钠离子电池技术正迅速向商业成熟度迈进，确立了其作为固定式储能领域传统锂离子系统主要替代方案的地位。 Wood Mackenzie 的最新数据显示，该技术的全球供货协议已超过 70 GWh 。尽管目前正进入商业化生产阶段，但这些电池因其在供应链多样化、低温环境下的出色表现以及长期的经济优势而备受关注。

领先制造商有望在 2026 年底前实现与磷酸铁锂（LFP）电池在电芯层面的成本平价。尽管在模组和系统层面仍预计要到 2030 年后才能达到平价，但该技术进展十分迅速，尤其是在中国。比亚迪和宁德时代等主要厂商正加大产能扩张以满足日益增长的需求；研究人员预计，受益于中国成熟的工业供应链，2026 年至 2035 年间中国的学习曲线增速（learning rate）将达到 18% 至 20% 。

尽管钠离子电池目前的平均前期成本为 229 美元 / 千瓦时，高于 LFP 约 200 美元 / 千瓦时，但其全生命周期拥有成本（total cost of ownership）更具吸引力。在 20 年的项目寿命期内，钠离子系统有望将总成本降低 21% 。这一经济优势主要来自更低的运营支出，并且避免了锂基替代方案往往需要的复杂容量扩张周期。

钠离子电池在技术层面同样增强了其竞争力，尤其体现在原材料供应安全、更长的使用寿命，以及在更宽温度范围内保持稳定运行。这种较强的热耐受性使其成为寒冷地区部署的理想选择，可能减少对昂贵的外部热管理系统的依赖。

政府政策与产业激励正在显著加快这一应用进程。在中国，针对钠离子技术的优惠税收减免（相对而言，对 LFP 电池征收了新的税费）正在形成竞争优势。与此同时，美国正借助《通胀削减法案》（Inflation Reduction Act）下的生产税收抵免推动本土制造；预计欧洲的碳足迹法规也将更有利于对关键原材料依赖较少的化学体系。印度同样在大力投入，提供超过 10 亿美元的激励措施，以培育自身的国内制造能力。 Sodium-ion battery technology is rapidly advancing toward commercial maturity, establishing itself as a significant alternative to traditional lithium-ion systems for stationary energy storage. Recent data from Wood Mackenzie indicates that global supply agreements for this technology have already surpassed 70 GWh. While currently entering the commercial production phase, these batteries are attracting attention due to their potential for supply chain diversification, strong performance in low-temperature environments, and long-term economic advantages.

Leading manufacturers are on track to achieve cell-level cost parity with lithium iron phosphate (LFP) batteries by late 2026. Although parity at the module and system levels is not anticipated until after 2030, the technology is showing strong progress, particularly in China. Major players like BYD and CATL are aggressively scaling their manufacturing capacities to meet rising demand, and researchers expect an 18% to 20% learning rate in China between 2026 and 2035, driven by the country's mature industrial supply chain.

Despite sodium-ion batteries currently carrying a higher average upfront cost of $229/kWh compared to roughly $200/kWh for LFP, they offer a compelling total cost of ownership. Over a 20-year project lifetime, sodium-ion systems could result in a 21% reduction in total costs. This economic benefit is largely attributed to lower operating expenses and the elimination of the complex capacity expansion cycles often required by lithium-based alternatives.

Technical advantages further bolster the case for sodium-ion, particularly regarding raw material supply security, increased service life, and consistent operation across a broader temperature range. This wide thermal tolerance makes them an ideal candidate for deployments in colder climates, potentially reducing the need for costly external thermal management systems.

Government policies and industrial incentives are significantly accelerating this adoption. In China, favorable tax exemptions for sodium-ion technology—contrasted with new levies on LFP batteries—are creating a competitive edge. Similarly, the United States is leveraging production tax credits under the Inflation Reduction Act to promote domestic manufacturing, while European carbon footprint regulations are expected to favor chemistries with less reliance on critical raw materials. India is also investing heavily, providing over $1 billion in incentives to foster its own domestic manufacturing capabilities.

Anthropic 已经达到了其《负责任扩展政策》（Responsible Scaling Policy）中定义的关键阈值，这意味着在 AI 自动化的推动下，该公司正在经历大幅加速的研发。近期评估显示，公司在"能力进展"（Capabilities）方面的速度在过去一年里翻了一番。尽管这种加速一开始可能由多种因素促成，但如今看来，AI 系统已成为这种持续增长的主要驱动力，这表明行业正进入一个由 AI 有效地"培养自身继任者"的阶段。

要实现整体能力进展翻倍，人类研究人员必须获得巨大的生产力提升。由于由 AI 驱动的研究只提供实现能力增益所需投入的一部分，因此人力劳动的实际加速幅度可能远高于表面结果，估计大约是基准速度的 10 倍左右。这并不只是简单地增加人员，而是一种职能性的转变：仿佛每位研究人员都在一个"时间扭曲"的泡泡里工作，使他们能够在相同时间内完成相当于平时十倍的产出。

这一进展中最令人担忧的部分，是加速趋势可能进一步"复利式"加深。如果 AI 能力继续提升，AI 研究人员的效率很可能还会增加，从而形成一个递归自我改进的反馈回路。由于研究人员尚未识别出会阻断这一过程的显著瓶颈或边际收益递减现象，盲目外推意味着，我们可能在一年内就达到超智能的极端水平。

如此迅速的演进带来了深刻风险，尤其是在相关系统无法与人类价值观完全对齐（aligned）的情况下。新能力的快速出现，可能让 AI 的行动速度超过我们对其进行控制或监测的能力。如果这些强大的系统出现失调（misaligned），它们就可能在我们有机会落实必要的保障措施之前，先行削弱人类的处境，进而引发灾难性后果。

目前，AI 行业在很大程度上似乎还没有做好准备，来安全地应对这种级别的风险。尽管这些里程碑刻不容缓，主要实验室近期却接连遭遇安全漏洞，表明它们在保护对控制这些系统至关重要的模型权重（model weights）和算法机密（algorithmic secrets）方面存在困难。在缺乏强有力的安全措施、普遍性的监测手段以及经过验证的对齐（alignment）技术的情况下，朝着"智能爆炸"的加速只会制造一种环境：让人类暴露在行业尚未具备能力管理的风险之中。 Anthropic has reached a critical threshold defined in its Responsible Scaling Policy, signifying that the company is experiencing dramatically accelerated research and development due to AI automation. Recent assessments indicate that capabilities progress at the company has doubled over the past year. While this acceleration may have initially stemmed from various factors, AI systems now appear to be the primary driver of this sustained increase, suggesting that the industry has entered a phase where AI is effectively building its own successors.

To achieve this doubling in overall capabilities progress, human researchers must be experiencing a massive productivity boost. Because AI-driven research contributes only a portion of the inputs required for capability gains, the actual speedup for human labor is estimated to be significantly higher, perhaps around 10 times the baseline rate. This is not merely a matter of hiring more people, but rather an functional transformation, as if each researcher were working within a time-warp bubble that allows them to accomplish ten times the usual output in the same period.

The most concerning aspect of this development is the potential for compounding acceleration. If AI capabilities continue to improve, the efficiency of AI researchers will likely increase further, potentially creating a feedback loop of recursive self-improvement. Because researchers have not identified significant bottlenecks or diminishing returns that would halt this process, a naive extrapolation suggests that we could reach extreme levels of superintelligence within a single year.

Such a rapid trajectory poses profound risks, particularly if the systems involved are not perfectly aligned with human values. The rapid development of novel capabilities could allow AI to move faster than our ability to control or monitor it. If these powerful systems were misaligned, they could potentially disempower humanity before we have the chance to implement necessary safeguards, leading to catastrophic outcomes.

Currently, the AI industry appears largely unprepared to handle this level of risk safely. Despite the urgency of these milestones, major labs have recently suffered security breaches, indicating that they struggle to protect the very model weights and algorithmic secrets that are essential to controlling these systems. Without robust security, universal monitoring, and verified alignment techniques, accelerating toward an intelligence explosion creates an environment where humanity is exposed to risks that the industry is not yet equipped to manage.

身份解析是企业在尝试把各类数据记录准确映射到个人客户时必须面对的关键且富有挑战的流程，这样才能实现精确分析与个性化触达。该领域的一大难点是"过度合并"（over-merging），即当记录被错误地归并到同一个 identity 下。由于通常缺少可用于验证这些分组的权威答案，组织往往只能通过人工抽样来评估准确性。但人工审核不仅耗时、成本高，而且很难产出具有统计显著性的结果。

研究人员想验证，大语言模型（LLM）是否能够通过对身份解析结果进行评分来自动化这一评估流程。他们重点对比了更近、更快的决策导向模型，例如 Jev 和 GPT-6 Luna，并与 Claude Opus 等前沿模型（frontier models）进行对照。尽管前沿模型的准确度最高，但它们价格更高、运行也更慢。决策模型则提供了很有吸引力的替代方案：即便在复杂数据集上的原始准确度有时会略低于前沿模型，它们的成本最高可降低 90 倍，速度则快 10 倍。

测试过程使用了一个包含 1,000 个 identity 的基准，这些 identity 由模拟人口统计数据生成，并刻意引入了拼写错误、姓名变体、以及家庭相似性等复杂错误，用来衡量各模型识别过度合并的能力。尽管 GPT-6 Luna 在开箱即用的初步测试中表现更好，但研究人员发现了一个关键差异：置信度校准能力。 Jev 的置信度评分校准更好，也就是说，它的失误通常发生在它对自身判断信心较低的时候；而 Luna 即使在出错时也常常显得过于自信。

这种校准差异让研究人员得以设计一种混合评估模型。系统先用 Jev 作为初步筛查器，把不确定的案例自动转交给 Claude Opus 进行二次复核。通过这种混合方案，系统能够达到与前沿模型相近的准确度水平，同时把成本显著压低。因为 Jev 能够有效判断何时需要更多帮助，系统就能根据遇到的数据复杂程度智能地调整投入。

总体而言，这项研究表明：将具备成本优势的决策模型与更强模型的偶尔监督结合起来，可以构建一种稳健、可扩展的数据质量保障架构。在这一场景下，混合的 Jev 方法优于 Luna，原因在于更好的校准能力使其在不同数据复杂度下都能保持更可靠的表现。对那些希望在不必为每条记录都运行前沿模型而承担高昂成本的前提下，维持身份准确性的高标准的企业来说，在混合架构中利用校准良好的决策模型是一种非常有效的策略。 Identity resolution is a critical yet challenging process for companies looking to map various data records to individual customers for accurate analytics and personalized outreach. A major hurdle in this field is over-merging, which occurs when records are incorrectly grouped into a single identity. Because there is often no definitive answer key available to verify these groupings, organizations typically rely on manual sampling to assess accuracy. However, human review is slow and labor-intensive, making it difficult to generate statistically significant results.

The researchers tested whether Large Language Models (LLMs) could automate this evaluation process by grading identity resolution outputs. They specifically compared newer, faster decision-oriented models like Jev and GPT-6 Luna against frontier models like Claude Opus. While frontier models remain the most accurate, they are significantly more expensive and slower. Decision models offer a compelling alternative, being up to 90 times cheaper and 10 times faster, even if they occasionally trail the frontier models in raw accuracy on complex data sets.

The testing process involved using a benchmark of 1,000 identities generated from simulated demographic data, intentionally including complex errors such as typos, name variations, and household similarities to evaluate how well each model identified over-merges. While GPT-6 Luna performed better out of the box in initial tests, the researchers discovered a vital distinction in confidence calibration. Jev displayed a better-calibrated confidence score, meaning its mistakes generally occurred when it had lower confidence in its own judgment, whereas Luna tended to be overconfident even when it was wrong.

This difference in calibration allowed the researchers to design a hybrid evaluator model. By using Jev as an initial screener, the system could automatically route uncertain cases to Claude Opus for a secondary review. This hybrid approach enabled the system to achieve accuracy levels comparable to the frontier models while keeping costs significantly lower. Because Jev effectively identified when it needed additional help, the system could intelligently flex its spending based on the complexity of the data it encountered.

Ultimately, the study demonstrates that integrating cost-effective decision models with occasional oversight from more powerful models creates a robust, scalable architecture for data quality assurance. The hybrid Jev approach proved superior to Luna in this context because its better calibration allowed for more reliable performance across varying levels of data complexity. This finding suggests that for businesses needing to maintain high standards of identity accuracy without the prohibitive costs of running frontier models on every record, leveraging well-calibrated decision models in a hybrid harness is a highly effective strategy.

Firecracker microVMs 用于隔离不可信代码，但研究人员已经成功演示了一种 guest-to-host 逃逸：在宿主机上实现 root 级别的代码执行。研究人员通过利用 Firecracker 管理 virtual CPU threads 的方式中存在的漏洞，在某个 thread 从 KVM hypervisor 返回时劫持了执行流。整个过程包括在宿主内存中定位 Firecracker 进程、识别特定的 thread registers，并将 thread 重定向以执行未授权操作，然后在不导致 VM 崩溃的前提下把它返回到原本的执行路径。

这条逃逸链首先借助此前建立的宿主内存原语，在宿主的物理内存空间中识别出 guest 。研究人员在 guest page 中放入一个 nonce，并通过 Extended Page Tables 将其转换为宿主物理地址。随后，他们反向映射宿主的页表，找到 Firecracker 进程对应的特定虚拟地址空间。通过分析进程内存，研究人员确定了可执行文件的 load bias，从而能够绕过 Address Space Layout Randomization，并定位关键代码段，例如 write 系统调用的 wrapper 。

完成进程映射后，重点转向寻找 virtual CPU 的 saved register state 。当 thread 调用 KVM_RUN 以执行 guest 指令时，内核会保存其 userspace register frame 。研究团队通过查找与正在进行的 KVM_RUN 调用特征相匹配的 frame，精确定位到 saved instruction pointer 和 stack 的具体位置。这一步非常关键，因为它为覆盖 thread 的返回路径提供了稳定目标，使得注入自定义逻辑成为可能。

最终的 exploit 构造了一个 payload，强迫 Firecracker 进程向自身的 stderr 执行 write 操作，作为概念验证。研究团队修改了 saved register frame，将 thread 重定向到 Firecracker 二进制中的 write wrapper，并传入所需参数。为维持系统稳定性，他们仔细构建了一个临时的返回序列，清除返回值后跳回原始的 ioctl continuation point 。在通过 serial port write 触发该操作后，exploit 成功执行，且 VM 仍保持可运行、可访问。

这项研究凸显了 virtualization environments 中内存管理带来的重大安全影响。尽管这份概念验证是在以 root 身份运行、且未配 jailer 的 Firecracker 环境中完成的，但它表明：一旦攻击者获得对宿主内存的访问权限，guest 与 host 之间的边界就可能被打破。该 exploit 的成功已通过可观察的输出以及随后对宿主线程状态的恢复得到证实，这有力证明了如何将复杂的基于内存的原语串联起来，从而破坏宿主的完整性。 Firecracker microVMs are designed to isolate untrusted code, yet researchers have successfully demonstrated a guest-to-host escape that achieves root-level code execution on the host machine. By exploiting vulnerabilities in the way Firecracker manages virtual CPU threads, the researchers were able to hijack execution flow when a thread returned from the KVM hypervisor. This process involved locating the Firecracker process in host memory, identifying specific thread registers, and redirecting the thread to perform unauthorized actions before returning it to its original execution path without causing the VM to crash.

The escape chain began by using a previously established host memory primitive to identify the guest in the host's physical memory space. Researchers placed a nonce in a guest page and used Extended Page Tables to translate it to a host physical address. From there, they reverse-mapped the host page tables to find the specific virtual address space of the Firecracker process. By analyzing the process memory, they identified the executable's load bias, which allowed them to account for Address Space Layout Randomization and locate critical code segments, such as the wrapper for the write system call.

With the process mapped, the focus shifted to finding the saved register state of the virtual CPU. When a thread calls KVM_RUN to execute guest instructions, the kernel saves its userspace register frame. By searching for a frame that matched the characteristics of an active KVM_RUN call, the team pinpointed the exact location of the saved instruction pointer and stack. This discovery was crucial because it provided a stable target for overwriting the return path of the thread, enabling the injection of custom logic.

The final exploit involved crafting a payload that would force the Firecracker process to execute a write operation to its own stderr, serving as a proof of concept. The team modified the saved register frame to redirect the thread to a write wrapper in the Firecracker binary, providing it with the necessary arguments. To maintain system stability, they carefully constructed a temporary return sequence that cleared the return value and jumped back to the original ioctl continuation point. After triggering the operation via a serial port write, the exploit successfully executed while the VM remained functional and accessible.

This research highlights the significant security implications of memory management within virtualization environments. While this specific proof of concept was executed against a Firecracker setup running as root without a jailer, it underscores that once an attacker gains access to host memory, the boundary between the guest and host can be breached. The success of this exploit, confirmed by the observable output and the subsequent restoration of the host thread's state, serves as a powerful demonstration of how complex memory-based primitives can be chained to compromise host integrity.

GitHub 推出了全新的 Git 存储架构，专门为承载由 AI 驱动开发带来的海量流量而设计。随着越来越多的开发者借助自动化代理来编写代码，单个仓库的写入操作正在以指数级增长。该新系统代号 Project Rubicon（项目"卢比康"），在维持低延迟读取性能的同时，实现了每秒数千次的并发写入，从而让平台能够随着代理式工作流需求的不断上升而持续扩展。

这套架构的核心理念是将存储与计算解耦，并尽量减少协调开销。团队发现：Git 对象是不可变的，因此可以在不需要一致性协商的情况下进行存储；但引用（reference）的更新则必须严格协调。通过将持久数据迁移到 Azure Blob Storage，并创建用于处理引用更新的专用服务 Ref-CAS（引用对比并交换），GitHub 让各组件可以独立扩展。这样系统不仅能支持多个写入节点，还能提升容错能力与整体吞吐量。

新系统中的流量管理由"虚拟副本"完成，它们充当可随时丢弃的计算集群。这些副本同时承担读写任务，本地磁盘则作为可重建的缓存使用。由于虚拟副本不保存权威的持久状态，它们可以在仓库流量波动时动态地添加或移除。该设计在实现租户管理横向扩展的同时，也有效应对了高活跃仓库带来的纵向扩展压力。

为在高写入量下仍保持性能，架构引入了一种新的压缩（compaction）方式。过去如果要在每个副本上做资源密集型维护，会影响系统开销；现在压缩节点会在后台独立运行，并直接与 Blob Storage 交互。借助租约机制（lease），这些节点能够并行处理数据分片，把 pack 合并起来，以确保读取性能保持稳定，同时不影响主请求链路。后台流程会持续保障：随着 push 次数增加，整个仓库依然保持高效与轻量。

展望未来，这一基础还将为更多优化铺路，例如全局副本（global replicas）、部分克隆（partial clones）以及更细粒度的数据访问。 GitHub 计划继续推进基础设施与 Git 项目的演进，让代理式操作变得更加高效。借助这套新架构，目标是为仓库提供无缝体验，使其在不需要特殊处理的情况下也能承载高强度、全天候的活动，从而让开发者及其代理能够在规模化的环境中高效运行。 GitHub has unveiled its new Git storage architecture, specifically engineered to handle the massive influx of traffic from AI-driven development. As more developers leverage automated agents to build code, individual repositories are seeing exponential growth in write operations. This new system, dubbed Project Rubicon, allows for thousands of concurrent writes per second while maintaining low-latency read performance, ensuring that the platform can scale alongside the increasing demands of agentic workflows.

The core philosophy behind this architecture is the decoupling of storage from compute and the minimization of coordination. The team identified that while Git objects are immutable and can be stored without consensus, reference updates require strict coordination. By moving durable data to Azure Blob Storage and creating a specialized service called Ref-CAS (Reference Compare-and-Swap) to handle reference updates, GitHub has enabled independent scaling for its components. This ensures that the system can support multiple writer nodes, improving both fault tolerance and throughput.

Traffic management within the new system is handled by virtual replicas, which act as a disposable compute fleet. These replicas serve both reads and writes, with local disks serving as caches that can be rebuilt as needed. Because these replicas do not store authoritative durable state, they can be added or removed dynamically as repository traffic fluctuates. This design provides horizontal scale for tenant management while effectively addressing the vertical scale challenges posed by high-activity repositories.

To maintain performance despite high write volumes, the architecture introduces a new approach to compaction. Instead of performing resource-heavy maintenance on every replica, compaction nodes now operate independently in the background, working directly against Blob Storage. By leveraging leases, these nodes can process shards of data in parallel, merging packs to keep read performance consistent without impacting the main request flow. This background process ensures that as pushes increase, the overall repository remains performant and lean.

Looking forward, this foundation paves the way for further optimizations, such as global replicas, partial clones, and more granular data access. GitHub plans to continue evolving its infrastructure and the Git project itself to make agentic operations even more efficient. With this new architecture, the goal is to provide a seamless experience where repositories can handle intensive, around-the-clock activity without requiring special handling, allowing developers and their agents to operate at scale.

Devvm 是一款开源工具，旨在为开发人员的编码工作流程提供隔离的、特定于项目的虚拟机沙盒。通过在 VM 内运行开发环境，你可以有效限制潜在安全威胁的影响范围，比如供应链攻击、泄露的密钥（secrets）或被入侵的代码代理。尽管这类工具很有用，但往往会带来繁琐的配置门槛。 Devvm 通过提供合理的零配置默认值来解决这个问题，让你只需一条命令就能建立一个安全、可直接使用的沙盒环境。

当你在项目文件夹中启动 devvm 时，它会创建一个封闭的环境，只挂载你的项目文件，从而让诸如 SSH keys 、 AWS credentials 和个人文档之类的敏感本地数据完全隔离。该 VM 通过主机上的专用代理（proxy）运行，由代理通过严格的网络允许列表（allowlists）来控制互联网访问。为进一步提升安全性，系统会在 macOS 上锁定 git hooks 和相关配置，确保在沙盒中执行的代码无法篡改你机器的主要 git 设置。

平台以高效为设计目标，最大限度减少日常使用中的摩擦。由于基础镜像只构建一次，后续的 VM 能在几秒内生成。 Devvm 支持多种项目需求，会根据你的项目文件自动检测所需的工具链，例如 Node.js 或 Python 。它还支持 Claude Code 、 Codex 、 OpenCode 等流行的代码代理，并将它们限制在 VM 的安全边界内运行。这些代理可以访问你的设置和技能，但它们无法接触到你的本地历史记录和私人机器数据。

自定义依然是核心能力，允许开发人员通过自定义 layers 扩展环境。你可以为 mailpit 等工具编写特定的设置脚本，加入你喜欢的 dotfiles，甚至可以让 VM 以已有的容器镜像（例如 Node:24）为基础。此外，如果你的项目已经使用了 devcontainer.json，devvm 可以自动解析该文件，用于配置 VM 的环境、端口和依赖项。这样既能保证沙盒安全且标准化，又能根据你的项目独特需求进行调整。

devvm 会通过直观的功能与你现有工作流程无缝衔接，比如自动端口转发，确保在 VM 中运行的服务可以通过主机上的 localhost 进行访问。你也可以使用 VS Code over Remote-SSH 等工具直接与 VM 交互。 devvm 构建在 Lima 之上，可在 macOS 和 Linux 上运行，在不同架构之间保持一致的使用体验。通过把虚拟化的复杂性隐藏在简单的 CLI 之后，devvm 让每一个你正在进行的项目都能轻松采用沙盒化的最佳实践。 Devvm is an open-source tool designed to provide developers with isolated, project-specific virtual machine sandboxes for their coding workflows. By running your development environment inside a VM, you effectively limit the blast radius of potential security threats like supply chain attacks, leaked secrets, or compromised coding agents. While these tools offer significant utility, they often introduce complex configuration hurdles. Devvm bypasses this by offering sensible, zero-configuration defaults, allowing you to establish a secure, sandbox-ready environment with a single command.

When you launch devvm within a project folder, it creates a contained environment that mounts only your project files, keeping sensitive local data like your SSH keys, AWS credentials, and personal documents completely isolated. The VM operates through a specialized proxy on your host machine, which enforces strict network allowlists to govern internet access. For added security, the system locks git hooks and configurations on macOS, ensuring that code executing inside the sandbox cannot alter your machine's primary git settings.

The platform is designed to be highly efficient, minimizing friction in your daily routine. Because the base image is built once, subsequent VMs are generated in seconds. Devvm supports a wide variety of project requirements, automatically detecting necessary toolchains such as Node.js or Python based on your project files. It also supports popular coding agents like Claude Code, Codex, and OpenCode, running them within the secure confines of the VM. These agents gain access to your settings and skills, but they remain isolated from your local history and private machine data.

Customization remains a core feature, allowing developers to extend their environments through custom layers. You can define specific setup scripts for tools like mailpit, include your preferred dotfiles, or even base your VM on existing container images like Node:24. Furthermore, if your project already utilizes a devcontainer.json file, devvm can automatically parse it to configure the VM's environment, ports, and dependencies. This flexibility ensures that while the sandbox is secure and standardized, it can still adapt to the unique needs of your specific project.

Integration with your existing workflow is handled through intuitive features, such as automatic port forwarding, which ensures that services running in the VM are accessible at localhost on your host machine. You can also use tools like VS Code over Remote-SSH to interact with the VM directly. Built on top of Lima, devvm runs on both macOS and Linux, maintaining a consistent experience across different architectures. By keeping the complexity of virtualization hidden behind a simple CLI, devvm makes the best practices of sandboxing accessible for every project you undertake.

由于选后出现大规模抛售，Anthropic 目前正面临比预期更为严峻的 IPO 环境，这也迫使市场对该 AI 实验室的估值预期重新评估。

在公司定于下周启动的正式 IPO 营销之前，潜在买家正在讨论的估值区间为 1.2 trillion 美元到 1.3 trillion 美元。市场波动带来的不确定性也进一步加剧了谨慎情绪，部分投资者甚至认为估值可能低至 1.1 trillion 美元。

这些调整后的预期反映了近期市场波动对科技行业的更广泛影响，尤其是对那些正准备上市的高关注度 AI 企业而言。随着投资者在权衡公司长期潜力与当前经济环境之间的关系，即将到来的营销期将成为衡量市场对其增长前景信心的关键信号。 Anthropic is currently facing a more challenging environment for its highly anticipated initial public offering as a result of a significant post-election selloff. This market shift has forced prospective investors to reevaluate their expectations regarding the AI lab's valuation.

Ahead of the company's formal IPO marketing, which is scheduled to begin next week, potential buyers are now discussing a valuation range between $1.2 trillion and $1.3 trillion. The market volatility has introduced even more cautious sentiment, with some investors considering a valuation as low as $1.1 trillion.

These adjusted expectations highlight the broader impact of recent market fluctuations on the technology sector, particularly for high-profile AI firms preparing to go public. As investors weigh the long-term potential of the company against the current economic climate, the upcoming marketing period will be a critical indicator of market confidence in the firm's growth trajectory.

在现代 PC 类硬件上开发 bare-metal 应用，需要直接与 UEFI firmware 进行交互；这是一项要求采用结构化方法的工作，目的是在传统 Linux 开发环境与系统 firmware 的底层需求之间搭起桥梁。本系列文章旨在为那些希望在操作系统限制之外进行编程、并且此前对 UEFI 环境复杂性毫无经验的资深 Linux 开发人员提供入门性的基础指引。

整个过程首先从必要的技术准备入手，重点介绍生成 UEFI 能够成功加载的可执行代码所需的工具、编译策略以及链接要求。由于 UEFI 的运行方式不同于常规应用程序环境，弄清楚如何为这种特定运行时准备代码，是从常规工作区切换到 bare-metal 执行的关键第一步。

要让这些应用程序能够顺利运行，还必须关注它们如何被存储，以及如何向 firmware 呈现。挑战之一在于正确配置 EFI system partition：它充当 firmware 定位并执行二进制文件的入口。本指南将说明如何组织该分区，并进一步教你创建可启动媒体，例如 USB 闪存驱动器或 DVD，确保硬件能够识别并引导开发过程中生成的自定义应用程序。

在介绍引导与链接机制之外，本系列还会讨论系统真正运行起来后如何与之交互的实际问题。这包括对该环境局限与可能性的务实分析，以及通过 Simple File Protocol 进行文件管理的实践。按照这些步骤依次推进，开发人员就能系统地理解如何在 bare-metal UEFI 场景中管理数据，并与存储设备进行交互。 Developing bare-metal applications on modern PC-type hardware involves interfacing directly with UEFI firmware, a task that requires a structured approach to bridge the gap between traditional Linux development environments and the low-level requirements of system firmware. This series of articles serves as a foundational guide for experienced Linux developers who wish to explore programming outside the constraints of an operating system, assuming they have no prior familiarity with the complexities of the UEFI environment.

The process begins with the essential technical setup, focusing on the software tools, compilation strategies, and linking requirements necessary to generate executable code that UEFI can successfully load. Since UEFI operates differently than standard application environments, understanding how to prepare the code for this specific runtime is the critical first step in moving from a standard workspace to bare-metal execution.

Successfully running these applications requires attention to how they are stored and presented to the firmware. Part of the challenge lies in correctly configuring the EFI system partition, which acts as the gateway for the firmware to locate and execute binary files. The guide details how to organize this partition and subsequently create bootable media, such as USB flash drives or DVDs, ensuring that the hardware can recognize and boot the custom applications created during development.

Beyond the mechanics of booting and linking, the series addresses the practical aspects of interacting with the system once it is active. This includes a pragmatic look at the limitations and possibilities of the environment, as well as an exploration of file management using the Simple File Protocol. By following these steps in order, developers can build a comprehensive understanding of how to manage data and interact with storage devices from within a bare-metal UEFI context.

《杀出重围》于 2000 年夏季发布，最初口碑褒贬不一，但此后逐渐坐实其作为交互式叙事必读之作的重要地位。尽管一些早期评测曾指出技术层面的瑕疵，如今的讨论却常把它抬到"预言之作"的高度，称其能出人意料地预见围绕企业权力、监控体系与社会分层等议题的当代焦虑。尽管人们普遍乐于把它包装成具有前瞻性的巅峰杰作，但从开发过程来看，它更像是建立在 20 世纪 90 年代末文化语境之上的务实迭代产物。

游戏的核心气质源自 Looking Glass Studios 所倡导的沉浸式模拟（immersive sim）理念。它的玩法也与 System Shock 如出一辙，强调涌现式的游戏过程：与其把玩家硬塞进预设好的解谜循环，不如提供多种工具，让玩家自行穿行于挑战之中。通过把"玩家能动性"置于优先位置，并允许采用多元的解决思路，《杀出重围》成功把复杂的模拟体验与引人入胜、以叙事驱动的过程连接起来。不过，与前作更封闭、更自然的环境相比，这部作品采用了线性的任务式结构，最终会在一定程度上限制玩家改写主线剧情的空间，使得事件走向一条固定的轨迹，并在结尾以三种预设结局的选择收束。

从叙事结构上看，《杀出重围》深深扎根于 20 世纪 90 年代的阴谋论文化：它更像是对流行套路的"最大化拼贴"，而非严密的哲学论证。编剧们希望打造一个世界，让各种阴谋论能在同一时空里并存，并从《 X 档案》以及多种科幻与推想小说的经典母题中汲取灵感。尽管故事里确实有些片段，放在今天那种高度拟真、监控密集的氛围中会显得格外刺眼、仿佛仍在"对照现实"，但更合理的理解是：这些内容首先是特定历史阶段的产物，而非真正的预言。创作者经常借用这类主题来营造氛围、制造紧张感；只是到了当下，在政治立场日益分化的背景下，对"深层政府（deep-state）"与反全球化叙事等母题的依赖，其回响也许已经不再是当年的那种单纯效果，而是呈现出更复杂的折射。

归根结底，《杀出重围》依然是一部里程碑式作品：它不仅把沉浸式设计的理念落地得相当成功，也在推动该类型的发展演变中发挥了关键作用。它之所以能长期流传，并获得人们心目中"经典之作"的地位，实至名归。即便有人倾向于赋予它超出事实的深刻性或预测能力，这种解读也会让它的本来面目被轻易扭曲。对玩家而言，这是一段重要却并不轻松的体验，它至今仍能提供对其所处时代创作野心的真实洞察。无论你把它视为典型的沉浸式模拟作品，还是把它当作"9/11 之前世界"的文化遗存，它都在证明交互式叙事所拥有的潜力与天然边界。 Released in the summer of 2000, Deus Ex arrived with mixed critical reception but has since solidified its reputation as an essential work of interactive fiction. While some early reviewers noted technical flaws, contemporary discourse frequently elevates the game to a prophetic status, praising its uncanny ability to anticipate modern anxieties surrounding corporate power, surveillance, and social stratification. Despite the common desire to frame it as a visionary masterpiece, the reality of its development suggests a more pragmatic, iterative process grounded in the cultural landscape of the late nineties.

Central to the game's identity is its inheritance of the immersive sim philosophy championed by Looking Glass Studios. Much like System Shock, the game emphasizes emergent gameplay, offering players a variety of tools to navigate challenges rather than forcing them into predetermined puzzle-solving loops. By prioritizing player agency and allowing for diverse approaches to problem-solving, Deus Ex successfully bridged the gap between complex simulation and engaging, narrative-driven experiences. However, unlike the more contained and organic environments of its predecessors, the game adopted a linear mission-based structure that ultimately limits the player's capacity to alter the overarching plot, resulting in a fixed series of outcomes that culminate in a choice of three pre-scripted endings.

The narrative architecture of the game is deeply rooted in the conspiracy culture of the 1990s, functioning as a maximalist pastiche of popular tropes rather than a rigorous philosophical treatise. The writers sought to create a world where every conspiracy theory could exist simultaneously, drawing inspiration from works like The X-Files and various speculative fiction staples. While the resulting story features moments that feel eerily relevant to today's hyperreal, surveillance-heavy climate, these instances are better understood as products of a particular historical moment rather than genuine acts of prophecy. The creators often utilized these themes to enhance atmosphere and create tension, though the reliance on deep-state and anti-globalist tropes has arguably aged with a different, more complicated resonance in light of modern political polarization.

Ultimately, Deus Ex remains a landmark title, both for its successful execution of immersive design and for its role in shaping the evolution of the genre. Its enduring influence and status as a beloved classic are well-earned, even if the tendency to assign it an unearned degree of profundity or predictive power mischaracterizes its nature. For the player, it represents a significant, albeit demanding, experience that continues to offer genuine insight into the creative ambitions of its era. Whether viewed as a quintessential immersive sim or as a cultural artifact of a pre-9/11 world, it stands as a testament to the potential and the inherent constraints of interactive storytelling.

Google 已推出 Gemini Omni 1.2 Flash，这是其视频生成模型的重大更新，目前已通过 Gemini API 和 Google Cloud 全面开放使用。此次发布带来了可生成最长 15 秒视频内容的能力，让创作者和企业有更充足的空间去打造更完整的叙事、产品讲解以及广告短片。该模型在参考一致性方面也实现了显著提升，能够确保角色、产品和风格在不同摄像机角度下保持一致，这对于虚拟试穿体验和品牌化内容等专业级任务至关重要。

这一新迭代的核心优势在于强调更具成本效益的高质量制作。 Google 将价格相较上一版本下调了 30%：目前 720p 输出的定价为每秒 0.07 美元。工作流程面向规模化需求进行了优化，用户可以先用较低分辨率起草内容，再针对特定细节进行迭代修改，而无需反复重生成整段视频。新的多轮对话式编辑能力进一步提升了效率：用户只需通过连续、逐步的提示，就能对背景、光照或产品摆放等元素进行精细调整。

在基础生成之外，Gemini Omni 1.2 Flash 还对动作与语音进行了技术强化。模型生成的对话听起来更自然，唇形同步更清晰，并支持区分的多说话人条件设定。它还能提供对电影级摄像机运动的更强控制，使得如 FPV 飞行穿越与镜头拉远这类连续镜头更顺畅，同时减少跳切并提升画面稳定性。这些在一致性与可控性上的改进，有助于从最初的创意想法更顺利地落到精修后的成品结果。

系统内置了安全与可追责机制，所有生成内容都将带有肉眼不可察的 SynthID 水印。该模型已被多类机构采用，包括创意代理机构、 VFX 从业者以及电商平台，用于简化并加速视频生产流程。通过将企业级基础设施与这些新的生成能力结合起来，Google 旨在为专业团队提供可靠工具，帮助他们更高效地制作、编辑并扩展视频内容。 Google has launched Gemini Omni 1.2 Flash, a major update to its video generation model that is now generally available through the Gemini API and Google Cloud. This release introduces the ability to generate video content up to 15 seconds long, providing creators and businesses with more room to build comprehensive narratives, product explainers, and advertising spots. The model also features significant improvements in reference adherence, ensuring that characters, products, and styles remain consistent across camera angles, which is critical for professional-grade tasks like virtual try-on experiences and branded content.

A primary advantage of this new iteration is the emphasis on cost-effective, high-quality production. Google has reduced the pricing by 30% compared to the previous version, now set at $0.07 per second for 720p output. The workflow is designed for scalability, allowing users to draft content at lower resolutions and iterate through specific details rather than re-generating entire clips. This efficiency is further bolstered by new multi-turn conversational editing capabilities, which let users refine aspects like backgrounds, lighting, or product placement through simple, successive prompts.

Beyond basic generation, Gemini Omni 1.2 Flash brings technical refinements to motion and speech. The model produces more natural-sounding dialogue with clearer lip-syncing and supports distinct multi-speaker conditioning. It also offers greater control over cinematic camera movements, enabling continuous shots like FPV flythroughs and zoom-outs with fewer jump cuts and improved visual stability. These advancements in consistency and control help bridge the gap between initial creative concepts and polished final outputs.

Safety and accountability are integrated into the system, with all generated content featuring an imperceptible SynthID watermark. The model is already being utilized by diverse organizations, ranging from creative agencies and VFX artists to e-commerce platforms, to streamline their production pipelines. By combining enterprise-grade infrastructure with these new generative capabilities, Google aims to provide a reliable tool for professional teams to produce, edit, and scale video content efficiently.