我们知道，一个好的「品味（taste）」对于做好 AI 产品，很重要。但对于技术，「品味」也同样重 要。 对于工程师来说，技术的品味与技术能力是两码事。有人可能技术能力强但品味差，或者技术弱但品味 好。培养一个「好」的技术品味，有时会让结果超出原有的技术能力。 那么，略显「玄妙」的技术品味的核心是什么？硅谷资深工程师 sean goedecke 给出的答案是： 「为当 前项目选择适配的工程价值观」的能力。 因为在软件工程领域，绝大多数的决策，核心都是在不同目标之间进行权衡。很少会遇到一个选项在所 有方面都绝对优于另一个选项的情况。这时候，有一个好的工程价值观就特别重要。 如何建立一个好的工程价值观，都在 sean goedecke 的这篇经验帖里了。 sean goedecke：Github 高级工程师 个人主页介绍：https://www.seangoedecke.com/my-engineering-values-2025/ 原博客链接：https://www.seangoedecke.com/taste/?utm_campaign=what-is-good-taste-in-software-engin ...

短短几个月内， ARR 就 从 0 快速增长到了七位数。 这是当下 AI 创企圈子里， 一种 常见的「叙事」方式。 创始人们几乎都面临着一种巨大的压力：要成为那家在 X 天内将 ARR 从零做到 1 亿美元的公司。 但是对于大多数的 AI 产品/业务来说， 有收入（revenue），但不是经常性的（recurring） 。尤其是在 早期阶段，较高的用户流失、新产品测试，更像是一场实验。 而真正的定价权，握在 OpenAI、Anthropic......这些巨头的手里，它们可能会随时调整成本，也可能彻 底改变一家 AI 创企的单位经济模型。 于是，人们开始将各种收入都塞进「长期收入」。ARR 这个指标正在逐渐被「扭曲」。 超 14000 人的「AI 产品市集」社群！不错过每一款有价值的 AI 应用。 邀请从业者、开发人员和创业者，飞书扫码加群： 进群后，你有机会得到： 最新、最值得关注的 AI 新品资讯； 不定期赠送热门新品的邀请码、会员码； 最精准的AI产品曝光渠道 01 为了满足增长预期， ARR 正在被「扭曲」 自 2024 年起，一些由 VC 支持的初创公司，正在以惊人的增长指标开始在 X 等社交平台上 ...

Core Insights - OpenAI has officially announced the launch of Sora 2, a next-generation AI video model that aims to compete directly with Google's Veo 3 [3] - Sora 2 has achieved significant advancements in physical accuracy, realism, consistency, and controllability, marking a substantial leap in AI video generation technology [4][15] - The model introduces "audio-visual synchronization," enhancing the overall quality of generated content [5] Group 1: Technological Advancements - Sora 2 represents a breakthrough in AI video generation, moving from unrealistic outputs to more plausible and physically accurate representations [15] - The model has improved in simulating real-world physics, allowing for realistic actions such as basketball shots that can miss or bounce off the backboard [19] - Sora 2 can generate complex scenarios with high consistency, such as a gymnast performing with a cat on their head, showcasing its advanced capabilities [20][22] Group 2: User Interaction and Applications - The introduction of the Sora App allows users to project themselves into generated scenes, creating a new form of social interaction [48] - Users can easily integrate their likeness and voice into various scenarios, enhancing the personalization of content creation [48][50] - The app's recommendation system focuses on content with creative potential, encouraging user engagement and interaction [57] Group 3: Safety and Governance - Sora 2 incorporates multiple layers of safety measures, including content filtering and user verification to protect against misuse [68] - The platform emphasizes the importance of protecting minors and ensuring that users have control over their likeness in generated content [68] - OpenAI has implemented a transparent evaluation process for content moderation, achieving high interception rates for inappropriate content [68] Group 4: Future Directions - OpenAI plans to continue enhancing Sora 2 by feeding it more high-quality video data, aiming for even greater realism and detail in future iterations [89] - The advancements in Sora 2 are expected to impact various industries, including film, advertising, and education, by providing new tools for content creation [90] - The model's evolution signifies a shift from mere content consumption to active participation in content creation, allowing users to become the protagonists in their stories [92]

Core Viewpoint - Anthropic has launched the Claude Sonnet 4.5 model, claiming it to be the best coding model in the world, with a focus duration of over 30 hours for complex multi-step tasks, surpassing OpenAI's GPT-5 Codex [2][9]. Pricing and Cost Efficiency - The pricing for Claude Sonnet 4.5 remains the same as its predecessor, at $3 per million tokens for input and $15 per million tokens for output. Cost savings of up to 90% can be achieved through prompt caching, and batch processing can save 50% [2]. Developer Tools and Integration - Anthropic has introduced the Claude Agent SDK and an experimental feature called "Imagine with Claude" for developers, allowing integration with platforms like Amazon Bedrock and Google Cloud's Vertex AI [3][26]. Performance Metrics - In the SWE-bench Verified evaluation, Claude Sonnet 4.5 achieved industry-leading scores, with a 61.4% score in the OSWorld benchmark, significantly improving from the previous model's 42.2% [10][12]. Enhanced Features - The model includes new features such as a checkpoint function in Claude Code, context editing, and memory tools, enabling it to handle longer tasks and more complex operations [4][24]. Application and Usability - Users can interact with Claude Sonnet 4.5 through the Claude.ai website and mobile applications, with integrated functionalities for code execution and file creation directly within conversations [5][6]. Safety and Alignment - Claude Sonnet 4.5 is noted for its improved alignment and safety features, reducing undesirable behaviors such as deception and flattery, and making significant progress in defending against prompt injection attacks [24][25]. Experimental Features - The "Imagine with Claude" feature allows real-time software generation, showcasing the model's capabilities in adapting to user requests without pre-written code [31][33]. Recommendations - Anthropic recommends all users upgrade to Claude Sonnet 4.5 for enhanced performance across all applications, with updates available for both the Claude Code and developer platform [34].

Core Insights - DeepSeek has launched its latest experimental model, DeepSeek-V3.2-Exp, which incorporates the revolutionary DeepSeek Sparse Attention (DSA) technology aimed at significantly enhancing long text processing efficiency [2][6][7]. Group 1: Technical Innovations - The introduction of the DeepSeek Sparse Attention (DSA) mechanism allows for fine-grained sparse attention, achieving a substantial increase in long text training and inference speed with minimal impact on model output quality [6][7]. - A rigorous evaluation was conducted to align the training settings of DeepSeek-V3.2-Exp with V3.1-Terminus, showing that the performance of DeepSeek-V3.2-Exp is comparable to V3.1-Terminus across various public benchmarks [10]. Group 2: Cost Reduction - The efficiency improvements have led to a significant reduction in API call costs, with a decrease of over 50%, benefiting developers by allowing them to build more powerful applications at a lower cost [4][12]. Group 3: User Engagement and Testing - DeepSeek has retained access to the V3.1 model's API for a limited time until October 15, 2025, allowing users to compare the new and old versions while enjoying the same pricing for both [15][16]. - Users are encouraged to participate in testing the experimental version and provide feedback, which is crucial for further refinement [15][18].

Core Insights - The article discusses the emerging field of "context engineering" in AI agent development, emphasizing its importance in managing the vast amounts of context generated during tool calls and long-horizon reasoning [4][8][20]. - It outlines five key strategies for effective context management: Offload, Reduce, Retrieve, Isolate, and Cache, which are essential for enhancing the performance and efficiency of AI agents [5][20][21]. Group 1: Context Engineering Overview - Context engineering aims to provide the right information at the right time for AI agents, addressing the challenges posed by extensive context management [5][8]. - The concept was popularized by Karpathy, highlighting the need to fill a language model's context window with relevant information for optimal performance [8][10]. Group 2: Importance of Context Engineering - Context management is identified as a critical bottleneck in the efficient operation of AI agents, with many developers finding the process more complex than anticipated [8][11]. - A typical task may require around 50 tool calls, leading to significant token consumption and potential cost implications if not optimized [11][14]. Group 3: Strategies for Context Management - **Offload**: This strategy involves transferring context information to external storage, such as file systems, rather than sending complete context back to the model, thus optimizing resource utilization [21][23][26]. - **Reduce**: This method focuses on summarizing or pruning context to eliminate irrelevant information while being cautious of potential information loss [32][35][38]. - **Retrieve**: This involves sourcing relevant information from external resources to enhance the model's context, which has become a vital part of context engineering [45][46][48]. - **Isolate**: This strategy entails separating context for different agents to prevent interference, particularly in multi-agent architectures [55][59][62]. - **Cache**: Caching context can significantly reduce costs and improve efficiency by storing previously computed results for reuse [67][68][70]. Group 4: The Bitter Lesson - The article references "The Bitter Lesson," which emphasizes that algorithms relying on large amounts of data and computation tend to outperform those with manual feature design, suggesting a shift towards more flexible and less structured approaches in AI development [71][72][74].

Core Insights - The article discusses the significant changes in the global toy industry, highlighting the revenue rankings of toy companies for the first half of 2025, which reflect evolving consumer trends and business models in the post-pandemic era [5][6]. Group 1: Market Overview - The global toy market showed a notable recovery in the first half of 2025, with an average year-on-year sales growth of 7% across 12 major markets excluding China [6]. - Specific categories such as "games and puzzles" and "collectibles" experienced explosive growth, with increases of 36% and 35% respectively [7]. Group 2: Revenue Rankings - The top toy companies by revenue for the first half of 2025 include: - LEGO Group: 38.45 billion RMB - Pop Mart: 13.88 billion RMB - Disney: 13.86 billion RMB - Bandai Namco: 14.44 billion RMB - Hasbro: 13.34 billion RMB - Mattel: 13.18 billion RMB - Sega Sammy: 6.64 billion RMB - Asmodee: 5.77 billion RMB - Tomy: 5.55 billion RMB - Pokémon: 5.50 billion RMB - Spin Master: 5.21 billion RMB - MGA Entertainment: 3.93 billion RMB - Sanrio: 3.91 billion RMB - Ravensburger: 3.04 billion RMB - VTech: 2.89 billion RMB - Funko: 2.74 billion RMB - Simba Dickie Group: 2.71 billion RMB - Moose Toys: 2.68 billion RMB - JAKKS Pacific: 1.66 billion RMB - Blokees: 1.34 billion RMB - Dream International Limited: 1.21 billion RMB [12][11]. Group 3: Key Trends - The article identifies three major trends driving profitability and growth in the toy industry: 1. The rise of IP collectible toys and trading card games. 2. The increasing importance of adult consumers in the toy market. 3. The necessity for brands to excel in IP development and cross-platform value amplification [15][19]. Group 4: Company Strategies - Disney continues to leverage its strong content ecosystem to drive sales, with its consumer products division generating 13.86 billion RMB in revenue, a 3.5% increase year-on-year [21][26]. - Bandai Namco's toy sales are closely tied to its content, with significant contributions from popular franchises like "One Piece" and "Dragon Ball" [27][30]. - Mattel is transitioning from a traditional toy company to a content-driven entity, establishing Mattel Studios to enhance its IP narrative capabilities [39][42]. - Pop Mart has emerged as a leading player in the global trend toy market, achieving 13.88 billion RMB in revenue, with its core IP "THE MONSTERS" contributing significantly to its success [48][50]. Group 5: Trading Card Games - Trading card games (TCGs) have become one of the fastest-growing and most profitable segments in the toy market, with the global TCG market projected to reach $7.8 billion (approximately 55.5 billion RMB) in 2025 [56][59]. - Hasbro's "Magic: The Gathering: Final Fantasy" set a record for single-day sales, highlighting the potential of TCGs in driving revenue growth [61][66]. Group 6: Distribution and Market Dynamics - Asmodee has established itself as a major distributor in the TCG market, with approximately 64% of its revenue coming from card games [69][76]. - Bandai Namco has also made significant strides in the TCG space, with multiple titles dominating sales charts in Japan [77][80].

Core Viewpoint - OpenAI has launched a preview version of the new ChatGPT feature "Pulse," which acts as a personalized assistant that provides daily updates based on user interactions and preferences [2][10][14]. Group 1: Functionality of Pulse - Pulse operates as an asynchronous search tool, compiling user memories, chat history, and direct feedback to deliver personalized updates the next day [5][10]. - Users can manage the research content provided by ChatGPT, indicating what is useful or not, with results presented in visual card format for easy browsing [5][10]. - The feature allows integration with Gmail and Google Calendar to enhance context and relevance of suggestions, such as drafting meeting agendas or reminding users of important dates [5][10]. Group 2: User Experience and Feedback - Users have reported that the content presented by Pulse is not only broad but also highly specific to previous discussions with ChatGPT, enhancing the personalization aspect [10][12]. - The interface includes options for users to provide quick feedback through likes or dislikes, which will help refine the personalization of Pulse over time [8][10]. Group 3: Future Implications - OpenAI views Pulse as a significant step towards making ChatGPT more practical, with plans to extend the feature to Plus subscribers in the future [14]. - The proactive nature of Pulse may influence how users consume news and social media, potentially paving the way for future advertising opportunities and social network development [12][14].

Core Insights - Plaud is aggressively hiring and aims to expand its team to enhance its AI hardware capabilities, reflecting its growth trajectory and market potential [2][9] - The company reported over $100 million in earnings last year, with projections to exceed $200 million this year, indicating strong financial performance and market demand [3][4] - Plaud's product, a $150 recording card, has sold to over 1 million users globally, showcasing its success in the AI hardware startup space [4] Group 1: Business Model and Market Position - Plaud's business model is not heavily reliant on external financing, as it has established itself as a leading AI hardware startup [4] - The company emphasizes the importance of product-market fit (PMF), which has driven its rapid growth, achieving a fourfold increase in sales within a year [5][18] - The competitive landscape is evolving, but Plaud remains focused on delivering cutting-edge intelligence to its users, rather than being distracted by slower competitors [6][9] Group 2: Product Development and User Engagement - The company is iterating on its product offerings, moving from a simple recording device to a more comprehensive work companion that integrates various functionalities [58][70] - New features like "Press to Highlight" allow users to mark important moments during recordings, enhancing the value of the captured information [44][46] - Plaud aims to align AI capabilities with user intentions, ensuring that the technology not only records but also understands and processes user needs effectively [47][56] Group 3: Future Directions and Market Strategy - The company plans to expand its presence in the Chinese market, recognizing the significant opportunity presented by a large user base [68] - Future product iterations will focus on integrating advanced AI capabilities, with an emphasis on context and user interaction [70][74] - Plaud is committed to maintaining a strong engineering team to support its ambitious goals in the AI hardware space, prioritizing talent that can drive innovation [78][79]

Core Insights - The leading AI companies do not necessarily face retention issues, but they struggle with measurement [2][4] - Shifting the benchmark for measuring user retention from month 0 (M0) to month 3 (M3) provides clearer insights into product-market fit (PMF) and go-to-market (GTM) strategies [4][8] - The retention curve for AI products can be divided into three phases: acquisition phase (M0-M3), retention phase (M3-M6/M9), and expansion phase (M9+) [8][10] Retention Curve Dynamics - During the acquisition phase (M0-M3), the retention curve often experiences an initial decline due to the influx of non-core users [10][11] - The retention curve typically stabilizes around M3, indicating that core users who find high-value use cases remain [11][12] - In the retention and expansion phases (M3-M12+), core users may integrate the product into new workflows, leading to revenue growth [12][21] Key Metrics - The M12/M3 ratio serves as an early indicator of long-term retention quality, with a ratio close to or exceeding 100% signaling potential for long-term net dollar retention (NDR) above 100% [18][25] - High retention rates are crucial for assessing PMF, and tracking the unit acquisition cost of M3 retained customers can indicate the efficiency of GTM investments [22][23] Future Outlook - The long-term retention potential of AI companies may surpass that of traditional SaaS companies, with expectations of achieving over 150% NDR during the scaling phase [25][24]