zerostack 是用 Rust 编写的极简编码代理，追求低内存占用和高性能。它支持多种 AI 提供商（包括 OpenRouter 、 OpenAI 、 Anthropic 、 Gemini 、 Ollama 及自定义提供商），提供文件操作、受权限控制的 bash 执行、会话管理和支持 Markdown 渲染的终端 UI 。代码量约 7,000 行，空会话内存占用约 8MB，工作时约 12MB，二进制体积约 8.9MB 。

权限系统有四个等级，从严格到宽松，可为各工具配置模式并设定会话白名单；内置 doom-loop 检测可防止重复调用相同工具。代理配备了提示（prompts）系统，内置模式包括 code 、 plan 、 review 、 debug 、 ask 、 brainstorm 、 frontend-design 、 review-security 和 simplify，运行时可切换，且可通过 Markdown 文件添加自定义提示；代理还会自动从项目目录加载 AGENTS.md 或 CLAUDE.md 。

zerostack 包含面向长周期任务的实验性循环功能，能让代理按计划迭代执行直至完成；并集成了 git worktrees，支持"每任务一分支"的工作流，可在聊天 UI 中创建、切换、合并和退出 worktree 。该工具支持 MCP 服务器以扩展工具能力，并集成了 Exa 搜索用于网络抓取与检索。安装依赖 Cargo 和 git，且可选通过 bubblewrap 开启沙盒模式以隔离命令执行。 zerostack is a minimalistic coding agent written in Rust, designed for low memory usage and high performance. It supports multiple AI providers including OpenRouter, OpenAI, Anthropic, Gemini, and Ollama, along with custom providers. The tool offers file operations, bash execution with permission controls, session management, and a terminal UI with markdown rendering. At around 7,000 lines of code, it uses roughly 8MB of RAM for empty sessions and 12MB during work, with a binary size of 8.9MB.

The permission system includes four modes ranging from restrictive to permissive, with configurable per-tool patterns and session allowlists. Doom-loop detection prevents repeated identical tool calls. The agent features a prompts system with built-in modes like code, plan, review, debug, ask, brainstorm, frontend-design, review-security, and simplify, which can be switched at runtime. Custom prompts can be added via markdown files, and the agent automatically loads AGENTS.md or CLAUDE.md from the project directory.

zerostack includes an experimental loop system for long-horizon tasks, where the agent iteratively works through a plan until completion. It also provides git worktrees integration for branch-per-task workflows, allowing users to create, work in, merge, and exit worktrees from the chat UI. The tool supports MCP servers for extended tooling and includes integrated Exa search for web fetching and searching. Installation requires Cargo and git, with optional sandbox mode via bubblewrap for isolated command execution.

OpenAI 与 the Government of Malta 宣布了一项里程碑式合作——名为 "AI for All" 的计划，将为所有 Malta 公民免费提供一年的 ChatGPT Plus 使用权限。这是由国家政府推动的全球首例。 OpenAI 首席执行官 George Osborne 表示，他们将智能视为一种类似电力的普遍公共服务，应该不分背景、人人可及。此项计划的核心是一门由 the University of Malta 专门设计的 AI 素养课程。

该课程侧重教授可直接应用于日常生活与工作的实用 AI 技能。学员将学习人工智能的基本原理、其能力与局限，以及负责任使用的原则。完成课程后，公民将获赠一年免费 ChatGPT Plus 访问资格。通过这种结构化安排，确保每位参与者都能以自信与能力使用 AI，帮助社会平稳迈入数字化时代。

Malta 的经济部长 Silvio Schembri 将此举称为确保没有公民被落下的积极措施。他强调，该项目通过先进的数字工具为家庭、学生和在职人员提供切实帮助。首阶段由 the Malta Digital Innovation Authority 负责管理，计划于五月启动；随着参与度提升，项目将逐步扩展到更多 Malta 居民及海外公民。

这一合作也契合了 OpenAI 的全球 "OpenAI for Countries" 倡议，该倡议与希望在国家层面推进 AI 战略部署的政府合作。 OpenAI 并非推行一刀切的通用方案，而是根据各国的具体优先事项进行定制，无论侧重教育、劳动力培训还是公共服务。 Malta 的做法有效将教育课程与对先进工具的直接接触及国家层面的 AI 素养培养结合起来。

该计划已被视为对欧洲及其他国家具有示范意义的先行范例。 George Osborne 指出，Malta 通过率先行动，为各国政府如何通过 AI 赋能民众树立了有力的先例。通过将教育资源与即时获取前沿技术相结合，Malta 的方案旨在加快 AI 的普及，并为其经济发展与民众日常生活带来切实益处。 OpenAI and the Government of Malta have announced a landmark partnership that will provide all Maltese citizens with access to ChatGPT Plus at no cost for one year. Titled the "AI for All" initiative, this program represents a world-first for a national government. OpenAI CEO George Osborne emphasized their vision of intelligence as a universal utility, comparable to electricity, that should be accessible to everyone regardless of background. The cornerstone of this effort is a specially designed AI literacy course developed by the University of Malta.
The curriculum is designed to teach practical AI skills applicable to daily life and work. Participants will learn the fundamentals of artificial intelligence, its capabilities and limitations, and guidelines for responsible use. Upon completing the course, citizens are rewarded with a free year of ChatGPT Plus access. This structured approach ensures that every participant is equipped to use AI with confidence and competence, supporting a smooth transition into the digital age.
Malta's Minister for Economy, Silvio Schembri, described the initiative as a proactive step to ensure no citizen is left behind. He emphasized that the program provides families, students, and workers with practical assistance through cutting-edge digital tools. The first phase is scheduled to launch in May, managed by the Malta Digital Innovation Authority. As participation grows, the program will expand to include more Maltese residents and citizens living abroad.
This partnership aligns with OpenAI's global "OpenAI for Countries" initiative, in which the company collaborates with governments pursuing strategic national AI adoption. Rather than implementing a generic solution, OpenAI tailors its approach to align with each nation's specific priorities, whether focused on education, workforce training, or public services. Malta's model effectively integrates an educational curriculum with direct access to advanced tools and national-scale AI literacy.
The initiative has been recognized as a pioneering example for Europe and other nations. George Osborne noted that by leading the way, Malta is setting a powerful precedent for how governments can empower their populations through AI. By combining educational resources and immediate access to cutting-edge technology, Malta's approach seeks to speed AI adoption and deliver tangible benefits to its economy and the daily lives of its people.

"Halt and Catch Fire"（HCF）最初是程序员之间的玩笑用语，用来形容那些会让 CPU 停止正常工作、必须重启才能恢复的机器码指令。尽管 AMC 的同名剧集讲的是计算机行业，这个术语本身却更早产生，源自工程师的笑话。它成了一个总称，用来指未记录或无效的操作码导致处理器死锁、用于模拟挂起的测试模式，以及真实存在的硬件缺陷。这种幽默沿袭了像 ADD 、 JMP 这类三字母汇编助记符的风格，类似的例子还有 EPI（Execute Programmer Immediately）和 DC（Divide and Conquer）。

这个说法在 Motorola 6800 处理器上得到了具体的案例。 1977 年 12 月，Gerry Wheeler 在 BYTE 杂志上发表文章，披露了该芯片的一些未记录指令。在 256 种可能的操作码中，有 59 种未在官方文档中说明。其中两个字节 $9D 和 $DD 会让处理器停止正常工作，程序计数器开始快速在内存地址间跳动并忽略中断。 Wheeler 把这种现象称为 "Halt and Catch Fire"，并指出虽然机器并不会真的着火，但在无法恢复、必须重启的意义上确实近乎如此。

"catch fire"这一表述并非全然夸张。 IBM System/360 在遇到某些无效操作码时，因持续访问特定内存位置，确实有过过热甚至着火的风险。在 Motorola 6800 上，这种现象更像是地址总线变成了一个 16 位计数器，毫无目的地顺序读取内存。工程师们后来还找到了实用用途，比如在硬件初次调试时用来扫描 RAM，把潜在的漏洞变成了"意外之喜"。现代的研究也证实了这些细节，发现计数模式开始前存在延迟，而且不同未记录操作码的行为各不相同。

类似的概念并不限于 Motorola 。 6502 处理器存在会锁死 CPU 的非法操作码，Pentium 的 F00F bug 允许精心构造的指令冻结处理器；在一些架构上，某些指令对可能会永远等待永远不会到来的中断。现代对 x86 的模糊测试仍通过向复杂处理器输入随机或异常数据来发现无效状态，说明即便技术进步，这类问题依然存在。

随着软件层次越来越高，人们更容易忽视底层硬件的实际情况。 HCF 提醒我们计算机毕竟是物理系统，可能以戏剧性的方式出错。这一短语之所以经久不衰，是因为它既反映了技术现实，又带有工程师式的黑色幽默。鉴于其吸引力，将 "HCF" 用于未来的项目或公司以延续这段计算史，也并不令人意外。 The phrase "Halt and Catch Fire" (HCF) originated as programmer humor, describing machine-code instructions that cause a CPU to stop functioning usefully, requiring a reset to recover. While the AMC show of the same name is about the computer industry, the term itself is much older, rooted in engineering jokes. It became a catch-all label for undocumented or invalid opcodes that lock up processors, intentional test modes that mimic hangs, and real hardware bugs. The humor fits the pattern of three-letter assembly mnemonics like `ADD` or `JMP`, with other examples including `EPI` (Execute Programmer Immediately) and `DC` (Divide and Conquer).

The phrase gained real significance with the Motorola 6800 processor. In December 1977, Gerry Wheeler published an article in BYTE magazine detailing undocumented instructions for the chip. Out of 256 possible opcodes, 59 were unaccounted for in official documentation. Two specific bytes, `$9D` and `$DD`, caused the processor to stop normal operation, with the program counter advancing rapidly through memory addresses while ignoring interrupts. Wheeler explicitly named this behavior "Halt and Catch Fire," noting that while the machine didn't literally catch fire, it came close in terms of being unrecoverable without a reset.

The "catch fire" part wasn't entirely fictional. The IBM System/360 could literally overheat and catch fire when encountering certain invalid opcodes due to constant access to specific memory locations. On the Motorola 6800, the behavior was more about the address bus turning into a 16-bit counter, reading memory sequentially without purpose. Engineers later found uses for this behavior, such as scanning RAM during hardware bring-up, turning what could have been a bug into a "happy accident." Modern investigations have confirmed these details, showing delays before the counting pattern begins and variations in how different undocumented opcodes behave.

The concept extends beyond Motorola. The 6502 processor had illegal opcodes that locked the CPU, and the Pentium F00F bug allowed carefully crafted instructions to freeze processors. Some instruction pairs on various architectures can wait forever for interrupts that never arrive. Modern x86 fuzzing continues to uncover invalid states in complex processors by feeding them random or unexpected data, proving that these issues persist even as technology advances.

As software becomes more abstracted, it's easy to forget the underlying hardware realities. HCF represents a reminder that computers are physical systems where things can go wrong in dramatic ways. The phrase has endured because it captures both the technical reality and the dark humor of engineering failures. Given its appeal, it wouldn't be surprising to see "HCF" adopted for future projects or companies, keeping this piece of computing history alive.

Kioxia 与 Dell 合作，推出了一款超高密度存储服务器：在仅 2RU 的机箱内集成了约 10 PB 的闪存容量。 Dell PowerEdge R7725xd 服务器配备 40 块 Kioxia LC9 E3.L 规格的 NVMe SSD，每块容量为 245.76 TB，总容量约为 9.8 PB；系统采用 AMD EPYC 9005 处理器，并可支持多达五块 400 Gbps 网卡以实现高速数据传输。 Dell 的 Arun Narayanan 指出，这一组合在不牺牲性能的前提下，提供了扩展 AI 基础设施所需的存储密度和能效。若在一个机架中部署 20 台此类服务器，理论总存储可达约 196 PB 。

Kioxia 的 Neville Ichhaporia 强调，这类服务器能为客户带来显著好处：支持海量数据摄入流、轻松扩展数据湖，并在更小的物理空间内处理大规模备份，从而大幅降低总体拥有成本。 LC9 SSD 是实现这一目标的关键，属于当前市面上容量最大的固态硬盘之一。其他主要存储厂商也在推进超大容量 SSD 领域的布局，包括 Micron 的 6600 ION 、 Sandisk 的 UltraQLC SN670，以及 SK Hynix 的 AIN D 和其子公司 Solidigm 的相关产品。

展望未来，Scality 透露 Samsung 正研发容量可达 1 PB 的近线级 SSD，被视为潜在的传统硬盘驱动器替代品，这反映出业界正持续朝着大幅提升固态存储密度的方向发展。 Kioxia 与 Dell 的合作展示了这些高容量 SSD 在需要兼顾性能与密度的场景（如 AI 基础设施）中的实际应用价值。 Kioxia and Dell have partnered to create an extremely dense storage server, packing 10 petabytes of flash capacity into a slim 2RU form factor. Dell's PowerEdge R7725xd server uses 40 of Kioxia's LC9 E3.L form factor NVMe SSDs, each with a capacity of 245.76 TB, to achieve a total of 9.8 PB. The system is powered by AMD EPYC 9005 processors and supports up to five 400 Gbps network interface cards for high-speed data transfer. Dell's Arun Narayanan highlighted that this combination delivers the storage density and power efficiency needed to scale AI infrastructure without compromising performance. A single rack containing twenty of these servers could hold a massive 196 PB of storage.

Kioxia's Neville Ichhaporia emphasized the benefits for customers, noting that these servers enable massive ingestion streams, effortless scaling of data lakes, and handling large backups in a fraction of the physical footprint, significantly improving total cost of ownership. The LC9 SSDs are central to this achievement, representing some of the highest-capacity drives available. Other major storage developers are also pushing into this ultra-high-capacity SSD space, including Micron with its 6600 ION, Sandisk with the UltraQLC SN670, and SK Hynix with its AIN D drive, along with its Solidigm subsidiary.

Looking even further ahead, Scality has revealed that Samsung is developing nearline-class SSDs with capacities reaching up to 1 petabyte, which are seen as potential replacements for traditional hard disk drives. This indicates a continued industry trend toward vastly increasing solid-state storage density. The collaboration between Kioxia and Dell demonstrates a practical application of these high-capacity SSDs, targeting demanding workloads like AI infrastructure where both performance and density are critical.

文章深入探讨了五种 HTML 列表类型，提醒开发者不要只局限于无序列表和有序列表。选择合适的列表更多是关于语义和含义，而不是仅仅为了视觉呈现。作者给出了一套决策框架：对用户输入字段使用控件列表；当顺序重要时使用有序列表；对键值对使用描述列表；对 UI 控件使用菜单；无序列表则作为默认的通用选择。

控件列表通过 <select> 配合 <option> 用于固定选择，而 <datalist> 用于提供建议输入。 <select> 支持使用 <optgroup> 进行分组、用 <hr> 做视觉分隔，以及通过 size 属性控制可见项数。 <datalist> 可与任何输入类型配合，包括范围滑块和周选择器，但作者提醒不同浏览器在为 datalist 选项应用样式时存在不一致，尤其是 Chrome 和 Firefox 之间。

当项目顺序会改变含义时，应使用有序列表（<ol>），例如食谱、算法或按字母排序的列表。文章介绍了有用的属性，如 reversed（用于降序编号）和 start（用于在多个列表间保持编号连续）。还演示了有序与无序列表如何嵌套，能构建出结构复杂但语义清晰的内容，即便在没有视觉渲染时也能保持可理解性。

描述列表（<dl>）在 HTML4 中曾被狭义地视为定义列表，但在 HTML5 中已扩展为适用于任何键值对关系。作者建议广泛使用它们来展示元数据、用户资料，甚至用于调试 JSON 对象。 HTML5 允许在 <dl> 中使用 <div> 将相关的术语和描述分组，这使得通过 CSS 进行样式化更为灵活。

<menu> 元素被推荐用于工具栏和控制按钮，在语义上不同于作为分区元素的 <nav> 。 <nav> 可以包含标题、段落等多种内容，而 <menu> 则专用于包含带有交互命令的列表项。文章澄清这两者并非互斥：<menu> 可以出现在 <nav> 内，但反过来则不成立。

当没有其他列表类型的特定语义适用时，无序列表（<ul>）仍是默认选择，作为顺序无关集合的语义"杂物抽屉"。 The article dives deep into the five different types of HTML lists, challenging developers to move beyond basic unordered and ordered lists. It emphasizes that choosing the right list type is about semantics and meaning, not just visual presentation. The author provides a decision-making framework: use control lists for user input fields, ordered lists when sequence matters, description lists for key-value pairs, menus for UI controls, and unordered lists as the default catchall.

Control lists are explored through `<select>` with `<option>` for fixed selections and `<datalist>` for suggested inputs. The `<select>` element supports grouping with `<optgroup>`, visual breaks with `<hr>`, and the `size` attribute for controlling visible items. Meanwhile, `<datalist>` can be paired with any input type, including range sliders and week pickers, though the author warns about browser inconsistencies in styling datalist options, particularly between Chrome and Firefox.

Ordered lists (`<ol>`) should be used whenever changing the item order would change the meaning, such as for recipes, algorithms, or alphabetical listings. The article covers useful attributes like `reversed` for descending numbering and `start` for maintaining continuity across multiple lists. It also demonstrates how ordered and unordered lists can be nested together to create complex, well-structured content that remains meaningful even without visual rendering.

Description lists (`<dl>`), once narrowly defined as definition lists in HTML4, have evolved in HTML5 to handle any key-value pair relationships. The author advocates for using them extensively for metadata display, user profiles, and even debugging JSON objects. HTML5 now allows `<div>` wrappers inside `<dl>` to group related terms and descriptions together, making them more flexible for styling with CSS.

The `<menu>` element is presented as the semantic choice for toolbars and control buttons, distinct from `<nav>` which is a sectioning element. While `<nav>` can contain various content types including headings and paragraphs, `<menu>` exclusively contains list items with interactive commands. The article clarifies that these elements aren't mutually exclusive, a `<menu>` can exist within a `<nav>`, but not vice versa. The unordered list (`<ul>`) remains the default choice when no other list type's specific semantics apply, serving as the semantic junk drawer for collections where order doesn't matter.

LLM 引导技术因 DeepSeek-V4-Flash 的出现又重新受到关注。这款开源模型足够强大，在代理型编码任务上能与一些低端前沿模型竞争。由于引导需直接访问本地模型的内部激活值，过去对大多数工程师而言并不现实。 DeepSeek-V4-Flash 改变了这一局面，开发者 antirez 已在 DwarfStar 4（为该模型特别精简的 llama.cpp 分支）中加入了引导支持。尽管目前的实现还很基础，该项目仅上线八天，但值得持续关注。

引导的原理是从模型的内部激活值中提取出某个概念，然后在推理时提升那些特定数值。最简单的做法是把同一条提示分别输入两次，一次正常输入、一次加上诸如"简洁回答"之类的修饰语，然后比较两次的激活值差异。这个差异就构成了一个"引导向量"，可以在任意层加到激活值上以产生期望效果。更复杂的方法会用稀疏自编码器等手段来识别模型行为中更深层的模式，类似 Anthropic 的相关研究。引导的吸引力在于，它像是找到了模型"大脑"的控制面板，可以用滑块直接调整冗长或细致程度等特质，而不是反复琢磨提示措辞。

尽管吸引人，引导并未广泛普及，原因有几方面。大厂可以通过训练直接改动模型，无需在推理时做这种笨拙的"手术"；普通用户通过 API 使用时也无法访问权重和激活值；而且许多基础的引导需求其实已经能通过更巧妙的提示解决——提示词元本身就能对模型行为提供极细粒度的控制。引导因此处于一个尴尬的中间地带：对大多数用户而言太复杂，对拥有完整模型访问的大厂来说又没必要。

引导最有希望的应用场景是提示失败时的补救。例如，"智能"这类能力过去还能靠"你是专家"之类的提示激活，但现在已经内嵌在新一代模型中。作者怀疑是否存在一个能实用地表示"智能"的引导向量，因为这类复杂概念很可能分布在几乎整个模型的权重上，解决它等同于训练出更聪明的模型。另一种可行性稍高的想法是把引导当作一种数据压缩手段：提取那些本来需要大量词元才能表达的概念，比如对某个代码库的深入知识。尽管略显可行，这种做法仍面临同样的根本挑战。

总体上，作者对引导技术抱有兴趣但持悲观态度，认为大多数收益更适合通过提示优化或微调来获得。不过开源社区对引导的探索尚浅，这一状况可能正在改变。如果引导确实有隐藏的实用价值，未来六个月内应该会逐步显现。未来开源权重模型发布时，也可能会出现社区提取的"可增强特征库"，类似目前量化版本和各种封装器的繁荣。

来自 Hacker News 评论的一条重要更新指出，引导能够改变提示无法触及的已训练行为，尤其是在移除模型拒绝回答方面已有实际效果——这也是一些开源模型去审查或所谓"abliteration"操作的实现方式之一。 antirez 指出，相比一次性修改权重，运行时引导对模型能力的损害更小且可按需启用，因此是一种更可取的轻量方案。 LLM steering is experiencing renewed interest thanks to DeepSeek-V4-Flash, a new open model that's powerful enough to compete with lower-end frontier models for agentic coding tasks. Since steering requires direct access to a local model's internal activations, it has historically been impractical for most engineers. DeepSeek-V4-Flash changes that equation, and developer antirez has already built steering support into DwarfStar 4, a stripped-down llama.cpp fork designed specifically for this model. While the current implementation is basic, the project is only eight days old and worth watching.

Steering works by extracting a concept from a model's internal activations and then boosting those specific numerical values during inference. The simplest approach involves feeding the same prompts twice, once normally and once with a modifier like "respond tersely," then measuring the difference in activations between the two runs. This difference becomes a "steering vector" that can be added to activations at any layer to produce the desired effect. More sophisticated methods use sparse autoencoders to identify deeper patterns in the model's behavior, similar to what Anthropic has published research on. The appeal of steering is that it feels like finding a control panel for the model's brain, with sliders for traits like verbosity or conscientiousness that could be adjusted directly rather than fiddling with prompt wording.

Despite its appeal, steering hasn't seen widespread adoption for several reasons. Major AI labs can manipulate their models directly through training rather than awkward mid-inference surgery. Regular users lack access to model weights and activations when using APIs. Most basic steering applications are already outcompeted by simply prompting the model more effectively, since prompt tokens already provide extremely fine-grained control over model behavior. Steering occupies an awkward middle ground that's too complex for most users but unnecessary for the labs with full model access.

The most promising potential for steering lies in cases where prompting fails. One example is "intelligence" itself, which used to be promptable with phrases like "you are an expert" but is now baked into current-generation models. However, the author is skeptical that an "intelligence" steering vector exists in any practical sense, since such a complex concept likely spans nearly the entire model's weights, making the problem equivalent to training a smarter model. Another possibility is using steering as data compression, extracting concepts that would otherwise require many tokens to express, like deep knowledge of a specific codebase. This seems marginally more plausible but still faces the same fundamental challenge.

The author remains fascinated but ultimately pessimistic about steering's practical applications, believing most gains can be more efficiently achieved through prompting or fine-tuning. However, the open-source community hasn't explored steering extensively yet, and that may be changing. If steering does have hidden practical value, the next six months should reveal it. It's possible that future open-weight model releases will come with community-extracted "libraries" of boostable features, similar to how quantized versions and wrappers currently proliferate.

A notable update from Hacker News comments revealed that steering can modify trained-in behaviors in ways prompting cannot, most significantly for removing model refusals. This is already how some uncensoring or "abliteration" is done for open models. Antirez pointed out that weight modification can damage model capabilities more than runtime steering, which can be applied only when needed, making the lighter-touch approach preferable.

SANA-WM 是 NVIDIA 研究人员开发的一款 26 亿参数开源世界模型，能够从单张起始图像和相机轨迹生成高保真 720p 、最长可达一分钟的视频。该模型兼顾效率与质量：只需 64 块 H100 GPU 训练 15 天，推理时仅需单块 GPU 。其蒸馏版在 RTX 5090 上配合 NVFP4 量化，仅需 34 秒就能对一段 60 秒的 720p 视频完成去噪，使分钟级世界建模更容易普及。

SANA-WM 的表现源自四项核心创新。混合线性注意力将逐帧门控的 DeltaNet 与周期性 softmax 注意力结合，在保持长距离上下文连贯性的同时节省内存，避免了纯 softmax 模型在 60 秒时长下常见的内存溢出问题。双分支相机控制同时采用全局位姿分支与像素对齐的精细几何分支，以高保真度跟踪 6-DoF 相机轨迹。两阶段生成管线把第一阶段的结果输入到一个 170 亿参数的长视频精化器，用于提升整个序列的纹理、运动和一致性。最后，鲁棒的标注管线能从公开视频中提取精确的度量级 6-DoF 相机位姿，生成约 21.3 万个带有高质量时空一致动作标签的片段用于训练。

该模型擅长从静止的第一人称视角生成多样化的自主动画场景。演示涵盖雪山小径、水下古庙、外星沼泽到后末日高速公路等环境，其中漂浮的雪粒、摇曳的植被、闪烁的火焰和流动的水等独立运动元素在整段一分钟的视频中自然持续。 SANA-WM 还支持可控的相机轨迹，示例显示在相同起始帧上沿盐滩、冰湖和丛林峡谷等不同路径移动，精确遵循指定的 6-DoF 运动。

在作者提出的一分钟世界模型基准测试中，SANA-WM 在动作跟随准确性上优于此前的开源基线，同时在视觉质量上与 LingBot-World 和 HY-WorldPlay 等大型工业模型相当，但吞吐量提升了 36 倍。两阶段精化显著改善了后期时间窗的画质、纹理细节和运动平滑度，有效缓解了长时长视频生成中常见的退化问题。高效的训练流程、单 GPU 部署能力与开源可用性共同使 SANA-WM 成为向实用化、高质量世界建模迈出的重要一步。 SANA-WM is a 2.6B-parameter open-source world model developed by NVIDIA researchers that generates high-fidelity, 720p videos lasting up to one minute from a single starting image and a camera trajectory. The model is designed for efficiency and quality, capable of being trained on just 64 H100 GPUs over 15 days and running inference on a single GPU. Its distilled variant can denoise a 60-second 720p clip in only 34 seconds on an RTX 5090 using NVFP4 quantization, making minute-scale world modeling more accessible.

Four core innovations drive SANA-WM's performance. Hybrid Linear Attention combines frame-wise Gated DeltaNet with periodic softmax attention to maintain coherent long-range context while staying memory-efficient, avoiding the out-of-memory issues that plague all-softmax models at 60-second durations. Dual-Branch Camera Control uses both a global pose branch and a fine pixel-aligned geometric branch to precisely follow 6-DoF camera paths with high fidelity. A Two-Stage Generation Pipeline feeds stage-1 outputs into a dedicated 17B long-video refiner that sharpens texture, motion, and consistency across the full sequence. Finally, a Robust Annotation Pipeline extracts accurate metric-scale 6-DoF camera poses from public videos, producing about 213K clips with high-quality spatiotemporally consistent action labels for training.

The model excels at generating diverse, autonomously animated scenes from stationary first-person viewpoints. Demos showcase environments ranging from snowbound alpine trails and underwater ancient temples to alien swamps and post-apocalyptic highways, where independent motion like drifting snow particles, swaying vegetation, flickering flames, and flowing water continues naturally throughout the minute-long clips. SANA-WM also supports controllable camera trajectories, with examples showing different paths taken from the same starting frame across salt flats, frozen lakes, and jungle canyons, demonstrating precise adherence to specified 6-DoF movements.

On the authors' one-minute world-model benchmark, SANA-WM achieves stronger action-following accuracy than prior open-source baselines while delivering comparable visual quality to large-scale industrial models like LingBot-World and HY-WorldPlay at 36 times higher throughput. The two-stage refinement process notably improves late-window quality, texture detail, and motion smoothness compared to stage-one outputs alone, addressing common degradation issues in long-duration video generation. The combination of efficient training, single-GPU deployment, and open-source availability positions SANA-WM as a significant step toward practical, high-quality world modeling.

一位居住在中国的父亲制作了一套希腊字母卡片，通过视觉联想帮助年幼的孩子学习希腊语。核心理念是把每个物体画成与其名称所对应的希腊字母在形状上相似，形成双重记忆：字母的形状联想起物体，物体的名称又反过来强化对字母的记忆。研究表明，这种方法比死记硬背能更快地让孩子掌握字母表。

为寻找合适的词汇，他使用了 GreekLex——一个包含超过 35,000 个现代希腊语单词并带有频率数据的语料库。他筛选出长度在 3 到 10 个字符之间、且在语料库中出现次数至少为 100 次的单词，以确保这些词对孩子较为熟悉。即便如此，每个字母仍有数百到数千个候选词，他便把这些候选词分批输入 ChatGPT，让它挑出那些可以被合理地画成呼应字母形状的物体。例如，橄榄树（ελιά）可以被风格化为一根竖直的树干和向右延展的三个圆形枝丫，呼应 epsilon（ε）的三条臂。从入围项中，他又使用 OpenAI 的图像生成模型来生成插图，有时还提供希腊字母的参考图以引导输出。有些字母较难表达，比如 phi（φ），他最后不得不先手绘一条蛇的草图，再请模型以 Eric Carle 风格渲染成图。

该项目制作了两套卡片。物体卡展示与相应希腊字母形状相似的插图，卡片底部印有字母和物体名称；字母卡只显示字母，背面印着创作者孩子的照片，这样两套卡片可以搭配用于各种游戏。插图延续了 Eric Carle（《好饿的毛毛虫》作者）那种色彩丰富、拼贴感强的美学风格。他还为卡片排版写了自定义代码。

这家人用这套卡片有几种玩法。孩子们先认识物体并学习每个物体如何在视觉上呼应对应的字母——大多数物体都来自他们的日常生活。一个下午、分两次各半小时的学习后，他们就掌握了大约 18 个字母。他们会玩记忆配对游戏：把字母卡面朝下、物体卡面朝上，轮流翻牌配对；还会玩一个叫"火"的体感游戏：父亲拿着一张卡站着，假装身后有火焰，每回答正确他就向前走一步离开火源，答错则后退并假装被烧着，孩子们觉得好玩极了。

创作者承认自己并非第一个使用这种视觉联想法的人，英语已有类似产品。但他认为自己的作品很可能是首套针对希腊字母的此类卡片。他也觉得市面上大多数英语字母卡设计平庸，常见的是把物体放在字母背后而非让物体本身呈现字母形状。他的目标是做出视觉上更机智、设计更巧妙的卡片，而且最重要的是，他的孩子们确实很喜欢玩这些卡片。 A father living abroad in China created a set of Greek alphabet cards to help his young children learn the language through visual associations. The core idea is that each object is drawn so it physically resembles the Greek letter its name begins with, creating a dual memory link where the shape of the letter evokes the object and the object's name reinforces the letter. Research suggests this method helps children learn the alphabet significantly faster than rote memorization.

To find suitable words, the creator used GreekLex, a corpus of over 35,000 Modern Greek words with frequency data. He filtered for words between 3 and 10 characters long that appeared at least 100 times in the corpus, ensuring the vocabulary would be familiar to children. This still left hundreds to thousands of candidates per letter, so he fed them to ChatGPT in batches to identify which objects could plausibly be drawn to echo a letter's shape. For example, an olive tree (ελιά) could be stylized with a vertical trunk and three rounded branches extending right, mirroring the three arms of epsilon (ε). From the shortlisted candidates, he used OpenAI's image generation model, sometimes providing a reference image of the Greek letter to guide the output. Some letters proved stubborn, like phi (φ), where he ultimately had to hand-draw a snake sketch and ask the model to render it in the appropriate Eric Carle-inspired style.

The project produced two card sets. Object cards show an illustration resembling its Greek letter, with the letter and object name displayed at the bottom. Alphabet cards show just the letters, with the backs featuring photos of the creator's children so the sets can be used together in games. The illustrations follow the colorful, collage-like aesthetic of Eric Carle, author of The Very Hungry Caterpillar. The creator wrote custom code to handle the card layout.

The family uses the cards in several ways. First, the children learn the objects and the visual trick of how each echoes its letter, with most objects already familiar from their daily world. In one afternoon of two half-hour sessions, they learned about 18 letters. They play a memory game by laying alphabet cards face down and object cards face up, taking turns flipping and matching. They also play a physical "fire game" where the father stands holding a card and pretends there is a fire behind him. Each correct answer lets him step forward away from the fire, while wrong answers mean stepping back and pretending to burn, which the children find hilarious.

The creator acknowledges he is not the first to use this visual association method, as similar products exist for English letters. However, he believes his are the first for the Greek alphabet. He also argues that most English alphabet cards on the market are poorly designed, typically showing an object merely hiding behind a letter rather than being shaped like it. His goal was to create something visually cleverer, and most importantly, his kids genuinely enjoy playing with the cards.

Failed to crawl the webpage. (English version not available)

本文提出了 δ-mem，一种轻量级的记忆机制，旨在帮助大型语言模型在长期助手和智能体系统中累积并重用历史信息。不同于扩展上下文窗口（既计算开销大又常常效果有限），δ-mem 在冻结的全注意力主干上增加了一个紧凑的在线联想记忆状态。该机制将历史信息压缩为固定大小的状态矩阵，并通过 delta 规则进行更新；在文本生成时，从该记忆读出信息，为主干的注意力计算提供低秩修正。

尽管在线记忆仅为 8×8 的小规模，δ-mem 仍显著提升了性能：平均得分比冻结主干高 1.10 倍，比最强的非 δ-mem 记忆基线高 1.15 倍；在内存密集型任务上提升更明显，在 MemoryAgentBench 上达到 1.31 倍，在 LoCoMo 上达到 1.20 倍。重要的是，这些改进并未削弱模型的通用能力。

δ-mem 的一大优势是无需完全微调、替换主干网络或显式扩展上下文即可运行，使其成为一种实用且高效的增强记忆方案。实验结果表明，通过与注意力计算直接耦合的紧凑在线状态，就能实现有效的长期记忆能力。 The paper introduces δ-mem, a lightweight memory mechanism designed to help large language models (LLMs) accumulate and reuse historical information in long-term assistants and agent systems. Rather than expanding the context window, which is computationally expensive and often ineffective, δ-mem augments a frozen full-attention backbone with a compact online state of associative memory. This approach compresses past information into a fixed-size state matrix that is updated using delta-rule learning. During text generation, the system uses a readout from this memory to produce low-rank corrections to the backbone's attention computation.

Despite its small size, an 8×8 online memory state, δ-mem significantly boosts performance. On average, it achieves scores 1.10 times higher than the frozen backbone and 1.15 times higher than the strongest non-δ-mem memory baseline. The gains are even more pronounced on memory-intensive tasks, reaching 1.31 times on MemoryAgentBench and 1.20 times on LoCoMo. Importantly, these improvements come without compromising the model's general capabilities.

A key advantage of δ-mem is that it operates without requiring full fine-tuning, backbone replacement, or explicit context extension. This makes it a practical and efficient solution for enhancing memory in LLMs. The results demonstrate that effective memory can be achieved through a compact online state that is directly coupled with the attention computation.