National Transportation Safety Board 已发布对 2026 年 9 月 6 日在 Miami International Airport 发生的 21 Air Flight 7598（一架 Boeing 767-33A 货机）跑道冲出事故的初步调查更新。调查人员已成功回收并分析了驾驶舱语音记录器和飞行数据记录器的数据。回收的飞行数据约 54 小时，包含 400 多个参数；语音记录为四通道、两小时的高质量音频。

语音记录显示，机组在开始放襟翼并放下起落架后不久，获准在 30 号跑道着陆。进近过程中，驾驶舱对飞机速度表示担忧。录音结束前约 1 分 42 秒，一名飞行员提醒另一名飞行员速度过快。此后在剩余的飞行时间里，录音中反复出现对速度过高的评论，但另一名飞行员并未持续做出口头回应。

在下降末段，飞机触发了多项自动电子警告，包括下沉率警告和低地形警报。尽管出现这些警告，着陆程序仍继续。机头和右主起落架在地速 158 节时首次接地，随后机组尝试复飞。在语音记录结束前约 15 秒，一名飞行员宣布复飞并增大了油门。

飞行数据记录显示，复飞尝试时间很短。发起机动后约 4 秒内，油门被回至怠速并重新踩下刹车。随后飞机以约 117 节的地速偏离了铺装路面。对飞行数据的进一步分析表明，在此次过程中，速度板和反推装置均未部署。调查仍在进行中，相关人员将继续审查证据并整理驾驶舱通话的正式文字记录。 The National Transportation Safety Board has released a preliminary update regarding its ongoing investigation into the September 6, 2026, runway excursion involving 21 Air Flight 7598, a Boeing 767-33A cargo aircraft, at Miami International Airport. Investigators have successfully recovered and analyzed data from both the cockpit voice recorder and the flight data recorder. The recovered flight data spans approximately 54 hours and includes over 400 parameters, while the voice recording provides two hours of high-quality audio across four channels.

According to the voice recorder data, the flight crew received clearance to land on runway 30 shortly after they began configuring the aircraft with flaps and landing gear. As the approach continued, concerns regarding the aircraft's speed were noted in the cockpit. Approximately 1 minute and 42 seconds before the recording ended, one pilot alerted the other that they were traveling too fast. While the cockpit recording captured repeated comments about this excessive speed for the remainder of the flight, there was no consistent verbal response from the other pilot.

During the final stages of the descent, the aircraft triggered several automated electronic warnings, including alerts for sink rate and low terrain. Despite these warnings, the landing sequence proceeded. Following the initial touchdown of the nose and right main gear at a groundspeed of 158 knots, the flight crew attempted a go-around maneuver. Approximately 15 seconds before the end of the voice recording, a pilot called for this procedure, and throttle settings were increased.

Data from the flight recorder shows that the go-around attempt was brief. Within four seconds of initiating the maneuver, the throttles were reduced to idle and the brakes were re-applied. The aircraft subsequently departed the paved surface at a groundspeed of 117 knots. Further analysis of the flight recorder data indicates that neither the speed brakes nor the thrust reversers were deployed during this sequence. The investigation remains active as officials continue to review the evidence and prepare a formal transcript of the cockpit communications.

Agents API 为开发者提供了一种通过托管的 Codex harness 构建持久化云端 agent 的方式。该架构让 OpenAI 负责会话管理、编排、上下文压缩和恢复等复杂底层工作，而开发者则保留对执行环境以及 agent 可访问具体工具的控制权。将这些后端职责交由平台处理后，开发者可以专注于打造用于调查警报、审阅文档或与各类软件平台交互等专业任务的 agent 。

系统的核心由四个基本概念构成：agent 、 environment 、 session 和 events 。 agent 由其指定的模型、指令、工具以及可能需要的 Model Context Protocol (MCP) 服务器定义；environment 可以是由 OpenAI 托管的沙箱，也可以是自托管的沙箱，为 agent 提供运行代码、编辑文件和发挥能力的安全空间。 session 是 agent 持续工作的持久实例，events 则记录这些交互过程中产生的输入与输出。

托管的 harness 提供了一整套增强性能与可靠性的功能，支持在沙箱中执行代码、应用多种技能并连接外部数据源。 API 还便于对 agent 进行主动管理，包括在任务过程中引导 agent 、总结过往工作以优化上下文窗口，以及将复杂目标委派给子代理。

要开始使用，开发者需创建一个 session 、配置 agent 的能力并提供初始任务。环境准备就绪后 agent 会开始处理，开发者可通过流式传输或 webhook 监控进度。由于 API 会维护会话状态，agent 能无缝恢复工作而无需重建对话历史。需要注意的是，Agents API 目前要求数据驻留在 United States，且不支持 Zero Data Retention 。 The Agents API offers developers a way to build durable, cloud-based agents by utilizing a managed Codex harness. This architecture allows OpenAI to handle the complex underlying processes of session management, orchestration, context compaction, and recovery, while developers maintain control over the execution environment and the specific tools the agent can access. By offloading these backend responsibilities, developers can focus on creating agents that perform specialized tasks, such as investigating alerts, reviewing documents, or interacting with various software platforms.

At the heart of the system are four primary concepts: the agent, the environment, the session, and the events. The agent is defined by its specific model, instructions, tools, and any Model Context Protocol (MCP) servers it may require. The environment can either be an OpenAI-hosted sandbox or a self-hosted one, providing a secure space where the agent can run code, edit files, and utilize its capabilities. Sessions serve as durable instances where the agent maintains work over time, and events track the input and output produced throughout these interactions.

The managed harness provides a robust suite of features designed to enhance agent performance and reliability. It supports the execution of code in sandboxes, the application of various skills, and the ability to connect to external data sources. Furthermore, the API facilitates active management of the agent, including the ability to steer it during tasks, summarize past work to optimize the context window, and delegate complex objectives to subagents.

To get started, developers create a session, configure the agent's capabilities, and provide an initial task. Once the environment is ready, the agent begins processing, and developers can monitor progress via streaming or webhooks. Because the API maintains session state, agents can resume work seamlessly without the need to reconstruct the conversation history. While this setup provides significant power, developers should note that the Agents API currently requires data residency in the United States and does not support Zero Data Retention.

一项名为 Deathray 的新漏洞显示，一个简单的 WebGPU 着色器就能冻结 Mac 的桌面界面，使机器在强制重启前无法使用。用户只需访问恶意网站，就可能在 macOS 上的主流浏览器（包括 Chrome 、 Firefox 和 Safari）触发这种死机。虽然该手法会持续锁死 Mac，但对其他操作系统通常只会导致浏览器标签页卡死，关闭标签页即可恢复。

该漏洞通过同时使用 compute shader 和 render shader 造成 GPU 死锁：compute shader 进入无限忙循环，不断向共享缓冲区写入数据；与此同时，vertex shader 试图从同一缓冲区读取，由于 compute 进程永远不会完成，读取操作被阻塞。这种"交通拥堵"进一步蔓延到 WindowServer 进程，导致桌面环境无响应并偶尔出现图形伪影。

这一问题暴露了 Apple 面临的反复挑战。早在 2023 年就曾发现一个针对 WebGL 的类似漏洞 ShadyShader，当时产生了 CVE 并尝试通过输入校验进行缓解，但 Deathray 表明 WebGPU 目前缺乏防止此类失控循环的有效防护。专家认为，问题可能与 M-series chips 的架构有关：操作系统内核无法直接抢占 GPU，这一职责由独立协处理器上的固件承担。

Deathray 在 2026 年 7 月被披露给 Apple 。 Apple 最初承认问题并表示会修复，但随后改变立场，将此类行为归类为潜在的挂起或崩溃，而非正式的安全问题，因此不会列为优先处理。这一回应引发争议，凸显了严格的安全技术定义与公众期望浏览器作为可靠、隔离的 web 内容沙箱之间的冲突。

尽管 Deathray 不会导致数据窃取或远程代码执行，但它只需一次点击就能扰乱系统，带来严重不便。发现该漏洞的研究人员主张应通过长期修复来改进系统对无响应着色器的处理，而非简单禁用 WebGPU 。在采取此类修复之前，这一漏洞再次提醒我们：在现代硬件上浏览互联网时，用户对 web 浏览器的信任是多么脆弱。 A newly discovered vulnerability dubbed the Deathray demonstrates how a simple WebGPU shader can freeze a Mac's desktop interface, rendering the machine effectively unusable until a forced restart. By merely visiting a malicious website, a user can trigger this hang across major browsers including Chrome, Firefox, and Safari on MacOS. While the technique consistently locks up Mac systems, the impact on other operating systems remains limited to browser-tab freezes, which are easily resolved by closing the tab.

The exploit functions by utilizing a combination of compute and render shaders that essentially create a GPU deadlock. The compute shader executes an infinite busy loop, constantly writing data to a shared buffer. Simultaneously, a vertex shader attempts to read from that same buffer, becoming trapped because the compute process never completes. This traffic jam overflows into the WindowServer process, causing the desktop environment to become unresponsive and occasionally displaying graphical artifacts.

This issue highlights a recurring challenge for Apple, as a similar exploit involving WebGL, known as ShadyShader, was identified in 2023. While that incident led to a CVE and an attempt at mitigation through input validation, the Deathray suggests that WebGPU currently lacks the robust defenses needed to prevent such runaway loops. Experts believe this difficulty may stem from the architecture of M-series chips, where the OS kernel cannot directly pre-empt the GPU, leaving that responsibility to firmware on a separate coprocessor.

Following the discovery of the Deathray, the issue was disclosed to Apple in July 2026. Although the company initially acknowledged the problem and indicated an intent to fix it, they later reversed their stance. Apple now classifies the behavior as a potential hang or crash rather than a formal security issue, suggesting it will not be prioritized. This response has sparked debate, as it pits a strict technical definition of security against the expectation that a browser should act as a reliable, isolated sandbox for web content.

While the Deathray does not facilitate data theft or remote code execution, its ability to disrupt a system via a single click remains a significant inconvenience. The researcher who discovered the exploit advocates for a long-term fix that improves how the system handles unresponsive shaders, rather than simply disabling WebGPU. Until such a resolution is implemented, the vulnerability serves as a stark reminder of the fragile trust users place in web browsers when navigating the internet on modern hardware.

Music Theory for the 21st-Century Classroom by Robert Hutchinson 是一本综合教材，旨在弥合传统音乐教育与现代学生多样化需求之间的鸿沟。全书从音高、记谱法、音阶、调号与节奏等基础概念入手，打下牢固基础；并循序渐进地讲解音程、三和弦与七和弦，为更复杂的和声分析做好准备，包括在各种音乐语境中的和声进行与和声功能研究。

书中有大量篇幅用于结构与风格分析，既探讨旋律与曲式，剖析流行音乐的结构，又覆盖二部、三部、奏鸣曲式与回旋曲式等古典曲式惯例。作者将这些传统的分析工具与实用技法相结合，如声部进行、对位法与伴奏织体的处理，使学生既能理解音乐的微小动机，又能把握大型作品的构造。

课程还延伸到高级和声领域，讨论染色和声（如借属和弦、那不勒斯和弦与增六和弦）以及复杂的转调技术。书中纳入了关于爵士理论、印象主义、音集合 / 序列理论、序列主义与极简主义的章节，囊括了二十一世纪音乐的广度。这些后调性概念的引入提供了前瞻性的视角，使教材对当代课堂具有高度相关性。

全书每一主题均配有练习题，鼓励主动学习并巩固技能。从基础理论到复杂分析体系的逻辑推进，使本书成为学生与教师皆宜的多功能工具。总之，它既尊重音乐理论的历史，又能有效应对当今多元音乐风貌的需求。 Music Theory for the 21st-Century Classroom by Robert Hutchinson is a comprehensive resource designed to bridge the gap between traditional music education and the diverse needs of modern students. The book establishes a solid foundation by covering fundamental concepts such as pitch, notation, scales, key signatures, and rhythm. By moving systematically through intervals, triads, and seventh chords, it prepares students for more complex harmonic analysis, including the study of harmonic progressions and function within various musical contexts.

A significant portion of the text is dedicated to both structural and stylistic analysis. It explores melody and form, providing insights into popular music structures, as well as classical conventions like binary, ternary, sonata, and rondo forms. The book balances these traditional analytical tools with practical applications, such as voice leading, counterpoint, and the use of accompanimental textures. This approach ensures that students gain an understanding of how music is built, from the smallest melodic motives to large-scale compositions.

The curriculum extends into advanced harmonic territory, addressing chromatic harmonies like secondary dominants, the Neapolitan chord, and augmented sixth chords, alongside sophisticated modulation techniques. By incorporating sections on jazz theory, impressionism, set theory, serialism, and minimalism, the author embraces the breadth of 21st-century music. This inclusion of post-tonal concepts provides a forward-looking perspective that makes the material highly relevant for contemporary classrooms.

Throughout the text, practice exercises are provided for every topic, encouraging active learning and the reinforcement of skills. The logical progression from basic music theory to complex analytical systems makes this book a versatile tool for students and instructors alike. Ultimately, it serves as a robust guide that respects the history of music theory while effectively catering to the eclectic demands of today's musical landscape.

软件结合了速度、金钱、复杂性和高度抽象的独特特性，能深刻地影响人的心理。尽管大多数软件产品本质上不过是被美化的电子表格，但构建它们的过程常常会扭曲原本理性的人们的视角。想法与实现之间缺乏天然的摩擦，造成一种危险的氛围，使每一次潜在改动都显得迫在眉睫、具有战略意义且刻不容缓。与实体建筑不同，改动平面图会产生明显可见的成本，而改变软件的代价往往是隐形的，通过动力丧失、上下文切换和架构侵蚀逐渐累积。

这种易改性催生出一种文化：每一个"可以"很快就变成"应该"。软件开发中很少有明确的"完成"定义，团队因此常陷入无休止的迭代。每一个按钮、查询或基础设施组件在理论上都可以改进、优化或扩展。行业巨大的金钱潜力进一步放大了这种倾向，使人们对平凡任务赋予过度的情感重要性：一次按钮重设计不再只是界面调整，而被寄予未来财富或公司成功的希望。

复杂性也被当作一种心理拐杖，为那些构建、辩论和重写繁复系统的人提供了重要感与地位感。在许多情况下，这些系统自我强化，其存在更多是为了支撑构建它们的组织，而不是为了解决实际问题。由此形成一个反馈循环：经理、创始人和工程师们觉得必须不断拉动各种杠杆，把公司本身看作一段永远可变、未完成、需要下一次重构的软件。

对速度的痴迷从根本上排斥耐心。实际上，客户、工程师和企业都需要时间来成熟并找到稳定性。当组织把耐心解读为不作为时，就会用一阵阵发布、转向和重塑去取代它，而这往往掩盖了最初更简单的问题。讽刺的是，经过多年堆叠复杂性和反复转向后，团队常常又建议重建最初那个简单的产品，并确信自己发现了什么深刻洞见。

应对这种系统性疯狂的办法不是为了停滞而放慢脚步，而是要重拾一种比例感。这需要纪律性：认识到并非每一次减速都是危机，并非每一个想法都该上路线图，也并非每一个工具都必须成长为庞大平台。承认多数软件的最终目标不过是成为一个好用的工具，团队就能避免对已在正常运行的事物进行自我毁灭式的反复修补。真正的成功在于，当事情运转良好时学会按兵不动，把实用性置于那无尽而神经质的创新循环之上。 Software possesses a unique combination of speed, money, complexity, and abstraction that can profoundly affect human psychology. While most software products are essentially glorified spreadsheets, the process of building them often warps the perspectives of otherwise rational people. The lack of natural friction between an idea and its implementation creates a dangerous environment where every potential change feels immediate, strategic, and urgent. Unlike physical construction, where modifying a floor plan incurs clear, visible costs, the expense of changing software is often hidden, accumulating through lost momentum, context switching, and architectural erosion.

This ease of modification leads to a culture where every "could" quickly transforms into a "should." Because there is no clear definition of "done" in software development, teams often fall into the trap of endless iteration. Every button, query, or infrastructure component can theoretically be improved, refined, or scaled. This environment is exacerbated by the vast financial potential of the industry, which causes individuals to attach outsized emotional importance to mundane tasks. A button redesign is no longer just a UI change, but something weighted with the hope of future riches or company-wide success.

Complexity is further used as a psychological crutch, providing a sense of importance and status to those who build, debate, and rewrite intricate systems. In many cases, these systems become self-reinforcing, existing primarily to support the organizations that built them rather than to solve actual problems. This creates a feedback loop where managers, founders, and engineers feel the need to constantly pull levers. They view the company itself as a piece of software that is perpetually mutable, unfinished, and in need of the next refactor.

This obsession with velocity is fundamentally hostile to patience. In reality, customers, engineers, and businesses require time to mature and find stability. When organizations interpret patience as inactivity, they replace it with a flurry of shipping, pivoting, and reinventing that often obscures the original, simpler problem. The irony is that after years of layering complexity and changing direction, teams often eventually propose rebuilding the initial, simple product, convinced they have discovered a profound insight.

The solution to this systemic madness is not to move slowly for the sake of stagnation, but to regain a sense of proportion. It requires the discipline to recognize that not every slowdown is a crisis, not every idea belongs on the roadmap, and not every tool needs to become a sprawling platform. By acknowledging that most software's ultimate goal is simply to be a useful tool, teams can avoid the self-destructive impulse to tinker with things that already function. True success lies in the ability to leave things alone when they are working, ultimately prioritizing utility over the endless cycle of neurotic innovation.

Neki 是 PlanetScale 推出的全新分片式 Postgres 产品，旨在帮助团队突破单机扩展数据库的限制。凭借多年管理大规模 MySQL 集群的实战经验，PlanetScale 开发了 Neki，让用户能够将 Postgres 数据分布到多台机器，同时保持与标准 Postgres 驱动、 ORM 和连接字符串的兼容性。 Neki 在每个分片上都运行真实的 Postgres（而非自研存储引擎），因此能够完整保留现有扩展和 SQL 功能。

Neki 的核心架构由四部分组成。 Neki routers 负责接收和路由流量，充当解析查询、生成分布式执行计划并汇总结果的网关；每个 shard 由一个主库和分布在不同可用区的多个副本构成，以保证高可用性；通过 sidecars 实现的智能连接池可实时监测各个 Postgres 实例的实际承载能力，从而更高效地调度流量；最后由 control plane 负责全局健康管理，协调故障转移、模式变更和数据重分片等操作。

用户通过 data topology 管理数据分布，这是一份用 shard keys 定义的 JSON 配置，用来把逻辑表映射到物理分片上。这样的灵活性让团队可以先采用标准的非分片 Postgres 部署，只有在性能成为瓶颈时再迁移到分片模型。平台将这些扩展操作作为在线内置工作流来处理，因此诸如 schema 更新或版本升级等通常需要大量停机的操作，用户可以在不中断应用的情况下完成。

除了分片能力外，Neki 还集成了 PlanetScale 的成熟功能，例如 database Insights 、自动架构建议和分支工作流，旨在为那些已超出单机限制但又不想放弃标准 Postgres 体验或自行管理复杂应用级分片的组织提供平滑的迁移路径。

Neki 目前处于平台预览阶段。公司鼓励用户了解其架构和功能，但建议在此阶段不要在关键生产负载上运行，因为可能会有破坏性变更。感兴趣的用户可以登录 PlanetScale 账号选择加入预览，并查阅文档获取有关集群设置与配置的详细技术说明。 Neki is a new sharded Postgres offering from PlanetScale, designed to help teams overcome the limitations of scaling databases on single machines. Drawing on years of experience managing massive MySQL clusters, PlanetScale developed Neki to allow users to distribute their Postgres data across multiple machines while maintaining compatibility with standard Postgres drivers, ORMs, and connection strings. Because it uses real Postgres on every shard rather than a custom storage engine, it preserves full support for existing extensions and SQL functionality.

The core architecture of Neki relies on four distinct components. Neki routers manage incoming traffic, acting as a gateway that parses queries, builds distributed plans, and aggregates results. Each shard consists of a primary database with multiple replicas across different availability zones to ensure high availability. To handle traffic efficiently, Neki utilizes intelligent connection pooling via sidecars that monitor the actual capacity of each Postgres instance. Finally, a control plane manages the health of the entire system, coordinating tasks like failovers, schema changes, and data resharding.

Users manage the distribution of their data through a data topology, which is a JSON configuration that maps logical tables to physical shards using defined shard keys. This flexibility allows teams to start with a standard, non-sharded Postgres deployment and transition to a sharded model only when performance needs demand it. Because the platform treats these scaling operations as online, built-in workflows, users can perform tasks that would otherwise require significant downtime, such as schema updates or version upgrades, without taking their application offline.

In addition to its sharding capabilities, Neki provides users with access to established PlanetScale features like database Insights, automated schema recommendations, and branching workflows. This integrated approach is intended to provide a seamless transition for organizations that are outgrowing the constraints of a single machine but do not want to sacrifice the standard Postgres experience or manage the complexity of manual application-level sharding.

Neki is currently available in a platform preview, and while the company encourages users to explore its architecture and capabilities, they advise against running critical production workloads on it during this phase due to the likelihood of breaking changes. Prospective users are invited to sign into their PlanetScale accounts to opt into the preview and consult the documentation for further technical details on how to set up and configure their clusters.

The Cognition Team 推出了 SWE-2，这是他们迄今为止最先进的编码模型，旨在显著改善能力与成本之间的平衡。通过将强化学习扩展到万亿参数级别，SWE-2 在 FrontierCode 1.1 Main 基准测试中达到了 50.0% 的成功率。该性能使其接近当前的顶尖模型，同时成本比 Fable 5.1 等竞争选项低 64% 。该模型以 Kimi K3 为基础，经过大规模的后期训练以强化其代理式编码能力。

SWE-2 在解决问题的效率和工程判断力上都有提升。与前代 SWE-1.7 相比，新模型能更快聚焦代码库中的关键部分，减少过度探索和重复读取代码，从而用更少的步骤就能开始实现，大幅降低每个任务的平均成本。此外，SWE-2 在编写端到端测试方面更可靠，也更善于变通：它能利用现有数据绕过受阻路径，而不只是照搬静态指令。

SWE-2 的一项核心技术创新是在强化学习过程中采用 Pareto-informed cost penalties 。团队在一次训练中同时覆盖所有推理努力级别，并根据基础模型的成本 - 性能曲线斜率调整惩罚，从而成功优化出整个权衡前沿。这一方法保证了模型在处理简单任务时具有成本效益，同时在面对更复杂的挑战时保持高强度的复杂规划能力。为支持该策略，团队还引入了基于长度加权的奖励基线，以减小梯度方差、提升训练稳定性。

训练基础设施也进行了优化，优先提升吞吐量并尽量减少推理与训练之间的不匹配。团队整合了在线草稿模型训练，并采用诸如 FP8 kernels 的低精度、量化感知方法以高效管理内存。这些改进带来了更稳定的训练过程，使模型在不显著增加延迟的前提下保持高性能。数据质量同样被高度重视：通过扩展 RL environments 并实施递归飞轮机制，不断增强验证器以防止奖励操纵。

在可信性与对齐方面，团队对 SWE-2 在宣传、审查和对上下文依赖表述的脆弱性等方面进行了评估。模型表现良好，在政治敏感提示上的通过率达到 98%，且无论用户如何表述或使用何种语言，都未表现出统计学上显著的偏见或脆弱性。 SWE-2 已通过 Devin Desktop 应用、命令行工具和 Web 平台向用户提供，这是将顶尖编码代理变得更强大且更易获取的重要一步。 The Cognition Team has introduced SWE-2, their most advanced coding model to date, designed to significantly improve the balance between capability and cost. By scaling reinforcement learning to the multi-trillion-parameter regime, SWE-2 achieves a 50.0% success rate on the FrontierCode 1.1 Main benchmark. This performance places it near the current frontier models while maintaining a cost profile 64% cheaper than competing options like Fable 5.1. The model is built on the foundation of the Kimi K3 base model, which underwent extensive post-training to refine its agentic coding abilities.

The intelligence of SWE-2 is characterized by more efficient problem solving and improved engineering judgment. Compared to its predecessor, SWE-1.7, the new model reduces over-exploration and redundant code reads by focusing more quickly on the critical parts of a codebase. This allows SWE-2 to begin implementation in fewer steps, significantly lowering the average cost per task. Furthermore, the model demonstrates enhanced reliability in writing end-to-end tests and exhibits greater resourcefulness by navigating around blocked paths using available data, rather than simply relying on static instructions.

A central technical innovation in SWE-2 is the use of Pareto-informed cost penalties during the reinforcement learning process. By training all reasoning-effort levels in a single run with penalties tuned to the slope of the base model's cost-performance curve, the team successfully optimized the entire trade-off frontier. This principled approach ensures that the model provides cost-effective performance for simple tasks while maintaining high-effort, complex planning capabilities for more difficult challenges. To support this, they also implemented a length-weighted reward baseline that reduces gradient variance and promotes stability throughout the training phase.

The training infrastructure for SWE-2 was refined to prioritize throughput and minimize inference-training mismatches. The team integrated online draft-model training and utilized low-precision, quantization-aware methods, such as FP8 kernels, to manage memory usage efficiently. These improvements allowed for a more stable training process, resulting in a model that maintains high performance without significantly increasing latency. Data quality was similarly prioritized through the expansion of RL environments and the implementation of a recursive flywheel that continuously hardens verifiers against reward hacking.

Finally, the team evaluated SWE-2 on its trustworthiness and alignment, specifically regarding propaganda, censorship, and vulnerability to context-dependent framing. The model performed strongly, achieving a 98% pass rate on politically sensitive prompts while showing no statistically significant bias or vulnerability regardless of the user's framing or the language used. SWE-2 is currently available for users through the Devin Desktop application, command-line interface, and web-based platforms, marking a significant step forward in making frontier-level coding agents both more capable and accessible.

一项由 NASA 开发、最初用于分析 Martian 表面的创新成像技术，正成为考古学家研究地球上古老岩画的变革性工具。这种被称为 decorrelation stretch 的处理手段，最早由 Jet Propulsion Laboratory 的研究人员在 1978 年开发，旨在增强高度相关的卫星影像。与单纯提高对比度不同，该方法通过将原始颜色映射到更宽的范围，揭示出肉眼难以察觉的细微纹理和信息。

考古学家 Jon Harman 在看到火星图像后发现了这项技术的潜力，他意识到该算法可以改用于历史研究。 Harman 利用 Karhunen.Loève Transform（一种用于在数字信号处理中实现能量压缩最大化的统计工具）开发了 Dstretch 插件。凭借医学影像的背景和数学博士学位，他将 decorrelation stretch 方法针对不同色彩空间进行了调整，使其成为识别褪色或被覆盖的历史颜料的高效工具。

这项技术对考古学领域产生了深远影响。研究人员借助该插件在全球范围内识别出数百幅此前未被记录的图像：从在古老 Egyptian 墓地中显现出的动物图像，到揭示 Canada 战士留下的战术象形符号，软件为学者提供了关于古代文化习俗和被遗忘艺术传统的新见解。

随着越来越多的研究者使用 Dstretch，他们不断发掘出该软件的新用途。通过调整数字图像中的特定颜色参数，考古学家可以分离出不同的颜料层，如同剥开时间的层叠，重现被掩盖的历史叙事。这种太空探索技术与人类历史研究之间的意想不到的跨界融合，凸显了跨学科创新为我们共同过去带来的深刻新视角。 An innovative NASA imaging technique, originally designed to analyze the Martian surface, is proving to be a revolutionary tool for archaeologists studying ancient rock art on Earth. The process, known as decorrelation stretch, was first developed by researchers at the Jet Propulsion Laboratory in 1978 to enhance high-correlation satellite imagery. By mapping original colors to an expanded range rather than simply increasing contrast, the method reveals intricate details that are otherwise invisible to the naked eye.

Archaeologist Jon Harman, who discovered the potential of this technology after seeing images of the red planet, recognized that the algorithm could be adapted for historical research. By applying the Karhunen.Loève Transform, a statistical tool for maximizing energy compaction in digital signals, Harman created the Dstretch software plug-in. His background in medical imaging and his PhD in mathematics allowed him to tailor the decorrelation stretch method to various color spaces, making it a highly effective tool for identifying faded or obscured historical pigments.

The impact of this technology on the field of archaeology has been significant. Researchers have used the plug-in to identify hundreds of previously undocumented figures across the globe. From uncovering images of animals in ancient Egyptian cemeteries to revealing tactical pictographs left by warriors in Canada, the software has provided scholars with new insights into ancient cultural practices and forgotten artistic legacies.

As practitioners continue to use Dstretch, they are constantly finding new applications for the software. By adjusting specific color parameters within the digital images, archaeologists can isolate different pigments, effectively peeling back the layers of time to reveal hidden narratives. This unexpected crossover between space exploration technology and human history highlights how cross-disciplinary innovation can provide profound new perspectives on our collective past.

保存旧且常常神秘的科技线缆这种普遍困扰，出人意料地让人感同身受。很多人囤着多年未用的电线，抱着某天可能用得上的模糊希望不肯丢弃。 Tyler Gaw 的一句话为这种习惯提供了理由：当你终于用上一根保存了十多年的电缆时，就足以证明你从未扔掉它们是正确的。

这话打动了 Jim Nielsen，他决定把对这堆线缆的坚持正式化。他的盒子被妻子亲切地标注为 FAMILY TECHNO BOX，成了一个用来彰显这些"旧物"价值的新项目的核心。他把那条主张保留此类盒子的社交媒体帖子打印出来，用包装胶带贴在盒子正面，把这个存放容器变成了未雨绸缪的证据。

现在每次往箱里再塞进一根电缆，贴在盒子上的那段话都会提醒他：这些东西是有意保存的，而不是杂乱无章。它既带给他一丝快乐和动力，也能阻止家里人想把箱子清空的念头。

从长远看，这也是留给后代的一个"遗产"。他希望当孩子们某天在阁楼里翻出这些物品时，能理解这份收藏的意义——既俏皮又真诚地向那种能让小配件在过时后仍然派上用场的远见致敬。 The universal struggle of maintaining a collection of old, often mysterious technology cables is a surprisingly relatable experience. Many people keep a hoard of cords that haven't been used in years, held onto with the vague hope that they might one day be needed. This common habit recently gained validation through a sentiment shared by Tyler Gaw, who emphasized that the day you finally find a use for a cable you have kept for over a decade is the ultimate justification for never discarding your collection.

This advice struck a chord with Jim Nielsen, who decided to formalize his commitment to his own collection. His box, affectionately labeled by his wife as the FAMILY TECHNO BOX, became the focus of a new project meant to honor the utility of these saved items. By printing out the original social media post advocating for the preservation of these boxes, he transformed his storage container into a testament of preparedness.

The process involved printing the message and affixing it directly to the front of the box with packing tape. Now, every time he adds another cable to the pile, he is greeted by a reminder that his decision to keep them is purposeful rather than clutter-prone. It provides him with a sense of joy and motivation while simultaneously serving as a deterrent for any family members who might consider cleaning out the box.

Ultimately, this act serves as a legacy for the future. By preserving the box and its manifesto, he hopes that when his children eventually discover his belongings in the attic, they will understand the importance of the collection. It is a playful yet sincere nod to the foresight required to keep gadgets functional long after they have fallen out of fashion.

Shopify 宣布其移动工程战略发生重大转变：公司将放弃长期使用的 React Native，转而采用 Swift 与 Kotlin 进行原生开发。尽管 2020 年采用 React Native 在开发效率和功能一致性方面取得了显著成效，但编码代理与大型语言模型（LLM）的快速发展已从根本上改变移动开发的成本结构。由于 AI 现在能有效负责跨两个平台的功能翻译、实现与测试，以往需要"重复构建"的负担不再是决定性障碍。

此次回归原生源于对移动技术栈的原则性重新评估。尽管 React Native 仍是强大的框架，Shopify 认为原生开发能更好地利用平台特有能力与官方工具。借助 AI 代理承接大量编码工作，团队可以在不再承担过去高额维护成本的前提下，享受原生带来的性能与稳定性。

为推动这一转型，Shopify 对其核心应用（如 Shop 与主 Shopify app）采取 greenfield 路线。工程团队并非缓慢渐进迁移，而是借助名为 Helix 的专用系统，以结构化反馈循环推进：代码以小而可验证的检查点构建，每一段在继续之前都必须通过严格测试、视觉审查和对抗性代码分析。该流程确保最终代码库可维护、高质量且稳健。

迁移的关键环节之一是通过将业务逻辑与用户界面解耦来提升速度。构建可通过命令行以 headless 方式运行的架构，使 AI 代理能够在毫秒级而非分钟级完成测试与迭代，消除了传统模拟器交互的瓶颈，从而让代理自主且高效地工作。该架构既支持人类开发者也支持 AI，确保新的原生应用在长期内易于更新与维护。

在开源 React Native 库的支持上，Shopify 也在调整：React Native Skia 将过渡到社区主导，行业广泛使用的 FlashList 在 Shopify 寻找长期维护方期间仍由公司继续提供关键维护，Restyle 库则计划于 2026 年底归档。整个过渡过程将保持透明，以尽量减少对依赖这些工具的开发者社区的影响。 Shopify has announced a significant shift in its mobile engineering strategy, moving away from its long-standing commitment to React Native in favor of native development with Swift and Kotlin. While the company's 2020 decision to adopt React Native was highly successful in terms of developer efficiency and feature parity, the rapid advancement of coding agents and Large Language Models has fundamentally altered the economics of mobile development. Because AI can now effectively manage the translation, implementation, and testing of features across two separate platforms, the traditional burden of building twice is no longer a disqualifying factor.

The decision to migrate back to native stems from a first-principles reevaluation of the company's mobile tech stack. While React Native remains a strong framework, Shopify determined that native development offers superior access to platform-specific capabilities and first-party tooling. By utilizing AI agents to handle the heavy lifting of coding, the team can enjoy the benefits of native performance and stability without incurring the historical maintenance costs that once made building for each platform separately prohibitive.

To facilitate this transition, Shopify is implementing a greenfield approach for its major applications, such as Shop and the main Shopify app. Rather than performing a slow, gradual migration, the engineering team is utilizing a specialized system called Helix. This tool operates within a structured feedback loop where code is built in small, verifiable checkpoints. Each segment must pass rigorous testing, visual review, and adversarial code analysis before moving forward. This process ensures that the resulting codebase remains maintainable, high-quality, and robust.

A critical component of this migration involves optimizing for speed by decoupling business logic from the user interface. By building architectures that can run headlessly via a Command Line Interface, engineers allow AI agents to test and iterate on code in milliseconds rather than minutes. This strategy removes the bottleneck of traditional simulator interaction, enabling agents to work autonomously and efficiently. This architecture is designed to support both human developers and AI, ensuring that the new native apps remain easy to update and manage in the long term.

As part of this shift, Shopify is adjusting the support for its popular open-source React Native libraries. React Native Skia will be transitioned to a community-led model, while FlashList—a tool used widely across the industry—will see Shopify continue to provide critical maintenance while the company searches for a long-term steward. Meanwhile, the Restyle library will be archived at the end of 2026. This transition is being handled with transparency to ensure minimal disruption for the broader developer community that relies on these tools.