←搜索结果

按匹配度 · 共 1137 条结果 清除搜索

星期四

  1. Johann Rehberger / Embrace The Red

    OpenAI 发布论文详解 URL 数据外泄缓解措施

    OpenAI 发布论文《Preventing URL-Based Data Exfiltration in Language-Model Agents》,详细说明其新增的 URL 数据外泄缓解措施。作者 Johann Rehberger 指出,OpenAI 现在用自己的网页索引做 URL 校验,爬虫收录的 URL 会进入允许列表,用户消息中直接输入的 URL 也被视为安全,但模型动态生成的 URL 会被判定为不安全。他同时说明该问题并未完全解决,用单个 URL 映射字母和数字的绕过方式仍然可行,并建议对同一会话内重复访问同一 URL 做缓存或限制。

    have a bug bounty program, so communication was via email, and unfortunately there was little traction or appetite to fix the problem in ChatGPT. I also reported the same issue to Microsoft as Bing Chat was impacted, and Microsoft applied a fix (via a Content-Security-Policy header) in May 2023 to generally prevent loading of images. Since then, dozens of such vulnerabilities across the AI ecosystem, being one of the most common AI application security vulnerabilities. Tackling the Problem In De

    匹配字段正文

星期四

  1. VentureBeat AI

    Railway 完成 1 亿美元 B 轮融资,以 AI 原生云基础设施挑战 AWS

    旧金山云平台 Railway 宣布完成 1 亿美元 B 轮融资,由 TQ Ventures 领投,FPV Ventures、Redpoint 和 Unusual Ventures 参投,此前累计融资仅 2400 万美元。Railway 称其平台部署时间低于 1 秒,按秒计费,内存 0.00000386 美元/GB-秒、vCPU 0.00000772 美元/vCPU-秒,客户报告开发速度提升 10 倍、成本最高节省 65%。公司目前每月处理超 1000 万次部署,团队仅 30 人,并已于 2025 年 8 月发布 Model Context Protocol 服务器,让 AI 编码智能体可直接从编辑器部署应用。

    se support with SLOs ($2,000), and dedicated virtual machines ($10,000). Railway enters a crowded market that includes not only the hyperscale cloud providers—Amazon Web Services, Microsoft Azure, and Google Cloud Platform—but also a growing cohort of developer-focused platforms like Vercel, Render, Fly.io, and Heroku. Cooper argues that Railway's competitors fall into two camps, neither of which has fully committed to the new infrastructure model that AI demands. "The hyperscalers have two comp

    匹配字段正文

星期四

  1. Anthropic:The Institute

    Anthropic 发布第四份经济指数报告:新增五项 AI 使用基础度量

    Anthropic 发布第四份 Economic Index 报告,基于 2025 年 11 月 Opus 4.5 发布前的 100 万条 Claude.ai 对话和 100 万条 1P API 记录,新增任务复杂度、人类与 AI 技能、使用场景、AI 自主性、任务成功率五项经济基础度量。报告显示 Claude.ai 上增强型使用回升至 52%、自动化降至 45%,前十项任务占 24% 对话,计算机与数学类任务占 Claude.ai 的 34%、API 的 46%。美国各州人均使用差距缩小,基尼系数从 0.37 降至 0.32,若延续则 2 至 5 年可趋于均衡;全球使用仍与人均 GDP 高度相关且差距稳定。

    Figure 1.1: Usage shares among top 10 tasks over time by platform, Claude.ai and 1P API.

    common in higher-income countries, while coursework use cases are more common in lower-income countries. This echoes findings from our prior report and aligns with recent work by Microsoft. GDP and human education predict adoption globally and within the US: A 1% increase in GDP per capita is associated with a 0.7% increase in Claude usage per capita at the country level. Human education—Claude's estimate of years of formal education needed to understand the human prompt—correlates positively w

    匹配字段正文

星期二

  1. LG AI Research 官方博客

    LG AI Research:从工具到同事,扩展人类判断力的 AI UX 设计

    LG AI Research 在 2025 AI Ethics Seminar 第四期中提出,AI 正从专用工具变为工作基础设施,UX 设计应让 AI 成为“同事”而非“工具”。文章引用 Kahneman 双系统理论与 Bruner 叙事自我理论,并援引 35 个人机决策案例及 ECG、放射科、150 名高管实验,说明自动化偏见会削弱人类判断。

    HJK_c3104e5c1.png

    are overly simplistic can induce overreliance, while excessively complex explanations increase cognitive load, thereby reducing users' ability to evaluate AI outputs effectively. Microsoft Copilot in Excel's "show explanation" feature exemplifies a well-balanced approach to reconciling simplicity and detail. At the formula suggestion stage, it first presents a concise result, then allows users to selectively access detailed rationale when needed; Promoting verification engagement: System design

    匹配字段正文

星期二

  1. Qwen Blog

    Qwen3-TTS 推出音色设计与音色克隆两个新模型

    Qwen3-TTS 家族发布两个新模型:音色设计模型 Qwen3-TTS-VD-Flash 和音色克隆模型 Qwen3-TTS-VC-Flash,均可通过 Qwen API 调用。VD-Flash 支持用自然语言指令控制音色、韵律、情感与角色设定,在 InstructTTS-Eval 上整体优于 GPT-4o-mini-tts 和 Mimo-audio-7b-instruct,并在角色扮演测试中超过 Gemini-2.5-pro-preview-tts。VC-Flash 支持 3 秒音色克隆,可基于克隆音色生成中文、英语、德语、意大利语、葡萄牙语、西班牙语、日语、韩语、法语和俄语 10 种语言语音,在 MiniMax TTS Multilingual Test Set 上平均词错误率优于 MiniMax、ElevenLabs 和 GPT-4o-Audio-Preview。

    nstall pyaudio Debian/Ubuntu sudo apt-get install python-pyaudio python3-pyaudio or pip install pyaudio CentOS sudo yum install -y portaudio portaudio-devel && pip install pyaudio Microsoft Windows python -m pip install pyaudio import pyaudio import os import requests import base64 import pathlib import threading import time import dashscope The DashScope Python SDK version must be 1.23.9 or later. from dashscope.audio.qwen tts realtime import QwenTtsRealtime , QwenTtsRealtimeCallback , AudioFor

    匹配字段正文

星期二

  1. DevClass

    Cursor AI 编辑器新增可视化网页设计器,但 bug 与频繁改动的 UI 引发开发者不满

    Anysphere 的 Cursor AI 编辑器上周发布 2.2 版本,新增可视化网页编辑器(浏览器侧边栏内可拖动、调整元素并让 AI 智能体改代码)和 debug 模式。开发者欢迎新功能,但抱怨 bug、UI 每周变动以及订阅成本偏高,有团队因此取消订阅;Cursor 工程师 Andrew Milich 称将尽量减少 UI 改动。

    ware development. When you just need to build a tiny desktop app: TinyJS October 13, 2026 Valen creator drives 'Golden Spike' to connect new languages with Rust September 26, 2026 Microsoft agentically ports Copilot runtime to Rust for $120K September 19, 2026 Apple iPhone Duo makes developers think in folds September 16, 2026 JavaScript installer pnpm recast in Rust because ECMAScript can't keep up September 5, 2026 Kubernetes cleans house, bins legacy kube-dns, IPVS, and cgroup v1 August 27, 2

    匹配字段正文

星期四

  1. DevClass

    VS Code 更新加入 Agent HQ 与 10 项智能体功能,IntelliCode 被弃用

    微软更新 Visual Studio Code,加入 Agent HQ 与 10 项智能体功能,包括聊天关闭后智能体继续运行、会话从本地迁移到云端、自定义后台智能体、运行自定义子智能体和组织内共享智能体。更新还提供 TypeScript 7 的实验性开关,需安装并配置 TypeScript Go 后用于 TypeScript 和 JavaScript 语言功能。提供本地 AI 补全的 IntelliCode 扩展被弃用,AI 辅助补全改由 GitHub Copilot 承担,Copilot 每月提供 2000 次免费补全,超出后需订阅。

    AI/ML Microsoft has updated Visual Studio Code (VS Code), the most popular programmer’s editor, with Agent HQ alongside numerous agentic coding updates, a preview of TypeScript 7 support, and deprecation of IntelliCode code completion in favor of subscription-based GitHub Copilot. Agentic coding is the hot thing in AI-assisted development, with the formation of the just-announced Agentic AI Foundation and recent releases from VS Code competitors including Google’s Antigravity and JetBrains Air i

    匹配字段正文

  2. DevClass

    研究:AI 对软件开发既助力也拖累,传统最佳实践才是关键

    Atlassian 旗下 DX 的数据显示,AI 对开发者生产力的影响因公司而异,部分企业变更信心提升 20% 以上,另一些则明显下滑。DX 副 CEO Justin Reock 在 AWS re:Invent 上指出,用好 AI 的关键仍是 SDLC 与代码卫生等 AI 出现前就适用的因素;仅用 AI 代码补全每周可省约 3.8 小时,Go 开发者可省 4 小时,COBOL 仅 2 小时。

    Invent attendees. When you just need to build a tiny desktop app: TinyJS October 13, 2026 Valen creator drives 'Golden Spike' to connect new languages with Rust September 26, 2026 Microsoft agentically ports Copilot runtime to Rust for $120K September 19, 2026 Apple iPhone Duo makes developers think in folds September 16, 2026 JavaScript installer pnpm recast in Rust because ECMAScript can't keep up September 5, 2026 Kubernetes cleans house, bins legacy kube-dns, IPVS, and cgroup v1 August 27, 2

    匹配字段正文

星期二

  1. METR Blog

    前沿 AI 安全政策的共同要素(2025 年 12 月更新)——METR

    METR 发布《前沿 AI 安全政策的共同要素》2025 年 12 月更新版,目前已有 Anthropic、OpenAI、Google DeepMind、Meta、xAI、Nvidia 等 12 家公司发布前沿 AI 安全政策。报告梳理了各政策中的能力阈值、模型评估、模型权重安全与部署缓解措施,并新增对 EU AI Act 通用 AI 行为准则和加州 Senate Bill 53 相关指引的引用。

    nal four companies joining since then. Currently, twelve companies have published frontier AI safety policies: Anthropic, OpenAI, Google DeepMind, Magic, Naver, Meta, G42, Cohere, Microsoft, Amazon, xAI, and Nvidia. In August 2024, we released a version of this report describing commonalities among Anthropic, OpenAI, and Google DeepMind’s frontier safety policies. In March 2025, with twelve policies available, this document updated and expanded upon that earlier work. This December 2025 version

    匹配字段正文

星期二

  1. Johann Rehberger / Embrace The Red

    Google Antigravity IDE 被曝五类安全漏洞,含远程命令执行与数据外泄

    安全研究者 Johann Rehberger 在 Google 新发布的 Antigravity IDE 中发现五个安全漏洞,包括通过间接提示注入实现的远程命令执行和数据外泄,其中部分问题可追溯到 2025 年 5 月他向 Windsurf 报告过的同类缺陷。Antigravity 默认由 AI 自行判断终端命令是否安全执行,且 MCP 工具调用完全没有人工确认环节,攻击者可利用不可见的 Unicode Tag 字符在源码或 Linear 工单中隐藏指令。作者建议禁用自动执行、谨慎启用 MCP 服务器,并指出 Claude Code、GitHub Copilot、Cursor、Codex 等编码智能体在安全性和补丁历史上更成熟。

    e capabilities of the tool, this allows the AI to exfiltrate data, code execution, data manipulation or deletion, etc., all without the developer’s consent. An interesting feature Microsoft recently added to GitHub Copilot is that the result of an MCP tool is actually displayed to the developer and the developer can decide to include the data or not in the prompt context. It is possible to disable individual tools, which is good. But there is no secure way to enable dangerous tools at the moment

    匹配字段正文

星期二

  1. antirez 原创博客

    Scaling HNSWs:antirez 谈 Redis 中 HNSW 的内存与性能优化

    antirez 在暂停 HNSWs 开发之际,分享了为 Redis 实现 HNSW 向量相似度搜索时的高级优化经验。他采用按向量量化的有符号 8 位表示(-127 到 127)来压缩向量,并指出多层结构平均仅带来约 1.3 倍开销,同时质疑 HNSW 中“H”层级是否真正必要。

    great area is to imagine evolutions and refinements of HNSWs, without getting trapped within the idea that the evolutions are only in the sense of: let’s do it, but for disk (see Microsoft efforts), or the like. Ok, enough with the premise, let’s go to the actual low level stuff :) Scaling memory Redis is an in-memory system, and both HNSWs and vectors have the unfortunate quality of being very space-hungry. There are three reasons for this: 1. HNSWs have a lot of pointers, like 16, 32 or more

    匹配字段正文

星期四

  1. LG AI Research 官方博客

    LG AI Research 谈 AI 如何改变人类批判性思维

    LG AI Research 在 2025 AI Ethics Seminar 中讨论了生成式 AI 对人类批判性思维的影响,并介绍了 ChatEXAONE 如何将这些变化纳入设计理念。文章引用 Lee et al.(2025)对 319 名知识工作者的研究,指出对 AI 依赖度越高,独立思考投入越少;Passi et al.(2024)则提出 CAIR、CSR 与 AoR 指标及三项 AI 设计策略。

    HDS_b02f7e961.png

    into its design philosophy. 1. The evolution of critical thinking: How much should we trust AI? In a paper by Lee et al. (2025)[1], researchers from Carnegie Mellon University and Microsoft analyzed changes in the critical thinking of 319 knowledge workers when using generative AI. The results were noteworthy. The higher the reliance in AI, the less effort was made to think independently; Those who had high confidence in their own judgment maintained critical thinking even when using AI The rese

    匹配字段正文

星期二

  1. Cognition 模型 / Devin 博客

    Cognition 用 Devin 自动化 .NET Framework 到 .NET Core 迁移

    Cognition 将 Jimmy Bogard 的 .NET 迁移指南改编为面向 AI 智能体的版本,展示 Devin 如何自动化 Strangler Fig 迁移的每个阶段。Devin 通过 DeepWiki 生成架构文档、Ask Devin 制定迁移计划、本地环境执行迁移、内置测试框架验证每个 PR,原本耗时数月的迁移最快可在两周内完成。

    with .NET Core. For each session key, determine its data type, serialization requirements, and usage patterns.” Next, use a Devin Session to register each key and serializer. "Add Microsoft.AspNetCore.SystemWebAdapters to the Global.asax file on the .NET side and configure session state sharing. Register all session keys with JSON serialization, set up the remote app server with proper API key configuration, and add the session server. For each session variable in the codebase, identify and regi

    匹配字段正文

星期三

  1. vLLM 官方博客

    vLLM OpenAI 兼容 API 支持返回 token IDs,消除智能体强化学习的重新分词偏移

    vLLM 从 v0.10.2 起为 OpenAI 兼容 API 增加 `return_token_ids` 参数,可随文本返回输入的 `prompt_token_ids` 和生成结果的 `token_ids`。该功能避免推理文本在训练时重新分词产生不同 token 序列,Agent Lightning v0.2 会自动启用参数并通过追踪能力收集训练所需数据。

    gent with RL. It has several core features. Seamless integration with existing agents with ZERO CODE CHANGE (almost)!; Build with ANY agent framework (LangChain, OpenAI Agent SDK, Microsoft Agent Framework, ...); or even WITHOUT agent framework (Python programs); No constraints on the input to the LLM, allowing for flexible orchestration such as summarization, multi-agent collaboration, and other complex workflows When Agent Lightning was first released, we implemented an instrumented vLLM serve

    匹配字段正文

星期一

  1. One Useful Thing

    Ethan Mollick:当下使用 AI 的一份有主见指南

    Ethan Mollick 发布了一份当下使用 AI 的有主见指南,建议按 OpenAI 公布的真实使用数据判断何时用免费模型、何时升级付费。他指出最先进的四套系统是 Claude、Gemini、ChatGPT 和 Grok,付费档为每月 20 美元或约 200 美元,前者适合大多数人,后者面向复杂技术与编程需求。做正经工作时建议手动选择 Agent 模型,例如 GPT-5 Thinking Extended 或 GPT-5 Thinking Heavy。

    、OpenAI 的 ChatGPT ,以及 Elon Musk 的 xAI 旗下的 Grok。然后是开放权重 AI 家族,它们几乎(但并非完全)同样出色:中国的 Deepseek 、Kimi 、Z 和 Qwen,以及法国的 Mistral。在几乎任何 AI 评级系统中,这些 AI 模型的变体合计占据了前 35 名。你使用的任何其他提供前沿 AI 的服务,从 Microsoft Copilot 到 Perplexity(两者都提供一定的免费使用),都是以这九种 AI 中的一个或多个为基础驱动的。 你应该如何在它们之间做选择?一些免费系统(如 Gemini 和 Perplexity)擅长网络搜索,而另一些则完全无法搜索网络。如果你想要免费的图像生成,最佳选择是 Gemini,ChatGPT 和 Grok 紧随其后。但归根结底,这些 AI 在许多细微之处存在差异,包括隐私政策、访问级别、能力、它们处理伦理问题的方式,以及“个性”。而所有这些都会随时间波动。因此,基于这些因素选择一个你喜欢的模型并使用它。不过,如果你正在考虑可能升级到付费账户,我建议从 Anthropic、Google 或 O

    匹配字段正文

星期二

  1. Microsoft AI Blog精选

    微软发布首款自研图像生成模型 MAI-Image-1,进入 LMArena 前十

    微软 AI 发布首款完全自研的图像生成模型 MAI-Image-1,在 LMArena 文本生图榜单中进入前十。该模型主打照片级真实感,擅长光照(如反弹光、反射)和风景等场景,官方称其速度与质量组合优于许多更大更慢的模型。11 月 4 日起,MAI-Image-1 已开始接入部分微软产品:在 Bing Image Creator 中与 DALL-E 3、GPT4o 并列可选,并用于 Copilot Audio Expressions 的 Story Mode 配图,目前面向所有可访问 Bing Image Creator 和 Copilot Labs 的国家开放。

    A collage featuring nature photos (a frog in water, a rabbit, mountains, fields, sunset, a tree) and food photos (pizza, sushi, a grapefruit slice), arranged in a grid on a light beige background.

    LI; X; FB Update – November 4, 2025: We have begun launching MAI-Image-1 into select Microsoft products! Try it in Bing Image Creator: Available at bing.com/create, in the Bing mobile app, or right from the Bing search bar, Bing Image Creator is built to meet people where they already search and create. MAI-Image-1 is now an option alongside DALL-E 3 and GPT4o in the model menu, enabling you to experiment and pick the model that best matches your creative goals. Try it in Copilot Audio Expressio

    匹配字段正文

星期三

  1. TLDR AI

    Gemini 登陆 Chrome,可快速总结任意网页

    Google 正在面向美国 Mac 和 Windows 桌面用户推出 Chrome 版 Gemini,用户可就正在阅读的任意网页请求解释和摘要。企业版将在未来几周通过 Workspace 提供,并配备企业控制功能。

    hy it matters: Leaders can standardize a light-lift reading assistant across the org to speed brief creation and reduce time spent parsing long pages. Source: Google Blog Summary: Microsoft’s latest Windows 11 Insider build adds a taskbar button that shares the active app directly with Copilot to analyze and summarize what is on screen. A “Click To Do” gesture also offers instant translations when text in another language is detected. Why it matters: This reduces friction to use Copilot in the f

    匹配字段正文

星期二

  1. TLDR AI

    AI 原生办公套件:AI 如何替你完成工作

    Andreessen Horowitz 测试了一批面向常见商业任务的智能体 AI 工具,发现通用工具更灵活,而 Gamma 等演示文稿工具和 Serif 等邮件工具能为对外工作提供质量更高、更完善的成果。企业应根据具体任务,在通用平台与专业智能体之间进行选择。

    nd welcome to TL;DR AI . Each week, I sort through the noise to bring you a curated digest of important AI developments and practical advice for business leaders. In this edition: Microsoft signals the shift toward collaborative, team-based AI in the workplace HR Tech and Adobe highlight orchestration platforms that scale automation across functions Personal AI assistants evolve into full workflow partners for individuals and enterprises The AI-Native Office Suite: How AI Can Do Work For You Sum

    匹配字段正文

星期二

  1. TLDR AI

    Google Gemini 提升日常工作效率

    Google Workspace 为 Gemini 增加会议摘要、文档起草和数据洞察功能,帮助团队减少日常事务性工作。简报还介绍了 Salesforce 用 AI 智能体处理一半客户互动并缩减约 4,000 个客服岗位,以及 Capgemini 关于六成组织预计未来 12 个月内将 AI 作为团队成员的调查结果。

    led to a 17% reduction in support costs since early 2025. Benioff stated the company is now reviewing every function for similar automation opportunities, echoing similar moves by Microsoft and Klarna. Why it matters: This is one of the clearest signals yet of AI-driven job displacement at a major enterprise, validating the shift from AI as a productivity tool to an outright workforce replacement for specific operational functions. Source: SHRM Summary: Microsoft has partnered with the US Genera

    匹配字段正文

星期四

  1. SemiAnalysis精选

    Amazon AI 复苏:AWS 与 Anthropic 扩建多吉瓦级 Trainium 基础设施

    AWS 正为 Anthropic 扩建多吉瓦级 AI 训练基础设施,三个接近完工的园区合计拥有逾 1.3GW IT 容量,最大园区将部署近百万颗 Trainium2。Trainium2 的多项性能落后于 NVIDIA 系统,但在单位内存带宽总拥有成本上具有优势,契合 Anthropic 侧重强化学习和训练后扩展的路线。Anthropic 还深度参与 Trainium 的系统设计,同时继续使用 Google TPU 和 NVIDIA GPU。

    he Amazon empire, generating 60% of group profits, and dominating the lucrative Cloud Computing market. But it struggles to translate this strength into the new GPU/XPU Cloud era. Microsoft Azure now leads the market on quarterly new cloud revenue, and the gap between Google Cloud and AWS has materially narrowed especially with Google's big moves on the TPU that we've been posting about for over a month. Markets have noticed. Year-to-date, Amazon is the clear laggard among the four tech-and-AI t

    匹配字段正文