]. - However, there are concerns about AI's reliability, with developers noting that it can sometimes provide incorrect suggestions, leading to confusion and wasted time [13]. - Experienced developers can discern AI's limitations, while less experienced ones may struggle to identify errors in AI-generated code, raising concerns about the potential impact on learning programming skills [13][14]. Group 3: Future of Programming with AI - Industry leaders predict that AI will increasingly automate coding tasks, with estimates suggesting that AI could write up to 90% of code in the near future [14][15]. - Despite these advancements, the role of human programmers is expected to evolve rather than disappear, as they will transition to guiding AI in coding tasks [15]. - The article concludes that the focus should shift from whether AI will replace software engineers to how engineers can adapt and evolve alongside AI technologies [15].

Core Viewpoint - Xiaomi's first SUV, the YU7, is expected to be launched soon, with preparations underway for its market entry and public display in various cities [1][5][9]. Group 1: Product Launch and Availability - The Xiaomi YU7 was showcased at the 15th-anniversary strategic product launch event on May 22, but no pricing or pre-order details were provided [1]. - Xiaomi's CEO, Lei Jun, indicated on social media that the YU7 is close to its official launch, emphasizing the importance of final preparations [1][3]. - The YU7 has already arrived at multiple stores, with 13 locations in Beijing set to begin outdoor displays starting June 1, followed by other cities from June 2 [5][10]. Group 2: Pricing Expectations - Lei Jun mentioned that the YU7's pricing is expected to be at least 60,000 to 70,000 yuan higher than the Model Y, which starts at 263,500 yuan [11]. - Speculations suggest that the YU7 will not be priced below 250,000 yuan, with estimates ranging from 250,000 to 350,000 yuan for the entire series [11]. Group 3: Competitive Landscape - Other competitors in the market, such as Huawei and XPeng, have also launched new models recently, indicating a highly competitive environment in the smart automotive sector [13][20]. - The Huawei Zun Jie S800 was recently released with a starting price of 708,000 yuan, while the XPeng MONA M03 Max is priced between 119,800 and 139,800 yuan, showcasing a diverse range of offerings in the market [20][22].