点击下方 卡片 ，关注" 大模型之心Tech "公众号 戳我 -> 领取大模型巨卷干货 现在， Meta AI Research 携最新力作 DINOv3 震撼登场。 >> 点击进入→ 大模型技术 交流群 本文只做学术分享，如有侵权，联系删文 写在前面 还记得当年AlexNet在ImageNet上的一鸣惊人吗？ 它点燃了深度学习的浪潮，但背后是海量 人工标注 的心血——千万张图片，被逐一打上标签。自此，"数据饥渴"和"标注成本"如同两座大山，压在 计算机视觉发展的道路上。 研究者们一直在追寻一个梦想： 能否让模型像人类婴儿一样，仅通过"观察"世界就能学习强大的视觉理解能力，彻底摆脱对人工标注的依赖？ 这 就是 自监督学习（SSL） 的终极目标。 这条路上星光熠熠： 然而，挑战依然存在： MAE(Masked Autoencoders) ：如同BERT之于文本，让模型通过"猜"被遮盖的图像块来学习，展现了强大的潜力。 MoCo/SimCLR ：通过对比不同视角下的同一图像，让模型理解"什么看起来应该相似"。 DINO系列 (特别是DINOv2) ：带来了真正的突破！它不仅能学到优秀的 全局 图像特征（用于分类、 ...

大模型开始走入"中场战事"，商业化落地成为今年大模型的关注焦点之一。各家对落地终端的押注主要 集中在汽车、手机、机器人等领域。 一直以端侧模型著称的面壁智能，近期将汽车这一终端的重要性提上了一个级别。南都N视频记者从面 壁智能方面获悉，8月15日面壁智能CEO李大海在公司成立三周年之际，发布全员信称，7 月下旬面壁 已进行了新一轮组织升级，专门成立一级组织汽车业务线，旨在实现压强式突破，让MiniCPM端侧模 型的智能奔腾到更多汽车上。 面壁智能于2022年8月成立，由清华大学自然语言处理实验室技术成果转化而来。2024年初，面壁智能 定义并开拓端侧智能市场，推出面壁小钢炮MiniCPM系列端侧模型。从2.4B参数能力超越Mistral 7B模 型打响口碑开始，推出多模态端侧代表作V2.5、o2.6 等一系列有世界级影响力的模型，陆续形成基座、 多模态、全模态的MiniCPM 端侧模型完整谱系。 今年6月，面壁开源两款最"快"的 MiniCPM 4.0模型，8月接力开源MiniCPM-V4.0，让多模态能力流畅 运行在手机上。据内部信透露，在面壁技术路线图上，今年下半年将有一批端侧模型陆续发布。 目前，端侧模 ...

Group 1 - The first application of specialized AI models in China's space station [1] - The Tianyi 2.0 humanoid robot from Beijing Humanoid Robot Innovation Center won both first and second place in the humanoid robot sports competition [1] - The "White Rhino" robot set a Guinness World Record with a 100-meter run time of 16.33 seconds [1] Group 2 - Meta is planning a comprehensive restructuring of its artificial intelligence work team [1]

Core Viewpoint - Palantir Technologies has experienced significant stock growth, rising over 140% in 2025, leading to questions about the potential for future investment returns [1][2]. Company Performance - Palantir's commercial revenue grew 47% year over year to $451 million, while government sales increased 49% to $553 million in Q2, contributing to the company's first quarter with $1 billion in revenue [5][6]. - The company started as a government-focused entity and has successfully expanded into the commercial sector [5]. Market Valuation - Palantir's current market cap is $430 billion, and to achieve a $1 million return from a $10,000 investment, it would need to reach a $43 trillion valuation, which is deemed unrealistic given that Nvidia, the largest company, has a market cap of $4.4 trillion [4]. - The stock is considered expensive, trading at 281 times forward earnings and 134 times sales, indicating that significant growth is already factored into the stock price [7][9]. Future Projections - If Palantir's revenue and profits grow from $3.44 billion and $763 million to $26.1 billion and $7.83 billion over five years, a reasonable valuation would be around 50 times forward earnings, resulting in a market cap of $392 billion, which is lower than its current valuation [9]. - The analysis suggests that the stock price reflects more than five years of anticipated growth, making it a costly investment at present [10].

Core Insights - The report highlights the emergence of two types of high-growth AI startups: Supernovas and Shooting Stars, with distinct growth trajectories and sustainability profiles [2][11][12] - AI is transitioning from traditional record-keeping systems to action-oriented systems, leveraging memory and context as new competitive advantages [3][24] - Vertical AI is rapidly penetrating traditionally "technology-averse" industries such as healthcare, law, and education, demonstrating clear ROI and immediate value [4][27][30] - Generative video is expected to see explosive growth by 2026, potentially leading to the emergence of new social media giants driven by AI capabilities [5][40] - The need for enterprise-level AI assessment and data traceability is becoming critical, with a shift towards privatized and scenario-based evaluations [6][44] Group 1 - AI has entered the "First Light" phase, with Supernovas achieving $40 million ARR in their first year and $125 million in the second year, but with low gross margins averaging around 25% [9][10] - Shooting Stars, in contrast, grow from $3 million ARR to $100 million in four years, maintaining healthier gross margins around 60% [11][12] - The new growth benchmark for AI companies is Q2T3 (4x, 4x, 3x, 3x, 3x), replacing the previous SaaS standard of T2D3 [2][14] Group 2 - AI is disrupting traditional enterprise systems like CRM and ERP, reducing implementation cycles from months to hours through automated data collection and execution [3][24] - The emergence of Agentic AI may lead to a browser-based interface that transitions from passive navigation to active task execution [4][39] - AI-native tools in CRM and HR are not just replacing existing systems but are creating entirely new experiences that significantly enhance productivity [25][29] Group 3 - The report identifies a significant acceleration in vertical AI adoption across sectors like healthcare, law, and education, with companies like Abridge and EvenUp leading the charge [27][30] - AI tools are automating complex workflows, improving efficiency, and addressing previously unmet needs in these industries [28][31] - The potential for AI to transform enterprise software is evident, with startups challenging established record-keeping systems by creating action-oriented solutions [24][25] Group 4 - The report predicts that 2026 will be a pivotal year for generative video, with advancements in technology making video production more accessible and commercially viable [5][41] - The rise of AI-driven social platforms is anticipated, with generative AI capabilities likely to create new forms of social interaction and engagement [45][47] - A surge in M&A activity is expected as traditional companies seek to acquire AI capabilities to remain competitive in the evolving landscape [48][49]

Core Viewpoint - The emergence of Gen Z founders in the AI startup scene is highlighted, showcasing their innovative approaches and successful fundraising efforts, particularly through the story of So la Sol uti ons founded by two MIT dropouts [2][3][4]. Group 1: Company Overview - So la Sol uti ons was founded by Jessica Wu and Neil Deshmukh, both from MIT, who identified inefficiencies in traditional RPA (Robotic Process Automation) and aimed to create AI agents that can automate complex tasks with minimal human intervention [7][8]. - The company has successfully raised a total of $21 million (approximately 150 million RMB) through a $3.5 million seed round and a $17.5 million Series A round, attracting notable investors such as Conviction and a16z [9][10]. Group 2: Market Position and Vision - So la Sol uti ons aims to disrupt the RPA market dominated by larger players by offering a simpler tool that automates data processing across various sectors, including logistics, insurance, and healthcare [8]. - The founders envision their AI agents taking over repetitive tasks, allowing human workers to focus on more creative endeavors [8]. Group 3: Investment Landscape - The article emphasizes the increasing presence of female investors in the tech venture capital space, with notable figures like Sarah Guo and Kimberly Tan leading investments in AI startups [11][12]. - The trend of Gen Z entrepreneurs gaining traction in the investment community is noted, with several successful startups emerging from this demographic, indicating a shift in the entrepreneurial landscape [15][16].

来源：环球网 【环球网科技综合报道】8月16日消息，据外媒The Verge报道，在经济学家们纷纷猜测股市是否正处于 一场可能很快破裂的人工智能泡沫之际，OpenAI首席执行官萨姆·奥尔特曼坦言，他相信行业正处于这 样的泡沫之中。 "我们是否正处于投资者整体对人工智能过度兴奋的阶段？"奥尔特曼在接受 The Verge等媒体采访时说 道。"我的看法是肯定的。" 他谈到，一些仅靠"三个人和一个想法"就成立的人工智能初创公司竟然能获得如此高估值的融资，这简 直"疯了"。"这不是理性的行为，有人会为此付出代价。" 在过去的一年里，几家人工智能初创公司筹 集了数十亿美元的资金。 在这场采访中，奥尔特曼还将当前市场对人工智能的反应与 90 年代的互联网泡沫进行了比较。"当泡沫 出现时，聪明人会对一些真相感到过度兴奋，"奥尔特曼说，"如果你回顾历史上的大多数泡沫，比如科 技泡沫，就会发现它们都是真实存在的。"（思瀚） ...

这可能是对于大语言模型（LLM）原理最清晰、易懂的解读。 前段时间，Google DeepMind 的首席科学家兼研究总监 Denny Zhou 在斯坦福大学的 CS25 课程中，分享了大语言模型推理的深刻洞见。 机器之心报道 编辑：泽南、杨文 所有学LLM的人都要知道的内容。 作为人工智能领域的领军人物，Denny Zhou 通过这场讲座对 LLM 推理机制及其优化方法进行了系统阐述，揭示了大模型推理的核心原理和最新进展。 Denny Zhou 总结了 四个关键点 ： Denny Zhou 不仅是 Google DeepMind 的顶尖科学家，还曾在 Google Brain 创立并领导了推理团队（Reasoning Team），现在该团队是 DeepMind 的一部分，专注于 开发具备推理能力的大语言模型，以推动人工智能通用智能的发展。 LLM 中的推理仅仅意味着在得出最终答案之前生成一系列中间 token，这是否与人类推理相似并不重要，关键在于，Transformer 模型通过生成许多中间 token，可以变得极其强大，而无需扩展模型的大小。 预训练模型即使未经任何微调，也具备推理能力。挑战在于，基于推 ...

Core Insights - OpenAI has launched its flagship model GPT-5, which is expected to significantly enhance capabilities in programming, writing, and complex problem-solving [1][2] - Sam Altman predicts that by the end of 2027, AI will achieve major scientific breakthroughs, potentially including advancements in drug development and cancer treatment [1][2][8] Group 1: GPT-5 Capabilities - GPT-5 can instantly generate professional-level software and supports real-time iterative updates [2][3] - The model has shown significant improvements in health-related queries, providing more accurate medical advice [39][40] - GPT-5's ability to create software in seconds represents a major leap from previous models, allowing users to express ideas and implement changes rapidly [6][10] Group 2: AI's Impact on Society - AI is described as a "double-edged sword," with the potential to either enhance human cognition or lead to complacency [2][3] - The job market is expected to undergo disruptive changes, but new opportunities will emerge as AI tools lower barriers to entrepreneurship [2][5] - A societal contract may need to be restructured to address the distribution of AI resources and prevent conflicts over access [2][45] Group 3: Development Bottlenecks - AI development faces four major bottlenecks: computing power, data availability, algorithm optimization, and productization [2][24] - The current demand for AI capabilities exceeds supply, necessitating significant investments in expanding computational resources [26][27] Group 4: Future Predictions - Altman anticipates that AI will lead to significant advancements in healthcare, with the potential for AI to autonomously guide drug development processes [40][41] - The concept of "superintelligence" is defined as AI systems surpassing human experts in core fields, which may occur sooner than expected [2][13] - The evolution of AI will likely lead to a society where AI is integrated into everyday life, becoming an invisible yet essential part of human experience [46]

Group 1 - OpenAI is preparing to sell approximately $6 billion in stock at a company valuation of $500 billion, with negotiations still in early stages [1] - The stock sale will involve current and former employees selling to existing investors such as SoftBank, Dragoneer Investment Group, and Thrive Capital, with Thrive Capital potentially leading the round [1] - OpenAI's valuation has seen exponential growth, rising from $15 billion in late 2022 to over $300 billion following a $40 billion funding round announced in March [2] Group 2 - The recent funding round included $8.3 billion in new capital, led by Dragoneer Investment Group with a $2.8 billion investment, and participation from various prominent firms [2] - OpenAI launched its latest AI model, GPT-5, which is claimed to be smarter and faster, particularly in writing, coding, and healthcare, although some users reported issues with accessing previous models [2] - CEO Sam Altman acknowledged that the impact of certain features in GPT-4 was underestimated, despite GPT-5 performing better in most aspects [3]