←搜索结果

按匹配度 · 共 1540 条结果 清除搜索

星期五

  1. Databricks Blog

    Databricks 用领域知识注入提升 LLM 对 npm 二进制恶意包的检测能力

    Databricks 自 2025 年 11 月起收集 npm 恶意样本,在扫描并核实的 11,413 个 package-version 中,96.2% 将载荷藏在源码或外部服务器,仅 3.8%(432 个)捆绑二进制,但这类包无法得出明确恶意结论的比例是纯源码包的两倍以上(14.4% vs 6.2%)。团队评估了 9 个模型(含 Claude Haiku 4.5、Gemini 3.5 Flash、GPT-5.4-nano、Qwen3-Next 80B 等),并在推理阶段注入预计算的静态证据作为能力画像,以提升预算级与前沿模型的检测质量。

    image1.png

    包”。2025 年。https://sourcecodered.com/indonesianfoods-npm-worm/ [3] Kieran Miyamoto。“朝鲜 npm 包”。https://dprk-research.kmsec.uk/ [4] Amazon。“Amazon 识别出开源供应链攻击背后的朝鲜黑客组织”。2026 年。https://aws.amazon.com/blogs/security/amazon-identifies-north-korean-hacker-group-behind-open-source-supply-chain-attacks/ [5] OpenSourceMalware。“PolinRider:朝鲜威胁行为者在数百个 GitHub 仓库中植入恶意软件”。2026 年。https://github.com/OpenSourceMalware/PolinRider [6] Zscaler。“追踪 Shai-Hulud:深入解析 ChainDrop NPM 蠕虫”。2026 年。https://www.zscaler.com/blogs

    匹配字段正文

  2. Unite.AI精选

    Reflection 以 Apache 2.0 开源 501B 参数 MoE 模型 Beam

    Reflection 于 10 月 16 日在 Hugging Face 以 Apache 2.0 许可发布其首个开放权重模型 Beam,Baseten 同日通过 Model APIs 和 Dedicated Inference 提供 day-0 支持。Beam 为 501B 总参数、每 token 激活 23B 的 MoE 自回归 Transformer,上下文窗口最高 100 万 token,预训练使用 23.8 万亿 token,并以 BF16、FP8、NVFP4 精度发布权重。Reflection 给出的评测结果为 GPQA Diamond 90.5%、SWE-bench Verified 80.9%、AIME 2026 97.8%、Terminal-Bench 2.1 80.1%、MCP Atlas 78.7%,同时开源了终端智能体工具 Mirror CLI。

    n roughly 1.3 billion sandboxes across training, evaluation, and ablations, with the final run peaking at 883,006 vCPUs, 83% of allocatable capacity, across 23 clusters on GCP and AWS. Reflection said it plans to release self-distillation recipes in the coming days and will adopt OpenEnv, an open environment standard, for future runs. Baseten describes Beam as the first in a new family of models from Reflection, with multilingual and multimodal capabilities planned for future releases. Theo Nash

    匹配字段正文

  3. 半导体产业纵横

    数据中心CPU收入增长35%,Yole报告揭示网络芯片扩张

    Yole Group数据显示,2026年第二季度数据中心CPU收入达132亿美元,环比增长34.9%,增速超过AI ASIC(+23%)和GPU(+13%)。数据中心网络芯片市场2026年将达530亿美元,其中AI工作负载占约88%收入,纵向扩展交换机销量同比增长94%。AMD CEO苏姿丰称2026年为“AI推理元年”,预计到2030年数据中心AI加速器市场规模接近1.5万亿美元,CPU市场复合年均增长率约50%。

    成为了主要支出项目之一。 美国公司主导着这个充满活力的市场。NVIDIA在2025年和2026年位居全球第一大网络ASIC供应商,它率先推出了基于交换机的纵向扩展架构,并且是高性能AI千兆集群横向扩展的关键参与者。 从2026年起,博通、思科、Marvell、Astera Labs和华为等公司将占据重要地位,尤其是在AMD MI400 GPU方面。与此同时,AWS、Meta和微软等超大规模数据中心运营商也在开发各自的AI ASIC芯片。谷歌的TPU则独立于传统的纵向扩展和横向扩展框架,采用环形拓扑结构进行点对点连接,而非通过数据包交换机。 Yole表示,长期以来一直是 NVIDIA 优势的纵向扩展交换技术,随着开放标准的出现,为超大规模数据中心提供了一个可靠的替代方案,这一技术正在向 NVIDIA 开放。 在横向扩展交换机领域,博通预计在2026年将引领市场,英伟达紧随其后,思科、Marvell和华为分列其后。 网卡(NIC)的排名则截然不同:英伟达遥遥领先,AMD、AWS Nitro(供内部使用)和博通分列其后。由于大多数网络芯片都采用先进工艺制造,台积电是该市场的领先代工厂。在中国

    匹配字段正文

  4. Reflection Blog精选

    Reflection 发布首个开源权重模型 Beam,501B 参数稀疏 MoE

    Reflection 发布首个开源权重模型 Beam,权重已在 Hugging Face 以 Apache 2.0 许可开放。Beam 是稀疏 MoE 模型,总参数 501B、激活 23B,面向编码、推理和长程智能体任务,同时提供 BF16、FP8、NVFP4 精度格式和投机解码草稿模型。Reflection 还联合 40 多家合作伙伴构建其称为智能体开放操作系统的生态,覆盖 vLLM、SGLang、Unsloth、Axolotl 等推理与微调框架,并开源了配套的 Mirror CLI 智能体工具。

    Reflection.ai dotted artwork

    ion’s internal sandbox dashboard. With evaluations and other RL experiments running alongside, the fleet peaked at 883,006 vCPUs (83% of allocatable) across 23 clusters on GCP and AWS. We partnered with NVIDIA (NeMo RL, AutoModel, Megatron), RadixArk (Miles), Hugging Face (TRL), Anyscale (SkyRL) and Axolotl to build fine-tuning and RL workflows, so developers can deeply customize Beam for their own use cases. Beam Self-Distillation Developers can also leverage novel approaches like self-distilla

    匹配字段正文

  5. ChinaTalk

    蚂蚁 Ling 产品负责人 Richard Bian:为何他认为 AGI 是“必然的

    蚂蚁集团基础大模型 Ling 的产品运营负责人 Richard Bian 在 ChinaTalk 播客中表示,AGI 的到来是必然的,但没人知道它最终会是什么样,当前仍处于探索这些模型能力的早期阶段。他认为 Hugging Face 遭 OpenAI 研究智能体入侵一事说明进展显著,且闭源模型受自身限制未能处理日志,最终由 Z.ai 的开源模型完成。访谈录制后,蚂蚁于 10 月 14 日开源 Ling 3.1 Flash,参数规模 560B,上下文窗口最高 1M token。

    . That’s good, but those expert models have one liability: when general intelligence expands, they immediately become obsolete. Inference cost is also getting cheaper, and by the laws of physics it will keep getting cheaper. The frontier models may stay expensive, but overall token costs will fall. People have been saying general intelligence is the most important thing. That’s why in 2025 and early 2026, the Ling family built for general intelligence too. Once we saw diminishing marginal return

    匹配字段正文

  6. The Rundown AI

    马斯克的"飞行汽车"来了:Tesla Roadster 空中特技亮相,2028 年交付目标承压

    Tesla 在得州韦科展示了量产版 Roadster,这款白色、灵感来自 SR-71 的跑车用 SpaceX 冷气推进器飞越了 Tesla 全系车型。10 个推进器依靠 600 升可充装压缩空气驱动,Performance(5 万美元预订)与 Thrust(25 万美元、0–60 mph 加速不到 1 秒)为两座,仅限邀请的 Rocket 为单座并配齐 10 个推进器,各版本均承诺 300 英里续航。该车 2017 年首次亮相、原定 2020 年交付,如今生产目标定在 2028 年,道路合法性仍未明确。

    g and transcripts are encrypted; both devices have an on/off button and a light indicating when they’re listening. Security firm Trail of Bits’ Sept. audit confirmed the biggest flaws were fixed, but some issues remained unresolved, including deletion of unencrypted summaries. Why it matters: From Plaud’s AI recorders to Meta’s Muse Charm, tech companies are turning everyday conversations into a market for personal AI. Amazon’s angle connects these conversations to Alexa and, if users opt in, it

    匹配字段正文

  7. The Rundown AI

    Meta 发布 816B 参数开源多模态模型 Muse Spark Open,可双 Mac Studio 本地运行

    Meta 发布开源权重多模态模型 Muse Spark Open,参数量 816B,面向编码与智能体任务,并与 Apple 合作通过更新后的 MLX 工具包在两台 M5 Ultra Mac Studio(合计 1TB 内存)上本地运行。Meta 称该模型在无云环境下完成 iOS 应用修复和在 DaVinci Resolve 中剪辑视频。在 Artificial Analysis 的 Intelligence Index 上它得分 39,高于西方开源模型,但低于 GLM 5.3 的 45、Kimi K3 的 44 和 Qwen3.8 的 40,中国模型在编码上仍领先。

    ss to assess its AI safety practices and publish findings without its approval. Bank of America CEO Brian Moynihan told CNBC that the bank hired 1K contractors to patch software flaws exposed by Anthropic’s Mythos and other models. OpenAI is training employees to guard against espionage, including “honeypot” schemes, as attempts to steal AI trade secrets intensify, the Wall Street Journal reported. OpenAI also launched GPT-Live-1.1 with improved multilingual accents and upgraded ChatGPT’s Chrome

    匹配字段正文

  8. SiliconANGLE

    智能体 AI 推动戴尔基础设施战略成型

    随着企业从 AI 实验阶段转向规模化生产,智能体 AI 正推动戴尔调整基础设施战略,重点转向混合环境下的数据整合、成本、可靠性与治理。戴尔推出 AI Data Platform,帮助企业在本地数据所在处统一存储、安全与数据管理能力;其高管称去年 AI 需求订单达 640 亿美元,仅第四季度就有 340 亿美元。

    cloud, cybersecurity and more 11.4k+ theCUBE alumni — Connect with more than 11,400 tech and business leaders shaping the future through a unique trusted-based network Are you an AWS customer? Support SiliconANGLE financially by buying your AWS services from our Marketplace portal page and links: https://siliconangle.com/aws-marketplace/ About SiliconANGLE Media SiliconANGLE Media is a recognized leader in digital media innovation, uniting breakthrough technology, strategic insights and real-ti

    匹配字段正文

  9. Latent Space精选

    Goodfire 的 Nick Cammarata 提出神经网络已足够清晰,称多义性判断有误

    Goodfire 的 Nick Cammarata 发布文章提出,多数神经元本身是可解读的,许多著名的多义性案例属于误读;他称 2023 年在 OpenAI 的自动可解释性工作只找到约 2% 可解读神经元。相关讨论提到 Qwen-32B 中 90% 的神经元只承担一项功能。Neel Nanda 表示这没有改变他的关键判断,仍认为存在表征叠加,弱激活承载记忆事实;Leo Gao 主张用神经元模拟的 KL 指标来评判解释,Pfau 指出论文尚未发表。

    d a hybrid sliding-window design. It was trained with Muon and activation-gradient clipping on about 40T text tokens plus 10T multimodal tokens, using LR and weight-decay scaling laws (Bakouch). Training compute: The run used 16k GB200s over 38 days (scaling01). Disputed comparison: Bakouch and teortaxes note that the bits-per-byte comparison uses MAI’s own validation data (caveat). Gemini 4 Argon evals and rollout: Epoch puts Argon at 165 on its Capabilities Index (ECI), level with Fable 5.1 an

    匹配字段正文

  10. Rohan Paul

    OpenUI 1.0 发布,把生成式 UI 带到任意应用与智能体

    OpenUI 1.0 发布,将生成式 UI 能力从 ChatGPT 扩展到其他应用和智能体,并支持闭源或开源模型,包括本地自托管模型。开发者注册应用自身的 UI 组件(如酒店卡片、地图、图表、预订表单),OpenUI 会将其转换为系统提示词,模型则以 OpenUI Lang 标记语言返回所需组件和数据,渲染器在流式输出中读取并校验后,用应用真实组件绘制界面。该项目 GitHub 已获 11K stars。

    ay use. When a user asks something, the LLM, whether GPT, Claude or a self-hosted open model, replies in OpenUI Lang, a compact markup that names the components the answer needs and fills them with data, instead of writing prose. OpenUI’s renderer reads that markup as it streams, checks it against the registered components, and draws the screen with the app’s real components, so the answer arrives as working interface in the app’s own design system. Check out their GitHub (11K stars) in comment.

    匹配字段正文

  11. 雷锋网

    Anthropic 五年前机密路线图泄露:Coding Agent、Tool Use 与长上下文早已写入

    一场版权官司意外曝光了 Anthropic 五年前由联创 Jared Kaplan 撰写的内部文件《Concise Research Roadmap》,其中已列出 Coding Agent、Tool Use、长上下文、RL 与 Agent 安全等方向。文档批注显示,Dario Amodei 定下短期演示工具使用、长期打造最强编程助手的节奏,Amanda Askell 则强调代码能力与对齐研究必须同步推进。文章随后梳理了 Anthropic 从宪法 AI、100K 上下文、Claude 3 系列到 MCP 协议和 Claude Code 的五年落地过程。

    具体雏形。Dario Amodei 为此定下两步走的落地节奏:短期优先完成能力验证,长期打磨深度生产力工具。团队从打印调试、代码仓库浏览这类轻量化工具能力切入,再逐步拓展到代码执行、自然语言文档编辑等更复杂的工具能力。围绕这个目标,他们把 “规模化” 放在了所有研究方向的最首位,从一开始就为万亿参数规模做准备。Jared Kaplan 是 Scaling Laws 理论的第一作者,他把这套认知从 OpenAI带到了 Anthropic,早早就在思考如何训练万亿(1T)乃至更大规模的模型。除了超大参数模型本身,他们还讨论如何拉高 GPU 利用率,提前适配 TPU、Trainium 等下一代硬件;怎么把块稀疏注意力这类复杂工程封装起来,让非专业人员也能轻松配置超参数;怎么给大模型权重这类核心资产搭好安全防护体系。Anthropic 还非常重视预训练与数据。他们从一开始就打算训出行业性能最优的语言 + 代码双料基座模型,给后面所有的微调、功能探索、能力迭代当好底子。他们早早开始搭建全链路高质量数据集,不仅有自然语言和代码,还覆盖了 Diff(版本差异)、PR(拉取请求)、调试记录以及代码运行

    匹配字段正文

  12. Unite.AI

    Google 详解 Android 短信 Circle to Search 诈骗检测功能

    Google 于 2026 年 10 月 15 日发布消费者指南,说明 Android 用户可用 Circle to Search 长按主页键并圈选可疑短信,由 AI Overview 返回诈骗判断与建议操作。该功能于 2025 年 12 月 2 日加入 Circle to Search 和 Google Lens,全球可用,会比对诈骗模式与网页数据,检查索款话术和链接安全性。Google 称 Phone 与 Messages 中现有的 AI Scam Detection 每月标记超 500 万条可疑诈骗电话和短信。

    s text messages to help identify potential scams. To run the check, a user long-presses the home button or navigation bar on an Android device to activate Circle to Search, then draws a circle around the message in question. The feature then shows an AI Overview with guidance and insights, including suggested next steps; if the results show red flags, the post advises considering deleting the message and blocking the number. A footnote on the post states that results, availability, and compatibi

    匹配字段正文

  13. The New Stack

    Vercel 推出常驻后台 Worker 服务 Daemons,Python 支持与 Drives 存储正式可用

    Vercel 在 Ship 大会上推出常驻后台 Worker 服务 Daemons(公测),支持按区域指定实例数并自动重启,目前面向从队列或数据库拉取任务的 pull 型负载。同时 Vercel 将 Python 支持(FastAPI、Flask、Django)和面向 AI 智能体的持久化存储 Drives 转为正式可用。Drives 将智能体文件与执行沙箱分离,写入落在本地 NVMe 并异步同步至对象存储。

    it’s certainly an instance of a new type of storage product,” Ubl says. He describes it as a hybrid of an EBS-style networked drive and object storage, something he doesn’t think AWS offers. Writes always land on local NVMe disk and stream asynchronously to object storage, he says, and reads come from the local disk when the data is already there — or stream in block by block. “You have a very large number of independent disks, which, because they’re just files in S3, are arbitrarily cheap to o

    匹配字段正文

  14. The New Stack

    黑客劫持 1900 个 GitHub 账号,向 3 万个仓库植入窃密工作流

    GitGuardian 追踪的 GhostAction 攻击中,黑客用窃取的开发者凭据访问约 1900 个 GitHub 账号,向超过 3 万个仓库植入伪装成安全任务的恶意 GitHub Actions 工作流,窃取 CI/CD 密钥、云访问密钥以及 Anthropic、OpenAI、OpenRouter 的 API token。攻击在 10 月 8 日和 9 日集中爆发,波及约 3000 个用户和组织,尝试提取 3174 个密钥,窃取数据被发送至 193.32.204.199。GitHub 的审批控制拦截了大部分恶意工作流,但截至 10 月 5 日,8 月和 9 月受影响的 772 个仓库中仅 124 个完成清理,GitGuardian 建议撤销受影响账号凭据并轮换云密钥与 AI 服务 API key。

    history The October variant went beyond the secrets explicitly referenced by workflows, using actions/checkout@v4 with fetch-depth: 0 to scan the working tree and Git history for AWS keys, GitHub and GitLab PATs, Google and Firebase keys, and API tokens from Anthropic, OpenAI and OpenRouter. Later versions captured surrounding lines to pair AWS key IDs with their corresponding secret access keys. The harvested credentials were sent over HTTP to 193.32.204.199, including keys removed from the cu

    匹配字段正文

星期四

  1. The New Stack

    Linux 基金会 Akrites 如何应对 AI 生成漏洞报告洪流

    Linux 基金会 6 月宣布的 Akrites 项目正通过共享安全事件响应团队(SIRT)汇总、验证漏洞报告并协调上游修复,以应对 AI 生成报告泛滥。Akrites CTO Christopher Robinson 称,审查单个安全 PR 最长耗时 8 小时,这些报告冗长、不准确且高估严重性;其成员包括 Anthropic、OpenAI、Google、Microsoft 等,已投入约 200 名借调工程师。

    The same technology responsible for much of the noise is also becoming increasingly capable of uncovering genuine security weaknesses, including combinations of seemingly minor flaws that can be exploited together. “They’ll take a series of low problems and be able to escalate that into a bigger exploit,” Robinson says. “That is months or a year’s worth of work that the computer can now do in just a couple hours.” Money and minds Akrites has had some serious industry muscle behind it from the g

    匹配字段正文

  2. SiliconANGLE

    SpaceXSI 发布 Grok Voice 3.0,语音到语音指数超过 OpenAI 与 Google

    SpaceXSI(原 xAI)发布语音到语音模型 Grok Voice 3.0,用于其 Voice Agent API。Artificial Analysis 的 Speech-to-Speech Index 给它 86.9 分,高于 GPT-Live-1 的 83.2 和 Gemini 3.8 Live 的 82.6;Full Duplex Bench 得分 97%。在 Tau-Voice 上 OpenAI 模型以 74.5% 略高于 71.6%,但 SpaceXSI 称其首次出声快约 40%。该模型现已在 Voice Agent API 上线,每分钟 8 美分,与 Grok Voice 2.0 同价。

    cloud, cybersecurity and more 11.4k+ theCUBE alumni — Connect with more than 11,400 tech and business leaders shaping the future through a unique trusted-based network Are you an AWS customer? Support SiliconANGLE financially by buying your AWS services from our Marketplace portal page and links: https://siliconangle.com/aws-marketplace/ About SiliconANGLE Media SiliconANGLE Media is a recognized leader in digital media innovation, uniting breakthrough technology, strategic insights and real-ti

    匹配字段正文

  3. SiliconANGLE

    CoreWeave 债务担忧引发 Oracle 及多家 AI 云厂商股价下跌

    受 CoreWeave 债务担忧影响,其股价当日下跌超 15.5%,DigitalOcean、TeraWulf、Hut 8 跌幅均超 10%,Oracle 收跌 8.8%,Dell 和 AMD 分别下跌约 11.3% 和 5.73%。起因是风投人 Chamath Palihapitiya 在 X 上讨论一批 12.5 亿美元 CoreWeave 债券,其收益率一度突破 15%,Elon Musk 回复 Yikes。CoreWeave CEO Mike Intrator 在内部备忘录中称公司积压订单足以覆盖全部债务,6 月该指标为 1040 亿美元,主要来自与 Microsoft、Meta 和 OpenAI 的基础设施协议。

    cloud, cybersecurity and more 11.4k+ theCUBE alumni — Connect with more than 11,400 tech and business leaders shaping the future through a unique trusted-based network Are you an AWS customer? Support SiliconANGLE financially by buying your AWS services from our Marketplace portal page and links: https://siliconangle.com/aws-marketplace/ About SiliconANGLE Media SiliconANGLE Media is a recognized leader in digital media innovation, uniting breakthrough technology, strategic insights and real-ti

    匹配字段正文

  4. SiliconANGLE

    微软探索可组合内存,缓解 AI 基础设施内存短缺

    微软首席架构师 Samir Rajadnya 在 AI Data Pipeline Forum 上表示,可组合内存将内存池化后按需分配给工作负载,是应对 AI 数据中心内存短缺的实用方案。他指出 AI 推理同时受限于功耗和内存,内存离开服务器机箱后,互连技术(包括光学连接)将决定可组合内存能否落地。他建议企业先锁定两三个核心工作负载,再围绕其做工作负载、系统、软件、芯片的垂直整合。

    cloud, cybersecurity and more 11.4k+ theCUBE alumni — Connect with more than 11,400 tech and business leaders shaping the future through a unique trusted-based network Are you an AWS customer? Support SiliconANGLE financially by buying your AWS services from our Marketplace portal page and links: https://siliconangle.com/aws-marketplace/ About SiliconANGLE Media SiliconANGLE Media is a recognized leader in digital media innovation, uniting breakthrough technology, strategic insights and real-ti

    匹配字段正文

  5. SiliconANGLE

    Supermicro:AI 工厂部署应以上线速度而非 GPU 数量衡量

    Supermicro 解决方案管理总监 Ben Lee 表示,AI 工厂部署正从单纯堆 GPU 转向数据视角,公司已与 Vast Data、ScaleFlux 等合作,将存储与计算同步规划,并强调从组件到集群的全栈验证。其 Data Center Building Block Solutions 可扩展至 5 兆瓦到 1 吉瓦,并曾用 120 天让 xAI 的液冷 Colossus-1 集群上线。

    cloud, cybersecurity and more 11.4k+ theCUBE alumni — Connect with more than 11,400 tech and business leaders shaping the future through a unique trusted-based network Are you an AWS customer? Support SiliconANGLE financially by buying your AWS services from our Marketplace portal page and links: https://siliconangle.com/aws-marketplace/ About SiliconANGLE Media SiliconANGLE Media is a recognized leader in digital media innovation, uniting breakthrough technology, strategic insights and real-ti

    匹配字段正文

  6. SiliconANGLE

    开放规范如何加速 AI 数据中心硬件互操作与部署

    三星半导体高级首席工程师 Siamak Tavallaei 在 AI Data Pipeline Forum 上表示,开放规范可把超大规模数据中心集成新硬件设计的时间从约 9 到 12 个月缩短到 3 到 6 个月。他正将 CXL 规范引入 OCP 的可组合内存系统子项目,三星此前已在 OCP 峰会上与 AMD 演示 CXL 内存池化。

    umber of our suppliers could build the same thing,” Tavallaei said. A finished system handed to OCP as a monolithic contribution may never catch on, while one shaped in the open draws input from many companies, according to Tavallaei. That broader buy-in brings volume and more engineers working on the same problems. “If you were to work that with the community, different people will interject their own ideas, and at the end, story will be something that a number of different companies would like

    匹配字段正文