Skip to content
Nov 3Tue
  1. 量子位48

    海马云:模型谁都买得到,拉开差距的是AI生产力系统Chinese

    海马云依托 6 EFLOPS 算力池推出面向 AIGC 的垂直生产力系统 RunningHub,通过“渲推一体”调度将整体算力利用率提升 70%。该平台每月 Token 消耗量达 9 万亿,约 60% 的模型调用由智能动态路由完成;同时推出 RH 系列垂直模型,其中 H3 Lightning 将 5 秒视频生成时间压缩至 28.7 秒。

  2. Hacker News · AI55

    开发者推出基于基准测试的 AI 模型成本优化工具Chinese

    开发者 Jeffery Kline 推出一款交互式成本优化工具,利用数据包络分析(DEA)与自由处置壳(FDH)方法,在基准测试得分不降低的前提下为用户匹配成本更低的单个模型或加权模型组合。该工具以每千次基准任务的预估 API 成本与多项基准得分为依据进行无权衡优化,但不计入延迟、配额和切换等实际部署开销。

  3. Hacker News · AI61

    Why Claude Code Cloud Agents Face Restrictions on GitHub

    In Anthropic's hosted cloud environment, Claude Code isolates real credentials through a dedicated GitHub proxy, strictly limiting access solely to the currently mounted repository and blocking all GraphQL requests.

    Why it matters: The documentation outlines the network and GitHub permission boundaries of the Claude Code cloud sandbox, helping developers troubleshoot why automated scripts may be blocked.

  4. Hacker News · AI27

    生成式 AI 的危害Chinese

    生成式 AI 正通过迎合人脑的惰性与成瘾性弱点,对人类的深度思考与学习能力造成危害。以 Claude 和 Codex 等大语言模型服务为例,它们能直接提供现成的完整方案,使用户跳过自主调研与深度思考的脑力劳动。这种机制虽然能带来即时满足感并消除向他人提问的顾虑,但极易让人形成依赖并落入停止思考的危险陷阱。

  5. Hacker News · AI69

    CPS Prosecutor Cites Fictitious Legal Precedent in High Court Following AI Search

    A senior prosecutor from the UK's Crown Prosecution Service (CPS) cited entirely fabricated case law during a High Court hearing after relying on an AI-generated web search summary. In a disputed case involving the prosecution of a child, the prosecutor submitted a non-existent case citation and later admitted it was a citation error. The CPS stated that it has launched an internal review and reiterated that legal precedents must be cited strictly from trusted sources. The incident brings the total number of tracked court-related AI errors and bogus citations in the UK to 101.

    Why it matters: The report reveals details of a state prosecution service citing fictitious precedents in court due to reliance on web-based AI summaries, highlighting the tangible risks as AI tools make their way into judicial proceedings.

  6. The Decoder79

    Huawei Asks Clients to Source Their Own Memory Amid HBM Shortage, Driving Up Black-Market Sourcing and Smuggling Premiums

    Unable to secure sufficient High Bandwidth Memory (HBM) for AI chips such as the Ascend 960DT planned for early next year, Huawei has asked some customers to procure their own memory, driving buyers to black-market channels. Restricted by U.S. export controls, smuggling rings bundle HBM with low-end server processors, ship them to Hong Kong, and dismantle them there. This has driven the price of smuggled HBM3E chips up to roughly $4,000 each, pushing memory costs for a single AI chip to as much as $32,000.

    Why it matters: The report highlights the severe supply chain distortions and extreme price premiums faced by downstream buyers of high-end computing chips due to memory shortages amid export restrictions and limited domestic manufacturing capacity.

  7. Hacker News · AI54

    AI 或将通过自动挖掘隐藏漏洞冲击法律体系Chinese

    斯坦福法学院访问学者 Felix Ringe 等人提出,大语言模型已具备规模化识别法律条文不一致与漏洞的能力,可能像挖掘软件漏洞一样冲击现有法律体系。研究者以德国劳动法规中缺失授权引用等漏洞为例,指出廉价法律推理技术的普及或将迫使立法者借助 AI 强化法案审查;受访法律专家认为该方法逻辑合理,但漏洞有效性仍需严格的人工复核。

  8. Hacker News · AI50

    开源工具 Recursive Human Intelligence 发布,可分析大模型聊天记录并提供反馈Chinese

    开源桌面工具 Recursive Human Intelligence 发布,可读取用户与大模型的历史提问记录并通过模型生成直白的改进反馈。该工具支持自带 OpenAI、Claude 或 OpenRouter 的 API 密钥,在本地过滤系统上下文并对提问模式进行聚类。分析过程会对凭据进行脱敏,单次请求最多采样 80 条用户摘录,且每条反馈必须引用至少两条真实历史消息作为依据。

  9. Hacker News · AI61

    When Data Changes After Human Correction: Zingg Reviews System Design Boundaries for Human-in-the-Loop Decision-Making

    Open-source entity resolution platform Zingg has reviewed the system design challenges of human-in-the-loop decision-making, pointing out that when underlying data changes after human corrections to AI, the system must clearly define the lock scope and permission boundaries of human decisions. The team proposed that while systems may add new records to human-verified entity clusters, they must strictly prohibit automatic splitting or merging, treating grouping relationships rather than individual data records as the unit of approval. The article emphasized that model confidence scores must never dictate when human decisions expire; instead, triggers for re-review must be governed by external rules or business events.

    Why it matters: The article examines the permission dilemmas that arise when underlying data changes in human-in-the-loop decision-making, providing engineering design principles for defining the granularity of human decisions and mechanisms for triggering re-evaluations.

  10. Azure Blog · 云基础设施20

    微软 Azure:支撑下一次 AI 突破的底层平台Chinese

    微软 Azure 指出,AI 智能体正在承担长周期、动态且不可预测的复杂任务,要求芯片、算力、存储与网络作为单一系统进行协同设计和一体化优化。以 Levi Strauss & Co. 为例,该企业已将数据中心迁移至 Azure,并正在构建统一的“超级智能体”以协调超过 1,000 个特定领域的智能体。

  11. CNCF Blog68

    OpenKruise Agents v0.6.0 Released to Strengthen Agent Sandbox Security Governance and Automated Operations

    CNCF open-source project OpenKruise Agents has officially released v0.6.0, building a data plane and control architecture comprising six components and 12 CRDs to comprehensively enhance security and operational efficiency when running agent sandboxes at scale on Kubernetes.

    Why it matters: This release extends security governance to the request and tool-calling layers, providing teams running agent sandboxes at scale with a practical reference that balances credential isolation and cost optimization.

  12. Hacker News · AI28

    Flow AI Fx:涵盖视频、图像、音乐与语音的统一 AI 创意工作室Chinese

    Flow AI 是一个整合视频、图像、音乐与语音的 AI 创意工作室,帮助用户将灵感快速转化为创意初稿。平台提供 Flow AI Video、Flow AI Image、Flow AI Music 及 Flow AI Voice 工具,支持平滑运镜控制、图像生成、全曲创作与多语言旁白。用户通过日常语言提示词经四步流程即可完成调整并导出媒体文件。

  13. C114 通信网 · AI与算力(网页)31

    中国电信荣获2026年世界互联网大会社会贡献类杰出贡献奖Chinese

    中国电信在2026年世界互联网大会乌镇峰会上荣获社会贡献类杰出贡献奖。其在京津冀、长三角等地打造了四个全国产化液冷万卡智算集群,“息壤”平台调度算力规模超119EFLOPS。此外,公司构建星辰TokenHub汇聚140多个大模型,并发布自研星辰超级智能体TeleAgent及融合QKD与PQC的分布式密码体系。

  14. C114 通信网 · AI与算力(网页)35

    中国电信“城域云网关键技术研发及应用”项目荣获2026世界互联网大会领先科技奖Chinese

    中国电信“城域云网关键技术研发及应用”项目在2026年世界互联网大会乌镇峰会上荣获领先科技奖。该成果首创积木式架构与扁平化全互联组网模式,规模部署云化宽带接入控制系统,使服务效率提升50%,并获得222项授权专利。目前该技术已在30多个国家规模推广,赋能20多个行业并服务超3亿用户。

  15. Hacker News · AI56

    智能体落地真正风险在于脱离业务主线Chinese

    企业应用 AI 智能体的核心风险不在于技术落后,而在于脱离业务流程与客户价值盲目推进自动化,导致效率无法转化为利润。综合麦肯锡、MIT 与 Gartner 等多项调研及 Klarna 案例,大多数试点停滞的原因是工具未融入真实工作流、且过度偏向单纯降本而损害了交付质量;企业应当先梳理核心盈利流程与评估指标,建立治理边界,最后再针对明确环节选型智能体。

  16. Hacker News · AI51

    Sentrint 推出针对 AI 生成代码的安全扫描工具Chinese

    Sentrint 发布针对 AI 生成代码的安全扫描工具,支持检测由 Claude、Cursor、Bolt、v0 等 16 种 AI 编程工具构建的 GitHub 仓库。该工具通过静态规则与模型复核检查硬编码密钥、数据库访问规则、漏洞依赖与高危路径,并输出 100 分制评分及适配原工具的修复提示词。产品还提供 MCP 接口,支持在 AI 智能体对话中直接触发扫描并定位修改代码行。

  17. The Decoder56

    OpenAI 将漏洞赏金最高提升至 17.2 万美元并启动安全冲刺Chinese

    OpenAI 将漏洞赏金计划奖励提高至四倍,外部研究人员发现 ChatGPT 或 Codex 等核心产品的严重漏洞最高可获 $30,000,叠加忠诚度与活动奖金后单笔最高可达约 $172,000。新方案新增了五级忠诚度计划以及与新功能上线挂钩的限时奖励,并对过去 30 天内提交的漏洞追溯补齐差额。同时 OpenAI 正联手安全公司 Cathedral 展开安全冲刺,对其系统进行压力测试。

  18. MIT Technology Review AI53

    逆向博弈论如何约束 AI 智能体Chinese

    AI 安全研究人员正尝试引入源自经济学的“逆向博弈论”(机制设计),通过构建激励规则而非探究黑盒内部运作来约束 AI 智能体行为。该方案提出利用模型互相交叉预测答案、解题者与审计者角色制衡等机制,防范模型在安全测试中故意隐藏能力或装傻欺骗。不过,该方法在实际落地时仍面临大模型是否符合理性假设、智能体之间可能串通作弊以及市场竞争激励等多重挑战。

  19. Hacker News · AI71

    Reviewer-gates: Enforcing Claude's Code Quality with an Independent Reviewer Agent

    The author proposes a code quality control mechanism called reviewer-gates, which uses a dedicated reviewer agent to intercept and review code generated by developer agents against Markdown rule files.

    Why it matters: It offers a workflow concept for enforcing code quality checks via an independent reviewer agent, helping reduce the overhead of repeatedly reviewing AI-generated code in long-term maintained codebases.

  20. Hacker News · AI66

    Trim open-sources local proxy tool to compress prompts before sending them to LLMs

    Trim has introduced a local proxy for AI coding tools that cuts context size by compressing prompts before requests reach upstream large language models. Featuring a unified, OpenAI-compatible local interface, it integrates directly with tools such as Cursor, Continue, VS Code Chat, and Claude Code, and supports both a fast AST-based compression mode and a deep mode running LLMLingua entirely locally.

    Why it matters: By executing heuristic and model-based compression at the local proxy layer, the project provides a drop-in solution to reduce context token overhead for AI coding assistants.

  21. Hacker News · AI57

    MIT 等机构推出 ImpactBench 与 AI 营养标签,量化评估模型对人类福祉的影响Chinese

    MIT Media Lab 联合南加大、UC 伯克利等机构推出开源评测框架 ImpactBench 及其可视化工具 AI 营养标签,用于衡量 AI 系统对人类心理、自主性与健康福祉的影响。该框架通过 800 项指标和 26 个基准评估前沿模型,涵盖幻觉、谄媚倾向与用户自主权等维度。初步测试显示各模型在规避伤害方面优于促进积极赋能,其中对用户学习与技能发展的支持得分普遍最低。

  22. Hacker News · AI10

    数据科学与 AI 免费奖学金测试Chinese

    数据科学与 AI(Data Science with AI)相关项目公布了一项免费奖学金测试。原始信息目前仅公布了测试主题,未包含具体的考试大纲、参与规则或申请条件等详细内容。