一亿美元能买一栋别墅，但买不了梦想？ 最近，Meta 内部发生了一些有意思的事情 —— 一边是扎克伯格动辄上亿美金薪资招兵买马，高调组建超级智能团队；另一边是一些老员工宣布开启新的 「冒险之旅」，转投其他 AI 公司。 今天，有两位资深研究者宣布离开 Meta，一位是专注于强化学习的 Rishabh Agarwal（去向未定）；另一位是已经在 Meta 工作了 12 年、参与了 PyTorch 构建的 Bert Maher（确定加入 Anthropic）。 看来，除了小扎挖不到的人，还有一些他留不住的人。甚至有人嘲讽「钱买不到顶级研究员」。 不过，还有很多选择离开的人，可能是因为钱没给够。在超级智能实验室成立之后，Meta 内部的待遇差距多次引发争议。 前 Meta 研究员 Rohan Anil（现 Anthropic）曾发帖称「非超级智能研究者待遇次等，像巨型社会实验」。 有人认为 Meta 内部薪资差距（同事赚 1-2 亿美元）会杀死工作动力，导致更多离职。 当外部挖不来真正的梦想家，内部又因分配不均而人心浮动，这背后折射出的，可能是比薪酬更深层的结构性问题。 这让一些人联想到了「90 年代的微软」，即由 ...

Meta拥抱"多云"架构 多年来，Meta一直主要依赖其自建自营的数据中心来运行旗下社交应用及其他服务。 然而，在过去几年中，Meta已开始拥抱"多云"（multi-cloud）策略，与亚马逊云服务和微软Azure均签 订了协议，以租用云服务器并共同开发由Meta创建的AI应用开发工具集PyTorch。此外，Meta还从甲骨 文和CoreWeave等公司租用云服务器。 将谷歌云纳入其供应商体系，将使Meta成为与苹果、OpenAI并驾齐驱的全球最大云客户之一。此举不 仅能增强其在全球范围服务用户的能力，也能使其在与亚马逊和微软云业务的价格谈判中获得更多筹 码。 亦敌亦友，谷歌云赢得竞争对手青睐 谷歌与Meta的这笔交易，是科技巨头之间复杂关系的最新例证。 谷歌云计算业务取得里程碑式的胜利，与其主要竞争对手之一Meta达成了一项历史性协议。 据科技媒体The Information报道，有知情人士透露，谷歌已和Meta达成协议，允许后者使用谷歌云的服 务器、存储、网络和其他服务，协议为期六年、价值超过100亿美元。 据悉，该交易是谷歌云自成立17年以来，已知的规模最大的合同之一，标志着其在追赶亚马逊和微软的 道 ...

Core Viewpoint - The article highlights the rapid growth and potential of Fireworks AI, a cloud service provider founded by Lin Qiao and her team, which is seeking a $4 billion valuation in its upcoming funding round, marking a remarkable valuation increase within a year [2][10]. Company Overview - Fireworks AI was established in 2022 in Redwood City, California, by Lin Qiao and six co-founders, with Lin Qiao serving as the CEO [3]. - Lin Qiao has 24 years of industry experience, having worked at IBM, LinkedIn, and Meta, where she led the development of PyTorch, a leading open-source machine learning framework [4]. Business Model and Performance - Fireworks AI focuses on providing low-cost, high-performance, and customizable AI models, allowing developers to efficiently deploy AI solutions [5][11]. - The company has over 100 models and reported an annual revenue exceeding $200 million, with expectations to grow to $300 million by the end of the year [5]. Competitive Landscape - Fireworks AI's main competitor, Together, reported an annual revenue of $150 million and a valuation of $3 billion as of March 2023 [6]. - The company differentiates itself from traditional cloud service providers by offering more cost-effective and high-performance solutions [11]. Funding and Valuation Growth - Fireworks AI completed its seed round in 2022 and Series A funding in 2023, achieving a valuation of $552 million, with investments from top venture capital firms including Sequoia Capital and Nvidia [12]. - The company is in discussions with Lightspeed Venture Partners and Index Ventures for further funding, which could elevate its valuation to $4 billion, representing a sevenfold increase in just one year [12][10]. Industry Context - The article notes the competitive relationship between Fireworks AI and Nvidia, as both companies operate in overlapping markets, with Nvidia also investing in various AI-related ventures [13][14]. - The potential for Nvidia to acquire Fireworks AI is highlighted, given the dual nature of their relationship as both collaborators and competitors [15]. Chinese Entrepreneurs in AI - The article emphasizes the significant contributions of Chinese entrepreneurs in the AI sector, with several companies founded by Chinese individuals achieving unicorn status and collectively valued at over $30 billion [16][19]. - Notable examples include Scale AI, founded by Alexandr Wang and Lucy Guo, and Cognition AI, which has seen rapid valuation growth [17][18].

将投中网设为"星标⭐"，第一时间收获最新推送 在这个"创新与机遇"的时代，华人正在创造奇迹。 作者丨陈美 来源丨 投中网 如果要说，最快诞生独角兽的地方，恐怕要数AI创业圈了。 近日，据外媒报道，Fireworks AI——一家新兴的云服务提供商，正在寻求以40亿美元（约合288亿元）估值（包含融资金额）进行新一轮融 资。 知名风险投资机构Lightspeed Venture Partners（美国光速）和Index Ventures等正就领投事宜，进行深入讨论。一旦融资成功，Fireworks AI 创始人将创造1年估值增长7年，3年干出一个288亿元独角兽的奇迹。 华人女性创业，年化营收突破2亿美元 这是一位华人女性在加州创业的故事。 2022年，Meta前高级工程总监乔琳（Lin Qiao）与6位联合创始人，一起在美国加州雷德伍德市成立了Fireworks AI。联合创始人兼首席执行 官的乔琳（Lin Qiao），本硕毕业于复旦大学计算机科学专业，之后在美国加州大学圣巴巴拉分校（UC Santa Barbara）获得计算机科学博 士学位。 在创业前，乔琳（Lin Qiao）拥有24年的行业经验，可谓是一 ...

Core Insights - Fireworks AI, a cloud service provider, is negotiating a funding round with a valuation of $4 billion, which represents a more than sevenfold increase from the previous year [1][2] - The company was founded by former engineers from Meta and Google, and has previously raised approximately $77 million from investors including Sequoia Capital and Benchmark [2] Financial Performance - Fireworks' annualized revenue has surpassed $200 million, with a monthly average of $17 million, and is projected to reach $300 million by the end of the year [3] - The company's gross margin is approximately 50%, which is comparable to other inference service providers but lower than the 70%+ margins typical in subscription software businesses [3][5] - Fireworks aims to improve its gross margin to 60% by focusing on GPU optimization [5] Competitive Landscape - NVIDIA has emerged as a new competitor to Fireworks and other GPU cloud service resellers, having launched its own GPU cloud marketplace after acquiring inference service provider Lepton [4] - Fireworks competes with companies like Together AI and Baseten, which also resell NVIDIA-powered cloud servers [4] - The company differentiates itself by offering faster and more cost-effective solutions for customizing and running open-source models compared to traditional cloud service providers like Amazon and Google [3] Strategic Focus - Fireworks is concentrating on optimizing GPU resource utilization to address financial challenges and meet customer demand, which can fluctuate significantly [5] - The CEO emphasized the importance of building tools and infrastructure that enable application developers to customize models and enhance inference quality, speed, and user concurrency [5]

Core Viewpoint - The open-source project ZLUDA, which enables non-NVIDIA chips to run CUDA, has been revived after facing near bankruptcy due to the withdrawal of AMD's support. A mysterious organization has stepped in to provide assistance, allowing the project to continue its development and support for large model workloads [1][2][12]. Historical Development - ZLUDA was initiated by Andrzej Janik, who previously worked at Intel, aiming to allow CUDA programs to run on non-NVIDIA platforms [4][5]. - Initially, ZLUDA was taken over by Intel as an internal project to run CUDA programs on Intel GPUs, but it was soon terminated [6][9]. - In 2022, ZLUDA received support from AMD but was again halted in February 2024 after NVIDIA released CUDA 11.6, which restricted reverse engineering on non-NVIDIA platforms [10][11][12]. Recent Developments - In October 2024, Janik announced that ZLUDA had received support from a mysterious organization, focusing on machine learning and aiming to restore the project to its previous state by Q3 2025 [13][15]. - The project has added a new full-time developer, Violet, who has made significant improvements, particularly in supporting large language model workloads [17]. Technical Progress - ZLUDA is working on enabling 32-bit PhysX support, with community contributors identifying and fixing errors that may also affect 64-bit CUDA functionality [19]. - A test project named llm.c is being developed to run the GPT-2 model using CUDA, marking ZLUDA's first attempt to handle both standard CUDA functions and specialized libraries like cuBLAS [20][22]. - The team has made progress in supporting 16 out of 44 required functions for the test program, indicating a step closer to full functionality [25]. Accuracy and Logging Improvements - ZLUDA aims to run standard CUDA programs on non-NVIDIA GPUs while matching NVIDIA hardware as closely as possible. Recent efforts have focused on improving accuracy by implementing PTX "scan" tests to ensure correct results across all inputs [26][28]. - The logging system has been significantly upgraded to track previously invisible activities and internal behaviors, which is crucial for running any CUDA-based software on ZLUDA [31][33]. Runtime Compiler Compatibility - ZLUDA has addressed issues related to the dynamic compilation of device code necessary for compatibility with modern GPU frameworks. Recent changes in the ROCm/HIP ecosystem have led to unexpected errors, but the ZLUDA team has resolved these problems [34][36][38].

Core Viewpoint - The article discusses the launch of Stanford University's CS336 course "Language Models from Scratch," which aims to provide a comprehensive understanding of language models through practical development and implementation [5][7]. Course Overview - The course focuses on the foundational aspects of language models, which are essential for modern natural language processing (NLP) applications. It emphasizes the importance of understanding language models for scientists and engineers in the fields of AI and ML [5][7]. - The course is structured into five major modules: Foundations, Systems, Extensions, Data, and Alignment & Reinforcement Learning [7]. Course Requirements - Students are expected to have proficiency in Python, as most assignments will require extensive coding. The course will provide minimal scaffolding, resulting in a higher volume of code written by students compared to other AI courses [7]. - A background in deep learning and system optimization is necessary, particularly familiarity with PyTorch and basic system concepts like memory hierarchy [7]. - Foundational knowledge in calculus, linear algebra, probability, and statistics is required, along with a basic understanding of machine learning principles [7]. Assignments - The course includes several assignments that cover various aspects of language model development, such as implementing a BPE tokenizer, training models on specific datasets, and optimizing performance on GPUs [8]. - Assignments are designed to simulate real-world challenges, including data processing and model alignment, with a focus on practical application and hands-on experience [8]. Course Schedule - The course is structured with a detailed schedule that outlines topics, materials, and deadlines for assignments, ensuring a systematic approach to learning [9].

Core Viewpoint - The article emphasizes the rapid development of large models and the need for efficient migration and deployment solutions in the AI ecosystem, particularly through the use of MindSpore, which aims to facilitate seamless integration and performance optimization for developers [1][2]. Group 1: Migration Challenges - The first challenge is fast migration, enabling zero-cost migration of third-party framework models while ensuring complete alignment in model accuracy. MindSpore achieves this through a threefold compatibility approach, allowing for zero-code migration of mainstream models and improving training performance by 5% while maintaining distributed parallel strategies [4]. - The second challenge is rapid deployment, automating the entire training-to-inference process to make large model deployment as simple as executing a single command [2]. Group 2: Training and Inference Solutions - MindSpore supports Day 0 migration for training, providing a "no-sense intelligent translation" capability across frameworks. It utilizes tools like MindSpeed/Megatron for seamless PyTorch model migration, achieving near-zero migration loss for popular models [4]. - In inference deployment, the vLLM-MindSpore plugin allows for HuggingFace models to be deployed in under 30 minutes, with an 80% reduction in weight loading time for large models [5][6]. Group 3: Open Source and Community Engagement - Since its open-source inception on March 28, 2020, MindSpore has fostered a vibrant developer community, with over 1.2 million downloads and contributions from more than 46,000 developers across 2400 cities [7]. - The company promotes a collaborative ecosystem through community governance, providing free computing resources and knowledge sharing across 20+ technical special interest groups (SIGs) [8].

Core Viewpoint - The article discusses the tension surrounding the concept of "openness" in AI, highlighting the phenomenon of "open-washing" where organizations label their models as open-source while imposing restrictive licenses that limit true freedom of use [1][3][4]. Group 1: Open Source and AI - The rise of open-source AI has created a self-accelerating "virtuous cycle," but there is a silent war over the definition of "openness" [1][4]. - Matt White introduced the "Model Open Framework" (MOF) to clarify standards and distinguish true open-source contributors [4]. - The "OpenMDW License" aims to provide maximum freedom for users of AI models, addressing the inadequacy of traditional software licenses in the context of AI [4][7]. Group 2: Global Engagement and Community - PyTorch Day aims to foster a global movement, with significant user engagement from China, where 70% to 80% of traffic on documentation sites originates [6]. - The event serves as a platform for showcasing innovative open-source projects and facilitating knowledge exchange among local engineers and researchers [11]. Group 3: Licensing and Usage - The core of "openness" in AI should be viewed through the lens of licensing, determining what users can do with the models [7]. - Licenses designed specifically for open models consider various aspects, including model architecture, weights, datasets, and documentation, unlike traditional licenses [7]. Group 4: Collaboration and Standards - Collaboration among tech giants and new entrants is essential for advancing open-source AI, with PyTorch serving as a trusted platform for cooperation [9][10]. - The Linux Foundation plays a crucial role in establishing neutral standards that ensure long-term viability and widespread acceptance of protocols [10]. Group 5: Future Trends and Education - The rapid development of AI agents and architectures necessitates a focus on open standards, with organizations like PyTorch and the Linux Foundation playing pivotal roles [10]. - Educators must adapt to the AI era, learning how to effectively integrate AI tools into their teaching without compromising core skill development [13][14]. Group 6: Challenges and Responsibilities - The article emphasizes the importance of addressing the "digital content authenticity" crisis, as AI-generated content becomes increasingly indistinguishable from real content [15]. - The need for responsible AI practices is highlighted, particularly in the context of misinformation and the potential misuse of technology [15].

Group 1: Market Positioning - GpuGeek targets algorithm engineers with high-quality computing power needs, avoiding homogenization in competition by understanding core user demands [2] - The platform offers a full range of GPU resources from RTX 4090 to A800, catering to various computational needs from entry-level development to large-scale model training [2] - GpuGeek employs a differentiated hardware strategy, optimizing storage, memory, and network configurations for different application scenarios, supporting up to 8 GPUs working in tandem [2] Group 2: Innovative Service Model - GpuGeek addresses the complexity and time consumption of AI development environment setup by providing pre-installed mainstream deep learning frameworks, allowing users to start coding within half a minute [3] - The introduction of diverse billing options, including daily, weekly, monthly, and bidding models, enhances user experience and operational efficiency by allowing cost-effective access to computing resources [3] - The platform features an online IDE tool that enables development from anywhere, integrating seamlessly with major code repositories for version management and team collaboration [3] Group 3: Global Expansion - GpuGeek has strategically deployed global nodes in locations such as Hong Kong, Dallas, and Europe, facilitating international academic resource access and efficient collaboration for multinational research teams [4] - This global layout enhances the platform's market share and improves service stability and reliability [4] Group 4: Open Ecosystem Development - GpuGeek's model marketplace includes cutting-edge AI models, providing users with convenient access for model invocation and deployment, enriching platform functionality [5] - The platform supports open-source community building, encouraging users to share experiences and technical achievements, fostering a vibrant learning community [5] Group 5: Core Competitiveness - As a professional AI infrastructure service provider, GpuGeek has established core competitiveness in the AI infrastructure market through precise market positioning, innovative service models, global strategic layout, and open ecosystem development [6] - The company aims to bridge technology and application, focusing on the long-term development of AI technology and industry [6] - GpuGeek is positioned to play a significant role in promoting industry innovation and ecological prosperity as AI technology continues to evolve [6]