Core Viewpoint - The Unified Bus (UB) protocol aims to bridge the gap between bus and network architectures, addressing the limitations of both in high-performance computing environments, particularly in the context of deep learning and large-scale data processing [2][3][9]. Group 1: Necessity of UB - The necessity for UB arises from the fundamental contradiction in computer architecture between bus and network systems, which have historically operated as isolated entities [2]. - Traditional bus technologies like PCIe offer high performance within a limited physical scope, while network technologies provide scalability at the cost of increased latency and complexity [2][3]. Group 2: Architecture and Design Philosophy - UB's core mission is to create a unified interconnect that combines the ease of programming and performance of bus systems with the scalability of network systems [3]. - The protocol proposes a peer-to-peer architecture where all devices can access each other's memory directly, bypassing the CPU, thus achieving zero-copy data transfer and ultra-low latency [5][6]. Group 3: Key Features of UB - UB provides a unified memory semantic that abstracts the physical differences between bus and network technologies, allowing applications to access remote memory as if it were local [7][8]. - The protocol introduces innovative operations like "Write with Immediate" and "Send with Immediate," which combine data transfer and notification into single atomic operations, reducing overhead [17][18]. Group 4: Jetty Abstraction - The Jetty abstraction replaces traditional connection models with a more flexible, connectionless approach, allowing multiple applications to share resources efficiently [22][23]. - This model simplifies the management of communication states and enhances scalability by treating requests as independent entities rather than requiring dedicated connections [27][28]. Group 5: Transaction Ordering - UB distinguishes between execution order and completion order, allowing for flexible optimization in distributed systems while maintaining consistency [42][45]. - The protocol supports various transaction service modes, enabling applications to choose between performance and consistency based on their specific needs [45].

Core Insights - Dario Amodei, CEO of Anthropic, has transformed the company into a major player in the AI field, driven by personal experiences and a commitment to accelerate technological advancements [1][5][10] - Anthropic is negotiating a funding round of $3 billion to $5 billion, potentially raising its valuation to $170 billion, reflecting strong investor interest in AI [3][74] - The company has seen rapid growth in annual recurring revenue (ARR), increasing from $1.4 billion in March 2025 to nearly $4.5 billion by July 2025 [5][61] Company Overview - Anthropic was founded during the COVID-19 pandemic with a mission to create advanced language models while establishing safety protocols [49][52] - The company has raised nearly $20 billion in funding, including $8 billion from Amazon and $3 billion from Google, indicating strong investor confidence [52][75] - Anthropic's strategy focuses on selling AI technology to enterprises, which has proven lucrative and has attracted a diverse client base, including major corporations like Pfizer and United Airlines [58][59] Financial Performance - Anthropic's revenue has surged, with projections indicating a rise from $0 to $100 million in 2023, and from $1 billion to an estimated $4.5 billion in 2025 [61][66] - Despite high revenue growth, the company is facing significant losses, projected to be around $3 billion for the year, raising questions about the sustainability of its business model [62][66] - The average spending of enterprise clients has increased fivefold, showcasing the growing demand for Anthropic's AI solutions [61] Market Position and Competition - The AI industry is experiencing intense competition, with new models emerging that challenge established players, such as the DeepSeek R1 model, which is priced significantly lower than competitors [70][71] - Anthropic's models are designed to maintain a competitive edge in specific domains, particularly in programming, where early adoption can lead to substantial advantages [69] - The company is also focused on improving the efficiency of its models to reduce operational costs, which is critical for long-term viability [65][66] Future Outlook - Amodei emphasizes the need for rapid development in AI technology, with plans to accelerate the release of new models [77] - The company is investing in research to ensure AI systems align with human values and goals, addressing potential safety concerns as models become more advanced [82][86] - Anthropic's commitment to understanding AI's internal workings and ensuring its responsible use is a key part of its strategy moving forward [85]

AI Agent 领域也存在 scaling law，甚至还在加速。 2022 年 ChatGPT 刚发布时能够实现的代码任务差不多等同于人类耗时 30s 的任务，到今天， AI Agent 已经能够自主完成需要人类花费一个小时的 coding 任务。"任务长度"是一个相当直观地测量 AI Agent 能力变化的标准。 编译：haozhen 编辑：Siqi AI 独立研究机构 META 的数据分析发现，Agent 能够完成的任务长度正以指数级增长，大约每 7 个 月翻一倍，预计 2029 年 Agent 能够完成时长为 1 个工作月的任务。 有意思的是，最近这一趋势甚至还在加速，2024-2025 年 Agent 能完成的任务长度约每 4 个月翻一 倍，如果这种更快的趋势持续下去，Agent 可能在 2027 年就能完成长达一个月的任务。 本文是对 META、Forethought 和 AI Digest 研究对于 agent scaling law 的整理编译。AI 研究人员们认 为，AI scaling law 的终局是 AI agent 自主开发 AI agent，到了那个时候我们就会进入软件智能爆炸时 ...