「DeepSeek 一更新，我们就知道又要放假了。」 昨天，DeepSeek 宣布其 R1 系列推理模型小版本升级，最新版本 DeepSeek-R1-0528 参数量高达 6850 亿，模型在思维深度和推理方面的能力显著提升。 刚刚，DeepSeek 公布了 R1-0528 在各类基准测评上的具体得分情况。R1-0528 在数学、编程与通用逻辑等多个基准测评中成绩亮眼，整体表现接近 o3 与 Gemini-2.5-Pro。 | Benchmarks | DeepSeek-R1- | OpenAI- | Gemini-2.5- | Qwen3- | DeepSeek-R1 | | --- | --- | --- | --- | --- | --- | | | 0528 | o3 | Pro-0506 | 235B | | | AIME 2024 数学竞赛 pass@1 | 91.4 | 91.6 | 90.8 | 85.7 | 79.8 | | AIME 2025 数学竞赛 pass@1 | 87.5 | 88.9 | 83.0 | 81.5 | 70.0 | | GPQA Diamond 科学测试 pass@ ...

Core Insights - The article discusses the upcoming Founder Park event, which aims to connect AI entrepreneurs, developers, and investors in a collaborative environment [1][2][3]. Event Overview - Founder Park will feature 22 AI startup communities and will serve as a platform for networking and discussions among participants [1]. - The event is scheduled for June 21-22, 2025, at various venues within the 751 Park area [5][22]. Agenda Highlights - The agenda includes thematic discussions on AI hardware, global expansion strategies, and innovative entrepreneurial paradigms [3][6]. - Notable sessions include "How to Deliver Unprecedented User Value in the AI Era" and "Reconstructing the Paradigm of Overseas Entrepreneurship" [6][7]. Keynote Speakers - The event will host prominent figures such as Zhang Peng, founder of Geek Park, and other industry leaders who will share insights on AI trends and investment opportunities [6][14]. - Discussions will also cover the future of embodied intelligence and the impact of AI on revenue models [7][15]. Networking Opportunities - The event is designed to facilitate spontaneous conversations and connections among attendees, emphasizing the importance of informal networking in the tech community [2][24]. - Participants will have the chance to engage with various startups and innovation partners, enhancing collaboration within the AI ecosystem [24][39]. Investment Trends - The article hints at a new wave of global investment paradigms driven by advancements in AI technologies, with a focus on the 2025 AI Cloud industry trends report [14][19]. - The event will also feature discussions on how AI can enhance SaaS offerings and global case studies [19][22].

Core Insights - The main change expected in 2025 is the effective application of reinforcement learning (RL) in language models, particularly through verifiable rewards, leading to expert-level performance in competitive programming and mathematics [4][6][7]. Group 1: Reinforcement Learning and Model Development - Reinforcement learning has activated existing knowledge in models, allowing them to organize solutions rather than learning from scratch [4][11]. - The introduction of Opus 4 has significantly improved context management for multi-step actions and long-term tasks, enabling models to perform meaningful reasoning and execution over extended periods without frequent user intervention [4][32]. - The current industry trend prioritizes computational power over data and human feedback, which may evolve as models become more capable of learning in real-world environments [4][21]. Group 2: Future of AI Agents - The potential for AI agents to automate intellectual tasks could lead to significant changes in the global economy and labor market, with predictions of "plug-and-play" white-collar AI employees emerging within the next two years [7][9]. - The interaction frequency between users and models is expected to shift from seconds and minutes to hours, allowing users to manage multiple models simultaneously, akin to a "fleet management" approach [34][36]. - The development of AI agents capable of completing tasks independently is anticipated to accelerate, with models expected to handle several hours of work autonomously by the end of the year [36][37]. Group 3: Model Capabilities and Limitations - Current models still lack self-awareness in the philosophical sense, although they exhibit a form of meta-cognition by expressing uncertainty about their answers [39][40]. - The models can simulate self-awareness but do not possess a continuous identity or memory unless explicitly designed with external memory systems [41][42]. - The understanding of model behavior and decision-making processes is still evolving, with ongoing research into mechanisms of interpretability and the identification of features that drive model outputs [46][48]. Group 4: Future Developments and Expectations - The frequency of model releases is expected to increase significantly, with advancements in reinforcement learning leading to rapid improvements in model capabilities [36][38]. - The exploration of long-term learning mechanisms and the ability for models to evolve through practical experience is a key area of focus for future research [30][29]. - The ultimate goal of model interpretability is to establish a clear understanding of how models make decisions, which is crucial for ensuring their reliability and safety in various applications [46][47].

Core Viewpoint - The article discusses the transformation of Google's search business towards AI-driven search modes, highlighting the challenges faced by traditional search engines in the face of emerging AI technologies and competition from Chatbot-integrated platforms [4][24]. Group 1: Google's AI Search Transformation - Google announced the launch of its AI Mode powered by Gemini, which allows for natural language interaction and structured answers, moving away from traditional keyword-based searches [2][4]. - In 2024, Google's search business is projected to generate $175 billion, accounting for over half of its total revenue, indicating the significant financial stakes involved in this transition [4]. - Research suggests that Google's search market share has dropped from over 90% to between 65% and 70% due to the rise of AI Chatbots, prompting the need for a strategic shift [4][24]. Group 2: Challenges for AI Search Engines - Perplexity, an AI search engine, saw its user visits increase from 45 million to 129 million, a growth of 186%, but faced a net loss of $68 million in 2024 due to high operational costs and reliance on discounts for subscription revenue [9][11]. - The overall funding for AI search products has decreased, with only 10 products raising a total of $893 million from August 2024 to April 2025, compared to 15 products raising $1.28 billion in the previous period [11][12]. - The competitive landscape for AI search engines has worsened, with many smaller players struggling to secure funding and differentiate themselves from larger companies [11][12][25]. Group 3: Shift Towards Niche Search Engines - The article notes a trend towards more specialized search engines, focusing on specific industries or use cases, as general AI search engines face increasing competition from integrated Chatbot functionalities [13][25]. - Examples of niche search engines include Consensus, a health and medical search engine, and Qura, a legal search engine, both of which cater to specific professional audiences [27][30]. - The overall direction for AI search engines is towards being smaller, more specialized, and focused on delivering unique value propositions to specific user groups [13][26]. Group 4: Commercialization Challenges - The commercialization of AI search remains a significant challenge, with Google exploring ways to integrate sponsored content into its AI responses while facing potential declines in click-through rates for traditional ads [43]. - The article emphasizes the need for AI search engines to deliver more reliable and usable results, either through specialized information or direct output capabilities, to remain competitive [43][24].

Core Viewpoint - The Browser Company is transitioning from its Arc browser to a new AI-native product called Dia, driven by the belief that traditional browsers will become obsolete as user interaction evolves towards AI interfaces [4][35]. Group 1: Arc Browser Launch and Initial Success - Arc browser was launched in 2023, introducing innovative features such as a customizable sidebar, smart tab management, and quick webpage previews, attracting over a million engaged users [2][3]. - Following the rise of ChatGPT, Arc quickly integrated AI capabilities with the launch of Arc Max, allowing users to interact with AI for webpage explanations [2][3]. Group 2: Transition to Dia - In October 2024, the company announced that Arc would enter maintenance mode as they focus on developing Dia, a new AI-native browser aimed at a broader audience [4][6]. - The decision to pivot was met with skepticism from existing users, who feared abandonment of the Arc product [5][6]. Group 3: Lessons Learned from Arc - The company identified three major mistakes made with Arc: delaying the decision to stop investment in Arc, not fully embracing AI sooner, and failing to communicate effectively with users [14][16][30]. - Arc's complexity led to a "novelty tax," where users faced high learning costs without proportional benefits, resulting in low engagement with many features [23][24]. Group 4: Future Vision and Product Strategy - The company believes that the future of desktop interaction will not solely rely on traditional web browsers but will integrate AI capabilities, creating a hybrid interface [35][36]. - Dia aims to prioritize simplicity and speed, addressing the shortcomings of Arc by ensuring a user-friendly experience while maintaining robust performance and security [30][31]. Group 5: Market Positioning and Expectations - The Browser Company envisions Dia as a potential successor to traditional browsers, with the belief that the most used AI interface in five years will replace the current default browsers [40][41]. - The company acknowledges the risks involved in this transition but remains committed to its vision of redefining how users interact with the internet [40][41].

AI 开源领域的核心玩家 Meta 近期面临的争议不断。 超 4000 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： 在继 Llama 4 模型被爆出实际性能与宣传不符，在测试集上进行训练的「丑闻」之后，Meta 近期又被爆出其 AI 团队的 近八成员工已离职。 据 businessinsider 报道，Meta 的 AI 团队正面临严峻的人才流失挑战，Llama 模型创始团队的 14 名核心成员仅剩 3 名在 职。而在已经离职的 11 名核心研究人员中，5 名跳槽去了法国 AI 开源模型创企 Mistral。 进群后，你有机会得到： 同时，Meta 还面临着 DeepSeek、Qwen 等开源模型追赶迅速的竞争压力。Meta 在 AI 领域投入了数十亿美元，但至今仍 没有推出专有的「推理」模型，人们逐渐转向使用提供更先进功能的模型，Meta 在开源领域与竞争对手的差距也变得更 为明显。 01 Meta AI 人才流失严重， Llama 14 名核心作者仅剩 3 名在职 Meta 的 AI 团队正面临严峻的人才流失挑战，其核心 Llama 模型创 ...

文章转载自红杉中国公众号「红杉汇」，内容略有调整。 红杉中国开放了他们内部进行 AI 和 Agent 基准测试的工具「 xbench」，并发布了相应论文《xbench: Tracking Agents Productivity，Scaling with Profession-Aligned Real-World Evaluations》。 论文地址： https://xbench.org/files/xbench_profession_v2.4.pdf TLDR： | Benchmark | Category | 151 B | 8 2nd | g 3rd | Details | | --- | --- | --- | --- | --- | --- | | xbench-ScienceQA | AGI Tracking | 03- high 60.8 | Gemini 2.5 Pro 57.2 | Doubao-1.5-thinking- pro 53.6 | View > | | xbench-DeepSearch | AGI Tracking | 03 65+ | o4-mini-high 60+ | ...

Character AI 是个不合时宜的故事，不再多聊。 但 Character AI 的火爆，让资本市场关注到一个新的人群：OC。Original Character，用户在虚拟世界构建的有自己生活与故事的独立个体。 2018 年诞生的 OC 游戏 Gacha Life，拥有超过 2 亿玩家。在应用商店的排名里，它常常与《原神》等游戏并列。Gacha Life 完成了 OC 市场的 PMF 验证。 在国内，B 站、小红书、抖音随处可见 Gacha Life 玩家的分享内容。 这是一种越来越普遍的创作需求，它需要新的生产工具和分发平台。 与 C.AI 不同，OC 拥有一个明确的快速增长的核心人群，他们年轻，黏性极强，有着拦不住的分享欲，愿意为自己创作的角色付出。对他们来说，AI 能 够降低创作门槛，也能让自己的角色拥有更丰富的体验——它可以拥有自己的世界，与其他角色互动，或者以桌宠的形态为用户提供充满惊喜的情绪价 值。 但从一个特点的人群出发，需要你非常理解这些用户。 产品 & 公司介绍： Kotoko AI 成立于 2023 年，其产品 Bside 是一款融合了 UGC 与游戏化玩法的 OC 社交互动平台， ...

Founder Show 是 AGI Playground 2025 大会中的创业者特别分享环节。 每位创业者将通过 20 分钟时间，全方位分享产品进展、创业思考，与场上的「高年级创业者」实时互动交流。 我们将通过「资料初筛-线上预沟通-项目复审-入选通知」等环节，选出 9 支新锐团队登上 Founder Show 的舞台。 通过初筛及全部线下展示团队都将获得由 Founder Park & 变量资本提供的创业加速资源包。 谁来参与 招募要求 时间地点 活动形式 时间：2025 年 6 月 20 日下午 地点：北京｜751 图书馆 时间线 招募流程 资料提交-资料初审-线上面试-项目复审-入选通知 9 位新锐 Founder，独立开发/拥有团队均可 AGI Founders Fund 特邀 LP、Founder Park 的「高年级创业者」 泛 Gen AI 赛道，垂类场景和产品形态不限，有可展示的产品 Demo 更佳 如最终入选，可配合大会流程，进行约 20 分钟的产品展示及线下互动 报名时间：5 月 23 日-6 月 10 日 18:00 最终通知：6 月 13 日 18:00（过期未通知即为未入选） 资 ...

Core Insights - Sam Altman and Jony Ive are collaborating to create a new hardware device, which aims to be the third core device on users' desks after the MacBook Pro and iPhone [1][4][5] - OpenAI has announced the acquisition of Jony Ive's AI hardware startup "io" for nearly $6.5 billion, with plans to ship 100 million units of the new device [1][4][8] - The device is designed to reduce users' reliance on screens and is not intended to be a smartphone or wearable technology [1][5][10] Summary by Sections Acquisition and Collaboration - OpenAI's acquisition of "io" is seen as a significant opportunity, with Altman suggesting it could generate up to $1 trillion in additional value for the company [4][9] - The collaboration between Altman and Ive has evolved over the past 18 months, with a focus on developing a device that serves as a core interaction point between users and OpenAI [10] Device Concept and Design - The new device will be pocket-sized and designed for easy placement on desks, emphasizing a low-profile design [5][10] - Altman and Ive believe that existing devices do not meet user needs, and the new product aims to change how users interact with AI [10] Market Context and Competition - The announcement comes amid other tech giants like Google and Apple launching their own AI hardware products, including smart glasses [2][9] - Altman acknowledges the challenges of entering the hardware market, especially against established companies like Apple and Google [8][9]