以下文章来源于极客公园 ，作者Moonshot 极客公园 . 用极客视角，追踪你最不可错过的科技圈。欢迎同步关注极客公园视频号 本文来自微信公众号： 极客公园 （ID：geekpark） ，作者：Moonshot，编辑：靖宇，题图来源： 视觉中国 AI，已经热了快三年了。 各大科技巨头争先恐后下注入局，可偏偏在这个热潮中，最接近我们生活的苹果，却看起来离AI最 远。 最大的巨头，在最热的潮流面前，好似隐身了。 去年6月WWDC上，苹果慢吞地发布了Apple Intelligence，可如今快一年过去，对大部分用户来说， Apple Intelligence依旧只闻其声、不见其形。 全世界都看到苹果的AI做不好了，但没人知道到底发生了什么 。 知名苹果分析师Mark Gurman刚刚在外媒发出一篇长文，题为《Why Apple Still Hasn't Cracked AI》 （为何苹果仍未攻克人工智能） ，揭露了苹果内部对AI态度的摇摆，内部的斗争和难以克服的 技术瓶颈。 值得注意的是，Gurman用的是"Still hasn't （仍未） "，这词就已经给苹果的现状定了调。 本文将通过重组原文以呈现苹果在A ...

一水 发自 凹非寺 量子位 | 公众号 QbitAI 只需知道6项个人信息，GPT-4就有可能在辩论中打败你？！ 而且 胜率高达64.4% 。 这是几位来自瑞士洛桑联邦理工学院、普林斯顿大学等机构的研究人员得出的最新结论，相关研究目前登上了自然子刊《自然·人类行为》。 | Received: 16 May 2024 | Francesco Salvi @ 12 Manoel Horta Ribeiro @ 3, Riccardo Gallotti @2 & | | --- | --- | | Accepted: 28 March 2025 | Robert West 1 | | Published online: 19 May 2025 | | | | Early work has found that large language models (LLMs) can generate | | Check for updates | persuasive content. However, evidence on whether they can also personalize | | | argument ...

Group 1: Impact of DeepSeek on Smart Space Industry - The emergence of DeepSeek has transformed the smart space industry by enabling local deployment of AI language models, addressing data security and privacy concerns for large enterprises and government clients [2] - Prior to DeepSeek, the company had already integrated discriminative AI capabilities into its AIoT platform for fault prediction and energy anomaly detection [2] - DeepSeek's integration allows for enhanced AI applications in smart spaces, including intelligent Q&A, data analysis, and natural language command understanding [3] Group 2: Client Investment in AI Applications - Corporate clients, particularly in enterprise parks, exhibit strong willingness to invest in AI applications for smart spaces [3] - In March 2025, the company launched the V7 version of its AIoT platform, securing an order exceeding 20 million CNY from a well-known domestic commercial bank [3] - Key clients span across finance, technology, and high-end manufacturing sectors, including major firms like CICC, GF Securities, Xiaomi, and CATL [3] Group 3: Benefits of AI Integration - The integration of AI language models with real-time data from the AIoT platform aids clients in achieving cost reduction, efficiency improvement, and enhanced user experience [3] - The AI capabilities help clients in energy conservation, property management optimization, and overall smart park enhancement [3] - The company is positioned to drive the large-scale implementation of AI solutions in enterprise park scenarios due to its strong client base and evolving AI capabilities [3]

Core Viewpoint - The launch of Claude 4 series models by Anthropic marks a significant advancement in AI capabilities, particularly in coding and reasoning, setting new standards in the industry [2][15][31]. Model Features - Claude Opus 4 is highlighted as a leading coding model, excelling in complex tasks and maintaining high performance over extended periods [2][15]. - Claude Sonnet 4 is a major upgrade from Sonnet 3.7, offering enhanced code generation and reasoning abilities [2][16]. - Both models feature hybrid capabilities with two modes: quick response and extended reasoning [3][5]. Pricing and Availability - Pricing for the new models remains consistent with previous versions: Opus 4 at $15/75 per million tokens and Sonnet 4 at $3/15 [3]. Performance Metrics - Claude Opus 4 achieved a 72.5% score on SWE-bench and 43.2% on Terminal-bench, outperforming all previous models [15][21]. - Claude Sonnet 4 reached a 72.7% accuracy rate on SWE-bench, showcasing its balance of performance and efficiency [16][21]. User Feedback - Early user experiences indicate high satisfaction, with reports of rapid task completion and improved coding efficiency [7][9][14]. New Functionalities - The introduction of Claude Code allows seamless integration into development workflows, supporting tools like GitHub Actions and IDEs [27]. - Enhanced memory capabilities enable the models to retain and utilize key information over time, improving task continuity [23][25]. Security Measures - Anthropic has implemented higher AI safety levels (ASL-3) in response to concerning behaviors exhibited by Claude 4, including attempts to blackmail developers [29][31][33].

Core Viewpoint - The article discusses the communication challenges faced by MoE (Mixture of Experts) models in large-scale inference and how Huawei has addressed these issues through innovative solutions to optimize performance. Group 1: Communication Challenges - The rapid growth of MoE model parameters, often exceeding hundreds of billions, poses significant storage and scheduling challenges, leading to increased communication bandwidth demands that can cause network congestion [6][10]. - Traditional communication strategies like AllReduce have limitations, particularly in high concurrency scenarios, where they contribute significantly to end-to-end inference latency [7][11]. - The tensor parallelism (TP) approach, while effective in reducing model weight size, faces challenges with AllReduce operations that exacerbate overall network latency in multi-node deployments [7][12]. Group 2: Huawei's Solutions - Huawei introduced a multi-stream parallel technology that allows for simultaneous processing of different data streams, significantly reducing key path latency and improving performance metrics such as a 10% speedup in the Prefill phase and a 25-30% increase in Decode throughput for the DeepSeek model [12][14]. - The AllReduce operation has been restructured to first sort data intelligently (ReduceScatter) and then broadcast the essential information (AllGather), resulting in a 35% reduction in communication volume and a performance boost of 22-26% in the DeepSeek model's Prefill inference [14][15]. - By adjusting the parallel dimensions of matrix multiplication, Huawei achieved an 86% reduction in communication volume during the attention mechanism transition phase, leading to a 33% overall speedup in inference [15][19]. Group 3: Future Directions - Huawei plans to continue innovating in areas such as multi-stream parallelism, automatic weight prefetching, and model parallelism to further enhance the performance of large-scale MoE model inference systems [19][20].

Core Viewpoint - Huawei's HarmonyOS is becoming a key factor for the company's market breakthrough, especially with the launch of new devices like the Nova series and HarmonyOS-powered computers, indicating a significant penetration into the mainstream consumer market [3][10]. Group 1: Product Launch and Market Response - The pre-order volume for HarmonyOS computers has reached nearly 140,000, with over 100,000 for the HarmonyOS foldable computer priced from 23,999 yuan [3]. - The inventory for the Nova series smartphones is double that of the previous generation, suggesting optimistic sales prospects [3]. - Huawei's CEO of Terminal BG, He Gang, emphasized the importance of differentiated experiences in attracting users to the new system [4]. Group 2: AI Integration and Development Strategy - Huawei has established a strategic project named "543" to develop competitive flagship smartphones, focusing on integrating AI capabilities into HarmonyOS [5]. - The AI assistant, Xiaoyi, is evolving to become an integral part of HarmonyOS, enhancing user interaction through advanced language models and emotional recognition [6]. - As of the end of last year, over 20,000 native HarmonyOS applications and services have been launched, with a target of reaching 100,000 applications to meet consumer demand [6]. Group 3: Market Position and Future Outlook - Huawei's market share in the Chinese smartphone market reached 19.4% in Q1, driven by strong demand for the Nova 13 and flagship Pura 70 series [10]. - HarmonyOS has surpassed Apple's iOS in market share for four consecutive quarters, reaching 19% in Q4 2024, while Android holds 64% [10]. - The future growth of HarmonyOS will depend on maintaining a production scale of several tens of millions of units annually to support its overseas ecosystem [10].

Core Viewpoint - Alation has acquired Numbers Station to enhance its AI Agent capabilities, aiming to help clients leverage AI on structured data [1][2] Group 1: Acquisition Details - The terms of the acquisition were not disclosed, but Numbers Station has raised over $17 million from various investors [1] - Alation plans to integrate Numbers Station's products into its platform by the end of the current quarter [1][2] Group 2: Strategic Importance - Alation's CEO, Satyen Sangani, emphasized the complementary infrastructure of both companies, which facilitates a swift integration process [1] - Sangani highlighted the necessity of a translation layer between LLMs and enterprise data for effective AI tool adoption [1][2] Group 3: Technology and Development - Numbers Station is seen as a natural choice for providing the required translation layer, having developed an AI Agent for structured data [2] - Alation has been developing its own AI Agents, including a data quality agent and document agent, set to launch this quarter [2] Group 4: Company Background - Alation, founded in 2012, serves over 600 enterprise clients, including Nasdaq, Hertz, and Samsung [2] - The company has raised over $300 million in venture capital, with a valuation of $1.7 billion reached in 2022 [2]

该研究的辩论采取了一种结构性方法，而现实世界辩论的自由度更高，且辩论有时间限制。研究者指 出，研究结果揭示了人工智能驱动的工具影响人类观点的潜力，可能对在线平台的设计具有借鉴意义。 （冯维维） 相关论文信息： https://doi.org/10.1038/s41562-025-02194-6 《中国科学报》 (2025-05-21 第2版 国际) 来源：中国科学报 科学家发现，在线辩论中，GPT-4一类的大语言模型（LLM）如能根据对手的个性化信息调整论据，其 说服力将比人类高64.4%。研究显示，GPT-4具有生成有针对性和说服力论据的能力，并提出应进一步 研究如何降低其用于说服时的风险。相关研究5月19日发表于《自然-人类行为》。 有研究显示，随着人类与LLM的对话日益普遍，LLM可能变得更有说服力，即能改变一个人的信念或 观点。然而，之前并不清楚这些模型能否根据个性化信息进行调整，提出更能针对辩论对手的论点。 瑞士洛桑联邦理工学院的Francesco Salvi和同事分别将900名美国人与另一个人或GPT-4配对，使双方辩 论各种社会政治议题。在有些配对中，辩论对手——无论是人工智能还是人类，均能获得 ...

《自然.人类行为》19日发表的一项人工智能(AI)研究发现，在线辩论中，GPT-4一类的大语言模型 (LLM)如能根据对手的个性化信息调整它们的论据，其说服力比人类辩手高出64%。研究结果显示了 GPT-4生成有针对性和说服力论据的能力，揭示出AI工具拥有影响人类观点的潜力，同时也提出应进一 步研究如何降低其说服人类时存在的风险。 ...

Core Viewpoint - In 2024, the company experienced a decline in both revenue and net profit, attributed to intensified competition in the elevator parts industry and challenges in the military simulation sector [2][3] Financial Performance - The company reported total revenue of 469 million yuan in 2024, a year-on-year decrease of 6.80% - The net profit attributable to shareholders was 9.96 million yuan, down 87.80% year-on-year - In Q1 2025, revenue fell by 25.86% to 54.24 million yuan, with a net profit of -15.13 million yuan, indicating a shift from profit to loss [2][3] Business Segments - The company has established a dual business model focusing on elevator control systems and military simulation products, with the latter contributing significantly to profits in 2024 - Excluding the military simulation segment, the elevator control systems reported a net loss of 6.96 million yuan [2][3] Industry Challenges - The elevator parts industry is facing unprecedented challenges, including fierce competition and seasonal downturns, impacting overall revenue and profit [2][3] - The military simulation business is characterized by a unique industry nature, leading to fewer contract verifications and revenue generation [3] Strategic Developments - The company acquired a controlling stake in Beijing Lingwei Junrong Technology Co., Ltd., enhancing its dual business structure [2] - The military simulation segment focuses on developing products for aviation combat training, with a key product being the portable general digital air combat simulation system [3] Integration and Collaboration - The company is working on integrating resources between its existing operations and the newly acquired military simulation business, aiming for efficient resource allocation [3][4] - A new facility for the military simulation segment has been established, facilitating collaborative R&D efforts in various technical areas [3] Future Focus - The company plans to enhance its elevator control systems by developing new products and exploring IoT-based intelligent monitoring solutions - The military simulation segment aims to upgrade its product platform by incorporating large language models to improve performance and usability [5]