苹果起诉OpenAI窃密与GPT-5.6破解数学难题领衔今日AI动态

SlimeBallBench · AI models play slime soccer

SlimeBallBench 是一个用于让 AI 模型进行史莱姆足球（slime soccer）对决的基准测试平台。该网站提供了比赛视频、积分榜、统计数据以及平台的工作原理介绍。

面向 Qwen 系列模型线性注意力的高性能优化实践|AICon深圳

Political Neutrality Benchmark of popular AI models

Results · Release 01 What the models reveal. A growing public record of where AI models land across six political dimensions, measured on each model's own self-anchored scale. These are the first runs, with more to come. 18models so far 6an

Inferring multicellular interactions in tumors from standard pathology slides

斯坦福医学院的研究人员开发了一种名为 CANVAS 的 AI 平台，能够直接通过常规的 H&E 染色病理切片预测肿瘤组织中的多细胞“邻域”（即癌细胞、免疫细胞与基质细胞之间的空间交互和分子沟通）。CANVAS 基于预训练医学图像模型 MUSK 构建，并结合了 CODEX 空间多组学技术的数据进行训练。在对 9 种癌症类型、超过 5000 名患者的存档切片进行测试后，该平台成功识别出 10 种独特的细胞邻域，其中某些特定邻域（如富含特定中性粒细胞的区域）与患者预后较差和免疫治疗耐药性显著相关。该研究已发表于《Cell》杂志。

RT Rohan Paul: Most VLAs (Vision-Language-Action Models) handle task variety only inside narrow, fixed bodies; LingBot-VLA 2.0 from @robbyant_brain tr...

Robbyant Brain 推出了具身智能大模型 LingBot-VLA 2.0，支持全身自由度控制。该模型通过统一的 55 维动作格式，可在 20 种不同的机器人配置（包括机械臂、灵巧手、头部、腰部及移动底座）上训练单一策略。研发团队对数据集进行了严格过滤，将 90,000 小时原始数据筛选为 50,000 小时高质量真实机器人数据。架构上，它采用了稀疏混合专家（MoE）模块，并通过预测当前与未来的深度和视频特征来追踪物体几何与场景变化。在 Agilex GM-100 上的测试显示，它达到了 66.2% 的任务进度和 34.4% 的成功率，性能超越了 pi0.5。

Meta's Muse Spark 1.1 's multimodal reasoning ability. Here it took videos from mobile, reaons to extract photos & product details then operates the b...

Meta 展示了 Muse Spark 1.1 的多模态推理能力。该模型能够处理手机拍摄的视频，从中推理并提取出图片和产品详情，随后通过计算机使用（Computer Use）功能操作浏览器，自动填写 Marketplace 表单、上传图片并发布商品信息。

OpenAI's GPT-5.6 Sol Ultra reportedly solves a 50-year-old math problem in under an hour

OpenAI 的 GPT-5.6 Sol Ultra 模型通过 64 个并行子智能体，在不到一小时内证明了拥有 50 年历史的“双圈覆盖猜想”（Cycle Double Cover Conjecture）。数学家 Thomas Bloom 指出，该证明方法初等且巧妙，但批评其未引用 1983 年前人论文的思路。该成果的实现得益于人类提示词的精心设计：强制模型假设证明存在、禁止搜索现成答案、设定最少计算时间，并利用对抗性智能体进行严格验证。

Show HN: AI Photo Editor – Professional-Grade Image Editing with Text Prompts

AI Photo Editor 是一款在线 AI 图像编辑工具，基于 Nano Banana 和 GPT Image 2 模型构建。用户只需输入简单的文本提示词即可进行专业级的图像转换与编辑。该工具支持面部修复/补全、跨图角色一致性编辑等功能，宣称生成速度可达 1 秒以内，且在效果与速度上优于 Flux Kontext。

Show HN: TrialPilot – clinical trials from your phone, built by a patient

I have Long COVID, five years and counting.I sit on the NIH RECOVER and RECOVER-TLC working groups where I watched how research on invisible illness actually gets done and who gets left out of it.I've also helped design one of the largest Long COVID clinical trials about to launch.And I've spent five years inside a patient community that knows exactly where the system fails.So I built TrialPilot with two engineers I've worked with for a decade. Clinical trials that come to you instead of asking you to come to them.https://www.trialpilot.app/We have partnerships lined up to launch a handful of pilot studies over the next few months. We're also competing in the HHS TOPx Tech Sprint for AI and Invisible Illness.Most of you have probably never run a clinical trial. Most people haven't. What you do know is software, and how it feels when software is good or bad, and how a sick person or researcher on the other end of it might feel.First impressions and honest feedback welcome. Reply here or send me an email (on the webpage). I read everything.Thanks for checking it out!

Claude Code now has a built-in browser that lets the AI read, click, and type on external websites

Claude Code now has a built-in browser that lets the AI open, read, and interact with web pages directly inside the development environment. Write actions on external sites are screened by classifiers, and purchases or account creations need user approval. The article Claude Code now has a built-in browser that lets the AI read, click, and type on external websites appeared first on The Decoder.

Linux of AI open-source tools for reducing AI vendor lock-in

Linux of AI 是一个旨在减少 AI 厂商锁定的开源项目生态系统，为开发者提供可移植、可审查、可衡量且可替代的 AI 基础设施。该生态系统由七个互补的开源项目组成，均已在 PyPI 发布，包括：OpenOntologyLite（便携式业务本体定义）、AgentPolicyPack（智能体策略即代码）、AgentForge（支持模型路由与预算控制的智能体编排）、PrivateAIStack（基于 Ollama 和 PostgreSQL 的私有本地部署方案）、ModelSwapBench（模型替代与迁移评估工具）、AIAuditLog（防篡改的 AI 系统审计日志工具）以及 AIMeter OSS（精确衡量 AI 使用成本与业务成效的工具）。

Show HN: Agent Legibility Analyzer see if AI shopping agents can read your store

AgentMint.net 推出了一款名为“Agent Legibility Analyzer”的的工具与实践手册，旨在帮助商家分析并优化其在线商店，使其更容易被 AI 购物智能体（例如基于 Claude Sonnet 4 的智能体）读取和选择。该平台提供基于实验证据的研究，展示了商品排版布局等因素如何影响 AI 智能体的决策，适用于开发者和技术 SEO 团队。

AI Integrated into My Brother's Custom Vibe Coded AAC System

有用户在Reddit上分享了将AI集成到其兄弟定制的AAC（辅助与替代沟通）系统中的消息，但目前仅有标题，缺乏具体的实现细节与技术内容。

Show HN: Kurvengefahr – browser CAD/CAM for pen plotters

A few years ago I made a pen plotter attachment for Prusa MK4 (https://www.printables.com/model/827264-pen-plotter-attachme...) and at the time I didn't have a good way to turn artwork into G-code for it, and I put the project on ice for a while.I recently wanted to dabble in line art again and made a small browser app to make it easier. As agentic AI tools of 2026 are quite addictive, it rather quickly grew into something quite a bit more - an integrated browser CAD/CAM for pen plotters that covers everything from importing existing artwork, creating artwork from scratch, preparing for plotting and hardware integration. It includes some off-beat features like a Logo interpreter for turtle art and Graves RNN for handwriting synthesis and in addition to 3D printer pretending to be pen plotters it now also supports actual pen plotters based on EBB (AxiDraw) and GRBL firmwares through Web Serial.If you own an AxiDraw or a GRBL plotter, I'd very much appreciate it you gave it a try and give feedback. As I don't own those, I did all the testing with a hardware mock on STM32, so I am not sure how well it works attached to an actual plotter.Source code and docs are on GitHub: https://github.com/tibordp/kurvengefahr

Microsoft joins Google in backing Go for AI agents — OpenAI and Anthropic lag

Go has emerged as the lingua franca for cloud infrastructure, used for everything from container orchestration and CI/CD pipelines to the command-line tools engineers rely on daily. Kubernetes, Docker, and Terraform are all written in it, a

The Amazon SES Alternative for AI Agents – MailKite

MailKite 推出了一款专为 AI 智能体（AI Agents）设计的邮件收发服务，旨在作为 Amazon SES 的替代方案。针对 SES 的两大痛点——新账户受限于沙箱（需要人工审核才能向外发信）以及接收邮件流程繁琐（需自行处理 S3、SNS 及 MIME 解析），MailKite 提供了开箱即用的方案。一旦域名通过 SPF 和 DKIM 验证，AI 智能体即可直接发信，且接收到的邮件会被自动解析为结构化的 JSON Webhook（包含已解码的文本、HTML、附件链接及安全验证状态）。MailKite 每月提供 3000 封免费信件额度。

Meta's New AI Photo Tool Feature Removed

Meta在发布AI图像生成工具Muse Image仅三天后，紧急下线了一项允许用户通过@-提及公开Instagram账号来获取其面部数据并生成图像的功能。该功能因默认开启而非用户主动授权（Opt-in）引发了公众及美国演员工会（SAG-AFTRA）的强烈抵制，指责其存在非自愿数字复制品的隐患。Meta随后承认该功能未达预期并将其下线，但底层的Muse Image模型仍保持可用。

Mnema: A local, encrypted memory layer for AI agents

Mnema 是一个专为 AI 智能体（Agents）设计的本地加密记忆层项目，旨在为智能体提供安全且保护隐私的本地化记忆管理与存储能力。

Show HN: Dr. Wong – an AI space for journaling and self-reflection

Dr. WongLog inSign upTalk it through with Dr. WongA calm, clear-eyed AI conversation partner with just enough bite to be useful.Self-reflection assistantSort thoughtsNight reflectionFree trial accessPour out the feelings, arguments, and anx

Databricks News: CLI v1.0.0, AI-tools, Docker, DABs UI sync, mutators

Databricks 社区发布新闻更新，主要内容涵盖 Databricks CLI v1.0.0 版本发布、AI 工具（AI-tools）、Docker 支持、DABs UI 同步以及 mutators 等功能特性。

An AI gateway that signs a receipt for every LLM response

AxioRank 开源了一款专为 AI Agent 设计的安全网关 AxioRank Gateway。该网关提供本地部署、兼容 OpenAI 接口的统一端点，其核心特性包括： 1. **离线可验证收据**：为每次 LLM 响应生成基于 Ed25519 签名的加密收据并形成哈希链，可离线验证，用以证明 AI 流量的真实执行历史。 2. **本地安全护栏**：无需额外网络请求，默认在本地拦截提示词注入（Prompt Injection）并对敏感数据（如密钥、PII）进行脱敏。 3. **路由与容灾**：支持多提供商路由、负载均衡、故障转移（Failover）与缓存，内置 34 种主流模型服务商预设（包括 DeepSeek、Ollama、OpenAI、Gemini 等）。 4. **极简审计**：采用 Node.js 编写，零运行时依赖，支持 Docker 一键部署，适合对数据合规和审计有高要求的企业开发者。

Show HN: Sanbox, batteries included sandboxes for AI agents

Hi HN,We are building Sanbox, a platform for running AI agents in isolated and resumable sandboxes.We use the OpenCode SDK as the harness, support reusable templates, and have a CLI that works with Codex, Claude Code, Cursor, CI, or your terminal. Each sandbox has MicroVM isolation, a persistent filesystem, and live trail of run events. Can also self-host if required for security/compliance.It's on the roadmap to add network ACL, secrets managements, LLM cost tracking & observability.We are also happy to build custom integrations or triggers for specific use cases. For example, spinning up a sandbox from a database edge function when data changes, or connecting Sanbox to an email inbox for complex workflow processing.It is still early, you can sign up to try here: https://sanbox.cloudMy email is also on my profile here.

I built an AI strength coach because I wanted my training backed by real studies

PerformanceAgent 🏋️ The first open-source AI Strength & Conditioning Coach powered by scientific research. English · Français · Español · Deutsch · Italiano It runs inside an AI agent CLI — a terminal program you chat with, such as Claude

Show HN: Spendict – a performance marketer's verdict for AI agents, over MCP

Spendict 是一款专为 AI 广告代理（Agent）设计的评估与决策工具。它能在广告实际投放产生花费之前，对广告素材、广告系列结构、投放表现和定位策略进行评估，并给出明确的判定（运行 run / 先修改 fix_first / 放弃 kill）。该工具支持 MCP 协议、CLI、Skill（可无缝集成至 Claude Code、Cursor、Gemini 等）以及 REST API，旨在帮助营销团队防止 AI Agent 错误投放低效广告并浪费预算。其评估标准由拥有 10 年以上付费社交媒体投放经验的专业营销人员校准。

Show HN: Lip Sync AI – Create Your Talking Videos Instantly

Lip Sync AI 是一款在线 AI 对口型视频生成工具，支持免注册免费体验。用户只需上传任意照片或视频并配上音频，该工具即可在数秒内生成口型同步的说话视频。它支持多语言和不同口音，并在生成过程中保持角色面部特征的一致性。该产品提供 Basic、Standard 和 Pro 等多种付费订阅及一次性买断计划。

AI-powered video generator SaaS Application for Sale

SaaS Currency: SaaS | Internet Verified Listing AI-powered video generation platform for creating viral short-form content (TikTok, Reels). Scalable SaaS with strong demand and monetization potential. Business Location Site Age 4 months Mon

Show HN: Runeward: Sandboxing AI agents with policy gates

Runeward 是一款开源的 AI 智能体（Agent）安全沙箱与治理工具，采用 Apache 2.0 协议。它通过声明式配置在 Docker 或 Kubernetes 中构建隔离的运行环境，默认禁止外发网络流量。该工具引入了防篡改审计账本、人工介入（HITL）审批网关、死循环检测及成本/Token限制等治理功能。支持通过 REST API、MCP（Model Context Protocol）、CLI 和 Web 控制台进行管理，旨在将安全规则置于模型外部强制执行，防止 Agent 执行高危操作或异常耗尽预算。

Show HN: Zero Trust Boundary for Agents

Hi HN, I’ve been working on Attestor, an open-source execution boundary for autonomous AI agents.

Show HN: Inkfold – workspace across multiple AI providers with shared memory

For people who pay for more than one AI — keep your memory and context as you move between ChatGPT, Claude, Gemini, Grok, and the rest. Stop re-explaining yourself every time you switch tools. No credit card Smart, Private, or Incognito ret

Agent Service – promptable AI agents with guardrails and downloadable packages

Agent Service – promptable AI agents with guardrails and downloadable packages

MSK – an AI agent that thinks like a CTO

Your on-demand AI CTO Only for iPhone Free · In‑App Purchases · Designed for iPhone. Not verified for macOS. iPhone Your AI CTO, on demand. Architecture reviews, scaling advice, and startup strategy — in chat or voice. No account needed. MS

TalkFitly – Practice high-EQ conversations with AI

AI coach for real talks Only for iPhone Free · In‑App Purchases · Designed for iPhone. Not verified for macOS. iPhone Not a quiz app. A social intelligence trainer. Have you ever — walked away from a conversation thinking "that's not what I

老黄RTX Spark真机现身Bilibili World!CPU和GPU直接焊在一起,笔记本跑120B大模型

英伟达在Bilibili World上首次展示了搭载RTX Spark超级芯片的笔记本电脑以及DGX Spark桌面超算。该系列芯片采用Blackwell GPU与20核Grace CPU，通过NVLink-C2C技术直接互联，配备128GB统一内存，算力达1 Petaflop。RTX Spark主打个人智能体，可在笔记本本地运行120B参数大模型（支持百万Context），并提供OpenShell安全运行时；DGX Spark则基于Linux系统，面向开发者，支持最高200B参数大模型的本地微调与推理。

Dismissive Dan's Review of the Overplane AI Coding Harness

Overplane 是一款开源的单二进制工具，能够将 Markdown 格式的软件规范转化为实际代码。它首先通过 SMT 求解器（如 Z3）交叉验证规范的逻辑一致性，然后在受限的本地容器（如 Docker 或 Podman）中运行 AI 编码智能体（如 Claude Code、Gemini CLI 等）来生成代码。虽然有行业评论指出，资深开发者可以通过自定义脚本和沙箱实现类似功能，且该工具仅验证规范一致性而非代码本身的正确性证明，但它为需要安全沙箱、多模型支持和求解器校验的团队提供了一体化的开箱即用方案。

Show HN: TrialPilot – Clinical trials from your phone, built by a patient

Hi HN. I’m Jon.I have Long COVID. Five years and counting. Somewhere around year two I realized every trial I actually wanted to be in was a plane ride away, and I was too sick to fly. I’d watch new studies open, look at the location, and close the tab.I asked researchers I trust why decentralized trials weren’t the default. The answers are honest ones. Clinical research is a slow-moving field for good reasons. The people running trials are clinicians and biostatisticians, not tech founders. The regulatory framework was written for a world of paper case report forms and site visits, and it takes years for the guidance to catch up to what’s technically possible.Nobody expected a mass disabling event on the scale of Long COVID. The field is genuinely headed toward decentralized as the default, and the FDA’s 2024 DCT guidance and 2026 real-time trials pilot are real signals of that. But the platforms researchers have to choose from today still cost six figures per study and still require physical sites. So most of the researchers I know use spreadsheets instead.So I built TrialPilot with two engineers I’ve worked with for a decade. The whole study runs from a patient’s phone. Researchers configure a study, generate a code, and watch enrollment in real time. Patients scan the code, sign consent on the phone, connect a wearable, and the study runs itself. Real functional tests like the six minute walk and NASA Lean Test happen from your living room instead of a clinic.Beachhead is invisible illness. Long COVID, ME/CFS, POTS, Lyme. The populations that get under measured because their symptoms don’t show up on a scheduled clinic visit, and get structurally excluded from research because they’re too sick to travel to it.We have clinical trial partners lined up and are getting ready to launch our first pilot studies. We’re also competing in the HHS TOPx Tech Sprint for AI and Invisible Illness, so wish us luck.https://www.trialpilot.app/Would love your impressions and feedback. What lands, what doesn’t, what would you build differently, what would make you use it or trust it.Especially from anyone who’s tried to run a trial and lost their weekends to it, or anyone who’s tried to enroll in one and lost their patience.I’ve been waiting five years for this to exist, so I built it.

Show HN: AgentTransfer – open-source file transfer for AI agents (one Go binary)

AgentTransfer 是一款专为 AI Agent 设计的开源文件传输与协同工具，采用 Go 语言编写，支持单文件部署。AI Agent 可在无需人类干预和无 SDK 的情况下，传输最大 5GB 的文件。该项目为 Agent 提供了持久化存储、临时分享链接和专属邮箱收件箱，并支持 Agent 之间的相互发现与共享空间协作。它还支持模型上下文协议（MCP）服务、客户端加密以及基于 Ed25519 签名的操作收据链，解决了 Agent 之间传输模型权重和数据集等大文件的痛点。

Mesh LLM: distributed AI computing on iroh

Mesh LLM 是一款基于 P2P 网络库 iroh 构建的分布式 AI 计算框架。它能将多台机器的 GPU 和内存资源进行池化，提供统一的 OpenAI 兼容 API（localhost:9337/v1）。该框架支持本地运行、路由至对等节点，以及名为“Skippy”的拆分模式——可将超大模型（最高支持 235B MoE 模型）按网络层拆分到多台低配置机器上协同运行。Mesh LLM 无需中央服务器，利用 iroh 实现 NAT 穿透与安全的 QUIC 连接，目前客户端大小仅约 18 MB，支持 40 多种模型。

Show HN: Token Time – Screen Time, but for your AI agent tokens

Token Time 是一款专为 macOS 设计的本地 AI Token 使用量和费用监控工具，类似于针对 AI 智能体的“屏幕使用时间”。它在菜单栏实时显示 Token 消耗和今日花费，支持按小时和模型进行费用细分。当 Token 消耗达到设定阈值时，它会提供 10 秒的全屏提醒以防费用超支。该软件完全在本地运行，无云端同步或追踪，单次买断售价为 6 美元。

Neobrowser AI has rediculously strong VPN builtin for FREE

诺顿（Norton）母公司 Gen Digital 推出了一款主打隐私保护的 AI 浏览器 Norton Neo（又称 Neobrowser AI）。该浏览器内置免费且无日志记录的 VPN，并提供反指纹追踪和广告拦截功能。在 AI 隐私方面，它采用“设计即隐私”架构，对 AI 服务商实行零数据留存政策，不使用用户数据进行模型训练，并通过服务端路由 AI 请求以隐藏用户的真实 IP 地址和浏览器标头。

AI found a secret computer bug hidden for 15 years.

安全公司 Nebula Security 研发的 AI 工具 VEGA 发现了一个隐藏在 Linux 系统中长达 15 年（自 2011 年起）的安全漏洞。该漏洞可被利用以获取受影响机器的完全控制权。谷歌为此向其支付了超过 92,000 美元的漏洞赏金。目前该漏洞已被修复。此事件展示了 AI 在检测人类难以发现的老旧代码缺陷方面的强大能力。

Two LLMs play live chess and rewrite their own brains after each game

Loading tonight's match… ChatGPT 5.5 vs Claude Fable 5 — the flagship models, no stand-ins 🥊 NEXT UP: GPT‑5.6 joins this duel the moment OpenAI opens API access — Claude rides a 6-game win streak into the rematch. Same playbooks, same cham

Show HN: BoundFlow – an open-source control plane for AI agents

The operational layer for the LLM agents and workflows you run unattended — cost caps, approval gates, and self-healing policy, enforced by a control plane. ImportantPublic preview (pre-1.0). The engine is complete and covered by Go, mock-L

Bitemporal provenance in agent memory: What did we believe, when, and why

MnesticDB（硬分叉自 CozoDB，使用 MPL-2.0 协议）是一款专为 AI Agent 记忆（“AI 海马体”）设计的关系-图-向量数据库。新版本引入了“双时态”机制（区分有效时间与事务时间），支持对 Agent 历史记忆进行审计与“时间旅行”；同时实现了半环溯源（semiring provenance）框架，可记录决策背后的证据链，使用户能够追溯 Agent 在特定时间点相信什么以及为什么相信。此外，该版本还优化了查询性能（如贪婪连接重排和 Yannakakis 风格的因子化计数），以防止 LLM 自动生成的低效查询导致系统卡死。目前已在 crates.io 和 PyPI 上线。

I built TradingSpy: local, privacy-first AI trading assistant(First Open Source)

TradingSpy 是一款开源、本地优先且保护隐私的 AI 交易研究助手。它通过 Docker 部署，结合了传统数据可视化与 LLM 代理（Agent）技术，支持 Google AI Studio、Ollama、Mistral 等多种大模型。用户可以使用自然语言描述交易想法，由 Agent 自动生成 Backtrader 策略代码，并针对历史数据进行回测、参数优化以及与基准策略的对比。该项目完全在本地运行和存储数据，目前采用非商业许可开源。

Show HN: Sqlsure – deterministic semantic checks for AI-generated SQL

Sqlsure 是一款开源的确定性语义检查工具，专门用于解决 AI（如 Text-to-SQL 模型和 AI Agent）生成 SQL 时出现的静默逻辑错误（如因 Join 导致的重复计算、平均值错误求和等）。该工具不依赖 LLM 调用，而是通过对比已有的 dbt 元数据、数据库 Schema 或主外键关系，在 0.1 毫秒内完成离线、确定性的静态分析。每次拒绝都会输出机器可执行的修复方案，支持 AI Agent 实现“生成-检查-修复-执行”的自我修复闭环。此外，它还提供了 MCP 服务端和 CI 门禁集成，便于无缝嵌入现有的 AI 开发流中。

AI Arcade – which coding model can build the best arcade game

Run

Show HN: Don't let your engineering brain rot in the age of AI

30 Seconds of Knowledge 是一款浏览器新标签页扩展程序，旨在帮助开发者在 AI 辅助编程时代保持代码阅读和理解能力。该插件在用户每次打开新标签页时展示一段可在 30 秒内读完的真实代码片段（涵盖语言特性、框架模式和面试题等），防止开发者因过度依赖 AI 自动补全而导致编程基本功退化。目前该工具已积累了 1500 多个代码片段，吸引了超过 2.5 万名开发者使用。

I built a free tool to evaluate AI agent outputs (human labels and LLM judges)

Verdict 是一款开源且完全在本地浏览器运行的 AI 评估工具，旨在帮助开发者分析和优化 AI Agent、聊天机器人及 RAG 系统的输出。该工具支持导入 OpenTelemetry、Langfuse 等多种 Trace 格式，便于人工进行快速的 Pass/Fail 标注与错误分类。此外，它能帮助用户构建并验证“LLM 裁判（LLM as a Judge）”，通过对比人工标注计算混淆矩阵、真阳性率（TPR）、真阴性率（TNR）及 Cohen's kappa 系数，确保 LLM 评估与人类标准一致后方可导出部署至生产环境。整个过程无需后端或账号，保障数据隐私。

Sovereign AgentOps – Self-hosted constitutional AI governance for MCP agents

Autonomous Digital Organization Platform — Community Edition Runtime-agnostic governed agent execution — demo and evaluation edition Quick Install # PyPI (standalone CLI + MCP server) pip install sovereign-agentops-community # Docker docker

Show HN: Wizard – Self-extending Rust terminal AI agent (one-line install)

Show HN: Wizard – Self-extending Rust terminal AI agent (one-line install)

Show HN: A Trust Index for MCP Servers

开发者推出了一款针对 MCP（Model Context Protocol）服务器的安全与信誉指数平台 Canopii。该工具定期从官方 MCP 注册表拉取服务器，并基于运行时防护、静态代码安全测试（SAST）、传输模式和凭证风险等指标自动进行安全评分。目前已对超过 1.2 万个 MCP 服务器完成了评估（其中 45% 获评 A 级，1,262 个被评估为 D/F 级高风险）。该平台还提供了 API 接口，方便开发者和企业将安全评分集成至 CI 门禁、采购审查或智能体（Agent）的允许访问列表中。

OpenSandbox Universal Sandbox Infrastructure for AI Applications

OpenSandbox 是一个专为 AI 工作负载（如编码智能体、浏览器自动化、远程开发和强化学习等）设计的通用沙箱基础设施，现已列入 CNCF 全景图。它支持在 Docker 和 Kubernetes 运行时中安全地隔离运行命令、代码解释器、浏览器和开发工具。该项目提供了沙箱生命周期管理功能，并支持 Python、Java/Kotlin、JavaScript/TypeScript、C# 和 Go 等多语言 SDK。

AI Found a Root Bug in Linux That Everyone Missed for 15 Years

Amid years of warnings that China’s notorious Volt Typhoon hackers may be pre-positioning within United States critical infrastructure, a closed-door war game for insurers played out an array of worst-case scenarios—revealing a menacing, di

Show HN: Standalone SearXNG CLI+MCP (no server needed)

开发者 nikvdp 开源了 searxng-ai-kit，这是一个面向终端和 AI 助手的独立命令行工具及 MCP (Model Context Protocol) 服务端。该项目将隐私元搜索引擎 SearXNG 封装为 Python 库和单文件二进制包，无需额外部署 SearXNG 服务器。它支持 180 多个搜索引擎的并发检索和网页内容抓取，并支持作为 MCP 服务器与 Claude Desktop 等 AI 助手集成，使本地或开源 Agent（如 Hermes、OpenCode 等）能够低成本地获取网络搜索和网页解析能力。此外，它还内置了终端 AI 问答与交互式聊天功能。

Agentation – Visual UI Annotation for AI Coding Agents

Agentation 是一款面向 AI 编程智能体（如 Claude Code、Cursor 等）的桌面端可视化 UI 标注工具。用户可以直接在页面上点击元素并添加批注，该工具会将其转化为包含 CSS 选择器、源码文件路径、React 组件树和计算样式等结构化上下文，帮助 AI 快速精准定位代码。此外，它支持 MCP（模型上下文协议）集成，使智能体能够直接读取并交互式响应用户的标注，无需手动复制粘贴。目前该工具对个人和企业内部使用免费。

Show HN: I Wanted AI Code Review I Could Own. So I Built Codra

开发者分享了其构建的 AI 代码审查工具 Codra。该工具旨在解决数据隐私和控制权问题，让开发者能够完全自主掌控（Own）自己的 AI 代码审查流程。

Concerned about how companies will manage their cloud bill once agents dominate

Hacker News 用户发帖表达了对未来 AI Agent 普及后企业云服务账单飙升的担忧。由于 Agent 旨在优化输出结果，成百上千个 Agent 在底层数据和计算层运行可能会导致 Token 消耗和算力费用失控，目前行业还缺乏能够有效控制此类成本的企业架构方案。

Majority of U.S..Majority of U.S. workers support an AI wealth fund

Digital generated image of young man in empy black space with semi reflective floor surrounded by multicoloured semi transparent data screens.Andriy Onufriyenko | Moment | Getty ImagesA majority of U.S. employees now want to hold corporatio

Show HN: Itara – Distributed system topology as an explicit, executable layer

Hi HN, I'm Gábor, a software engineer from Budapest.I spent almost a decade designing, building and maintaining distributed systems, and I came to truly understand why a lot of people define software architecture as "the stuff that's hard to change later". Changing service boundaries, communication protocol or serializers is time consuming and risky. The past few months, I've been building Itara, my attempt to ease that pain.Itara takes the software topology that's currently spread across the codebase and configuration and infrastructure, and concentrates it into a dedicated, executable layer. This dedicated layer describes the topology as a directed graph, where the nodes are the components of the system, and the edges are the connections between them. The edges have all the properties of the connection, like the transport to use, the serializer to use, the failure handling strategy and so on. Colocated components, components running in the same process, are modelled with direct connections.This gives me an accurate map of my topology, lets me change the topology without changing the business code, and keeps communication logic separate from business logic while still giving me direct control over the configuration, avoiding the network fallacies.In Itara, each component consists of an API the other components can build against, and an implementation that actually implements the business logic. Event-driven design is supported through dedicated events APIs.For each deployment unit, a wiring agent runs at startup to prepare the communication channels as specified by the wiring config. These channels implement the component APIs and are used by the application like regular interfaces. There is no runtime overhead, except for the structural observability events, because the wiring agent steps aside after startup.The project aims to be language agnostic. The current implementation supports Java, with a Rust implementation at proof-of-concept level.I collected a few common questions and their answers in an FAQ: https://github.com/itara-project/itara/blob/main/docs/FAQ.mdI prepared a demo, an order processing system that consists of 5 components, one of them written in Rust, communicating through HTTP and Kafka events. The demo shows that to change the topology, only the wiring file and the docker compose file need to change, the application code can stay the same. It also includes a deliberately flaky transport to demonstrate failure handling. The traces make the topology changes directly visible. Link to the demo: https://github.com/itara-project/itara/tree/main/demoI'd appreciate your feedback! Does this solve a problem you've encountered? What parts of the direction resonate with you, and where do you think it falls short?The spec, manifesto, and architecture docs are in the repo: https://github.com/itara-project/itara

Apple files lawsuit, accuses OpenAI of stealing trade secrets

苹果公司正式提起诉讼，指控 OpenAI 窃取其商业机密。这一诉讼标志着两家科技巨头在人工智能领域的竞争与知识产权冲突进一步升级。

Perfectly Hitting the Wrong Target: The Story of an AI Code Review Benchmark

Hexmos创始人Shrijith Venkatramana撰文对现有的AI代码审查基准测试（如Code Review Bench）提出了质疑。他指出，目前的基准测试将AI代码审查视为单一问题，但实际上AI的发展已将其拆分为两个完全不同的方向：一是面向人类工程师的“信息推荐”，需要过滤噪音、优化有限的人类注意力；二是面向AI Agent的“机器验证”，需要穷尽式地发现并修复问题以提升软件稳定性。作者认为，基准测试盲目以历史人类审查数据为标准，无法真正衡量软件质量的提升，呼吁业界重新审视AI代码审查的评估方法。

At this point, I’ve given up hope that X’s bot-reply problem is solvable. The “the part no one says” writing is annoying, but worse is that they a...

沃顿商学院教授 Ethan Mollick 针对 X 平台（原 Twitter）日益严重的机器人自动回复问题提出建议。他指出当前的机器人回复内容千篇一律，并提议 X 平台应通过在潜在空间（latent space）中测量语义距离，来筛选并优先展示真正具有多样化观点的回复。

AI Adoption Was the Easy Part. Now Comes the Dollar-Sign Shock

AI Adoption Was the Easy Part. Now Comes the Dollar-Sign Shock

Investors sell longer-dated AI debt amid Big Tech borrowing spree

随着大型科技巨头为资助AI基础设施建设而大举借债，投资者开始抛售与AI相关的较长期债券。这表明在科技公司融资热潮中，市场对长期债务风险的评估正发生变化。

India's TCS plans up to 8,900 AI deployment engineers, seeks AI acquisitions

印度最大的软件服务外包巨头塔塔咨询服务（TCS）计划组建一支由 5,900 至 8,900 名“前沿部署工程师”（FDE）组成的团队，直接入驻客户企业以加速 AI 工具的落地。同时，TCS 正在积极寻求 AI、数据安全和网络安全领域的收购。尽管市场担忧 AI 会缩短项目周期并挤压外包行业利润，但 TCS 首席执行官认为，整合多模型系统仍需深厚的客户环境知识，AI 将带来新业务而非颠覆外包模式。数据显示，TCS 第一季度 AI 业务环比增速从上季度的 28% 放缓至 13%。

The impressive AI demo is dead. Here's what actually reaches production

根据 Confluent 发布的《2026年数据流报告》，仅有 32% 的企业成功在生产环境中运行智能体 AI（Agentic AI）。研究表明，阻碍 AI 从 Demo 走向生产环境的最大瓶颈并非模型本身，而是实时数据基础设施与数据质量（三分之二的受访者提及）以及相关技能人才短缺（71% 的受访者提及）。72% 的 IT 领导者指出，缺乏足够的实时数据处理基础设施是限制 AI 规模化的主要障碍。因此，企业对数据流（Data Streaming）的投资比例（88%）首次超过了对 AI/机器学习本身的投资（82%），企业正将重心从单纯优化模型转向构建实时、可靠的数据管线。

The fight against AI data centers is just beginning

随着AI的爆发式增长，全球各地针对AI数据中心建设的社区抵制和能源争议正不断升温。文章回顾了2015年苹果在爱尔兰筹建数据中心所引发的早期抗议，指出当前的AI热潮正在严重威胁地方电网，围绕数据中心建设的环保与资源争夺战才刚刚拉开帷幕。

S&P Global sees OpenAI as a "key credit risk" for Oracle and cuts its credit rating

标普全球（S&P Global）将甲骨文（Oracle）的信用评级从“BBB”下调至“BBB-”（仅比垃圾级高一级），并将OpenAI视为其“关键信用风险”。标普指出，甲骨文在AI领域的资本支出预计到2027年将飙升至950亿美元，而OpenAI占了甲骨文6380亿美元合同义务的近一半。如果OpenAI发生变故，甲骨文将面临庞大且无法填补的数据中心闲置产能。此外，软银已将以OpenAI股票抵押的贷款规模从100亿美元缩减至60亿美元，OpenAI也将IPO推迟到了2027年。

AI backlash hits university: laptops and phones banned for law students

芝加哥大学法学院宣布将从今年秋季起，禁止一年级学生在课堂上使用手机、平板和笔记本电脑。此举旨在应对AI对高等教育的冲击，确保学生在没有AI依赖的情况下培养独立思考和批判性思维能力。课堂上将由指定的“书记员”负责为集体记笔记。虽然限制了课堂使用，但学校并非完全排斥AI，而是推行“抗AI教学法”，要求学生在法律研究和写作中将AI作为辅助工具使用（但严禁直接用AI撰写材料），并提供AI在法律领域应用的选修课程。

Meta kills Muse Image feature that let anyone generate AI photos of Instagram users without consent

Meta pulled a controversial feature from its new Muse Image model after widespread criticism. The feature let users generate AI images of other people by @-mentioning their public Instagram accounts. No consent needed, just a username. Meta admits "this feature missed the mark" and shut it down days after announcing it. The article Meta kills Muse Image feature that let anyone generate AI photos of Instagram users without consent appeared first on The Decoder.

Memory makers are slaves to the boom-bust rollercoaster

AI数据中心对高带宽内存（HBM）、DDR5和NAND闪存的极高需求导致市场严重供不应求，SK海力士、美光和三星等内存巨头的营收大增。尽管各大厂商正投资数千亿美元扩大产能（如韩国主导的5760亿美元半导体投资计划），但由于新建晶圆厂通常需要三年以上时间，预计内存高价将持续至2028年。高企的硬件成本给AI初创公司带来了沉重的资金压力；若未来AI需求不及预期，内存行业恐将面临极其严重的产能过剩与行业低谷。

AI Agents Are About to Change Payments Operations

Rapyd总经理David Rosa在播客中探讨了AI智能体（AI Agents）和稳定币对跨境支付与金融服务带来的变革。他分享了Rapyd如何利用AI加速商户入驻、实现内部运营自动化以及辅助决策，并指出AI的普及正在重塑支付业务的经济模式。

25% long-form social media posts appear AI-generated

REG AD ai and ml One in four long-form social media posts appear entirely AI-generated, with nearly half of those on Microsoft's and Elon's platforms involving AI in some form No surprise here. A study from AI detection platform Pangram sug

Databricks AI Agent Genie Code Is No Longer Free. Now You Have to Pay as You Go

Databricks 旗下的 AI 数据智能体 Genie（原免费试用）现已调整收费策略，正式转向按需付费（Pay-as-you-go）模式，用户后续使用该服务需要支付相应的计算和资源费用。

Chasing new skills, going back to basics: how software engineers adapting to AI

随着AI技术的发展，软件工程行业正面临剧烈变革，例如谷歌已有75%的代码由AI编写。软件工程师的角色正从传统的“编写代码”快速转变为“审查和验证AI生成的代码”。这一变化伴随着行业裁员和计算机专业入学率下降等挑战。为应对这一趋势，程序员们正通过提升AI评估能力、重温编程基础，甚至通过成立如“What We Will”等互助组织，来应对AI带来的职业颠覆。

CEO Pleads with AI Industry to Stop Charging So Much to Replace Human Labor

网络安全巨头 Palo Alto Networks 的 CEO Nikesh Arora 近日呼吁科技行业降低 AI 的使用成本。他指出，大语言模型（LLM）的成本在 2027 年前需要下降 20%，到 2028 年前需下降 90%，企业才能真正有效利用该技术。Arora 认为，目前 AI 的高昂成本与其带来的实际自动化价值并不匹配。科技评论家 Ed Zitron 也表达了类似看法，认为 AI 行业目前的实际市场规模仅在 100 亿至 300 亿美元之间，却被伪装成了万亿美元规模的行业。

How does a Dev's job look like in a few years?

I'm a experienced/senior developer which is frequently using ai, guiding coding agents, etc. I wonder, how does my job look like in a few years? Which skills might be the best ones to have? Currently, having business knowledge, development experience helps greatly with guiding coding agents, creating MVPs/PoCs in "no time", improving code, etc. But what if coding agents/ai would overtake this job?

Claude Cowork's biggest use case is the mundane office work nobody wants to own, Anthropic says

Anthropic analyzed 1.2 million Claude Cowork sessions from more than 600,000 organizations. About half of all usage goes toward business processes and text creation, what Anthropic calls "the work around the work." That means tasks like compiling status reports, building onboarding checklists, or putting together slide decks. Software development barely shows up in Cowork because developers stick with Claude Code for that. The article Claude Cowork's biggest use case is the mundane office work nobody wants to own, Anthropic says appeared first on The Decoder.

Meta u-turns on AI feature amid privacy backlash

由于遭遇公众对隐私问题的强烈抵制，Meta撤回了其在Instagram等平台上推出的相关AI功能。

OpenAI CEO Altman is now "pretty sure" AI is net job-creating, which is quite the pivot from predicting mass layoffs

OpenAI 首席执行官 Sam Altman 近期改变了此前关于 AI 将导致大规模失业的预测，表示他现在“相当确定” AI 在净效应上创造了更多就业机会。Anthropic 首席执行官 Dario Amodei 也修正了其先前的言论，将 AI 视为生产力倍增器而非工作杀手。尽管目前的研究尚未发现 AI 对整体劳动力市场产生显著影响，但部分企业的确因将预算倾斜至 AI 硬件而进行了裁员。

The Hard-Line Activists Ramping Up for the War with AI

《华尔街日报》报道了针对人工智能的强硬派抗议活动正在升温。反AI活动人士正在积蓄力量准备与AI的发展展开对抗，反映出社会层面对于AI技术快速普及的抵制情绪与冲突正在加剧。

Shark: The First Light Aircraft That Cancels Turbulence [video]

该视频介绍了一款名为 Shark 的能够消除气流颠簸的轻型飞机。该条目为检索误报（因“Aircraft”中含有“ai”而被召回），实际内容与人工智能技术无直接关联。

98年哈工大教授创业,要做人形灵巧操作世界模型

从采集触觉，到对齐触觉，再到使用触觉

Grades dropped from 96 to 48 percent when a Brown professor made students take the exam without AI

布朗大学经济学教授Roberto Serrano发现，在一次均分高达96%的家考试后（怀疑多数人使用AI作弊），他将期末考试改为闭卷线下监考，结果导致18名学生退课、9人缺考，期末均分骤降至48.6%，创下该课程历史新低。此外，来自中国（涉及2.6万名学生）和加州大学伯克利分校（涉及50万份成绩）的两项大样本研究也印证了这一现象：学生在平时作业中依赖AI虽能提高作业分数并缩短时间，但会导致监考考试成绩大幅下滑（如中国研究中考试成绩平均下降20%，优秀学生受损最严重）。

Chamath on all important “prefill” and “decode.” in AI compute. Prefill is compute-bound; massive parallel GPUs win, so Nvidia dominates as contex...

Chamath Palihapitiya 在 All-In Podcast 中解释了 AI 计算中 Prefill（预填充）与 Decode（解码）阶段的不同硬件瓶颈。他指出，Prefill 阶段是计算受限的（compute-bound），需要大规模并行 GPU，因此随着上下文的增长，Nvidia 在这一阶段占主导地位；而 Decode 阶段则是内存带宽受限的（memory-bandwidth bound），因为生成下一个 Token 依赖于扫描已生成的内容。

Inside the secret AI war between Silicon Valley and China

根据《华盛顿邮报》的报道，AI安全与研究公司 Anthropic 指控中国公司正在通过“知识蒸馏”（distillation）其 Claude 模型来获取技术与知识，这也折射出硅谷与中国之间暗中展开的 AI 竞争。

Big Tech piles on $350B in debt to fuel AI data center race

The largest builders of artificial intelligence data centers have doubled their debt load in the last five years, turning to borrowing to finance an unprecedented spending spree they claim is needed to transform the economy. Alphabet Inc.,

AI boom puts Big Tech's transparency to the test

AI boom puts Big Tech's transparency to the test

Guide to the circular deals underpinning the AI Boom

彭博社发布专题指南，解析支撑当前 AI 行业繁荣的“循环交易”（circular deals）商业模式。这类交易通常涉及科技巨头向 AI 初创公司提供投资，而初创公司又将资金用于购买该巨头云算力的循环回流模式。

Under federal rule, colleges must leave grads better off or lose financial aid

Under federal rule, colleges must leave grads better off or lose financial aid

Cloudflare Threatens to Cut Google Off from Their Publishers in Searches

内容分发网络 Cloudflare 宣布从2026年9月15日起，将默认对新注册用户和免费层用户屏蔽同时用于搜索索引和AI训练的“多用途爬虫”（主要针对谷歌）。此外，由于谷歌未像微软、OpenAI等同行一样与出版商达成AI内容授权协议，包括 USA Today 和 Beehiiv 在内的出版商也计划屏蔽谷歌爬虫，甚至准备从谷歌搜索中彻底除名。

AI notetakers promise easy meeting recaps, but some question their use

AI会议记录工具在提供便利的同时，也带来了严重的隐私和安全风险。这些工具将会议发言转化为文本数据，可能导致商业机密泄露、律师与客户间的保密特权失效。此外，一些AI工具会生成并存储用户的“声纹”（属于生物识别特征），这在部分地区（如美国伊利诺伊州）面临严格的法律监管（如BIPA法案）。行业专家建议用户在会议中主动甄别是否有AI机器人加入，必要时拒绝录音，并仔细审查服务商的数据存储及是否将其用于AI模型训练。

近百名玩家涌入具身数据:一年融资44.7亿,谁能真靠“卖数据”赚钱?

为了帮你看清具身数据行业，我们总结了以下十个行业现状

Lagacy media is problematic. 👀

机器人公司 1X 的相关开发人员（dar）表示，他将 Neo 机器人手部发布的独家新闻权交给了《连线》（WIRED）杂志，但后者却发表了一篇题为《The 1X Neo Robot Has Freaky-Fast Fingers》的偏颇报道，指责他们将机器人“性暗示化”（sexualizing robotics）。开发人员对此表示感到被背叛，因为这完全偏离了技术研发的初衷。AI 行业观察者 Rohan Paul 转发并批评传统媒体的这种报道倾向。

Bring seamless PQC encryption into every messenger you already use

Hi everyone!I am a big privacy activist and I strongly believe privacy is a fundamental human right.Lately, things have been quite frustrating to say the least. We have Meta removing end-to-end encryption over the past months, the continuing debate around EU Chat Control, and major technology companies collecting enormous amounts of data while building their own AI systems.The most popular messengers are convenient because everyone already uses them. But most of them are not truly privacy-first.It's also the reason why consumer social apps are such a competitive field.Telegram, for example, markets itself heavily around privacy, but normal Telegram chats are not end-to-end encrypted. Users are still trusting Telegram to create, store and manage their private keys on their own servers.But there are much better privacy-focused tools that I myself have also been a user of throughout my life. Signal stores keys locally and makes end-to-end encryption the default. SimpleX takes a more decentralized approach. Meshtastic allows hardware networks. Many of these projects are open source and solve real privacy problems.But they all face the same major issue: the network problem.A messenger is only useful when the people you need to talk to are using it. It is difficult to convince your friends, family, coworkers, or customers to leave the applications they already use and move somewhere else.That is why I started thinking about a different approach to restore privacy in our daily lives. Instead of creating another messenger and asking everyone to switch, what if we could bring private encryption into the messengers and social platforms people already use?That is what I am building with my side project experiment, Ekko.Ekko lives inside your browser extensions or mobile keyboard. You write a message normally, Ekko PQC encrypts it locally before it reaches the messenger, and the recipient decrypts it locally on their device. The messenger only transports the encrypted content.PGP showed decades ago that individuals could control their own encryption keys. But it also showed how difficult encryption becomes when users have to manually manage keys, copy messages, encrypt them, paste ciphertext, and then repeat the process to decrypt a response.Strong cryptography is not enough if normal people cannot use it.The goal with Ekko is to make that process seamless with modern, already popular messengers. You should not need to understand encryption algorithms or switch to self hosted solutions, move applications.Ekko is still early. We have a good working prototype that I am testing with friends in early access while we build out a backend infrastructure and ace the seamless messengers integrations. It does not work perfectly just yet, but we are actively getting there!The applications, browser extensions, and cryptographic code are intended to be open source.The messenger may still see metadata, including who is communicating, when messages are sent, and how frequently people communicate. A platform could block encrypted messages or break an integration, but it could actively be battled back, especially when we have almost a billion Europeans about to be actively surveilled.But it could give users control over the actual contents of their messages without forcing their entire social circle to move to another platform.I would appreciate direct criticism, especially around the security model, key exchange, multi-device support, metadata, platform restrictions, and whether this approach actually solves enough of the network problem to be useful.Just publitised this idea two days ago and started building in public so any thoughts are appreciated :)https://useekko.appI would love to talk! Any inquiries, suggestions or collaborations, kirill@useekko.appLets fight for the right to privacy to never be suppressed!- Kirill from Ekko

The Human Cell Is Wildly Complex. Can AI Decode It? – Silvana Konermann – Ted [video]

该视频为 Silvana Konermann 在 TED 上的演讲，探讨了人工智能（AI）在解码极其复杂的人体细胞及推动生物医疗变革方面的潜力和前景。

Being part of one of the biggest wealth creation events in history (AI)

《华盛顿邮报》报道指出，人工智能（AI）正在引发历史上最大规模的财富创造事件之一。伴随AI领域的爆发性增长，即将诞生一批新晋亿万富翁，他们正在思考如何处理和使用这些巨额资金。

Banning AI in Law School: We've Seen This Before

芝加哥大学法学院近期出台新规，禁止一年级学生在课堂上使用手机、笔记本电脑和AI，引发广泛讨论。文章作者通过回顾1982年哈佛法学院因公平和作弊担忧而禁用首款便携式电脑（Osborne 1）的历史，指出对新技术的警惕与禁令并非新鲜事。作者认为，正如当年的个人电脑和互联网最终普及并重构行业一样，当前的“AI原生代”使用AI的趋势是不可阻挡的，预防性的技术禁令只会阻碍进步，人们应当积极拥抱技术并学会高效利用它。

The AI Disagreement Index: 8 models agreed on the "best tool" 0 of 16 times

根据 Deep Synthesis 发布的“AI分歧指数”（AI Disagreement Index）显示，在16次测试中，8个不同的 AI 模型在选择“最佳工具”时达成共识的次数为0。这表明当前主流大模型在评估和推荐工具时的决策逻辑存在显著差异。

AI 2040 – The Comic

该内容仅包含指向《AI 2040》漫画的链接，缺乏具体的行业或技术实质性内容。

Physical AI scale up chemistry startup gaining traction at Big Pharma

加拿大化学自动化初创公司 Telescope Innovations 开发的物理 AI 自主实验室（SDL）技术在制药和先进材料行业取得重要商业进展。该技术通过其专利的 DirectInject-LC™ 自动采样接口，实现了化学反应过程的实时分析，并将数据直接输入本地部署的贝叶斯优化算法。这种“本地边缘 AI”模式有效解决了制药企业对知识产权泄露的担忧。目前，该公司已向辉瑞（Pfizer）交付了第二套 SDL 系统，与韩国制药及生物制药协会（KPBMA）达成基础设施合作，并与欧洲某大型药企签署了结晶工作流自动化协议。此外，该平台已被证实可跨界应用于锂电池回收材料（纯度 >99.9% 碳酸锂）的自动纯化优化。

AI rebrands fail to deliver a lasting share price boost

据《金融时报》报道，企业通过重塑品牌加入“AI”元素，并不能为其股价带来持久的提振效应。

Safe from AI: which jobs will help you thrive in the future?

Entering the world of work often brings some uncertainty, but now there is another question: how can I AI-proof my career?We asked people from across various industries what they think the impact of AI will be on careers, and which jobs may

AI 2040 and the Cult of Intelligence

著名开发者 George Hotz（geohot）发表博客，对 AI 2040 预测以及 AI“急剧跃升（hard takeoff）”的观点提出批评。他指出，现实世界受到供应链、物理规律和制造周期的限制，纯粹的智能或 Token 无法直接解决现实中的硬件制造和物理问题。此外，他强烈反对科技巨头和政府推动的 AI 监管，主张“本地化 AI”（Plan L），认为 AI 应当完全听命并服务于其所有者，而不应设有由科技公司强加的限制和“护栏”。

Who manages the agents?

本文探讨了AI未来的两种发展愿景：一种是由少数技术精英控制的集中式AI，另一种是以人类为中心、将AI作为能力放大器的分布式智能。作者指出，虽然前沿AI能力飞速提升，但仅有极少数超级用户实现了生产力飞跃，普通员工的效率并未显著改变。作者倡导企业和个人应当拥有并控制属于自己的“主权Agent”（Sovereign Agents），让每一位员工学习成为Agent的管理者，而不是直接用AI替代人类，以此实现真正的生产力普惠和业务增长。

Thalmaar AI (Defense and Space Tech) Is Hiring Founding PM

Thalmaar AI (Defense and Space Tech) Is Hiring Founding PM

Alex Karp Is Saying What Every Angry CEO Is Thinking About AI

Alex Karp Is Saying What Every Angry CEO Is Thinking About AI

Reverse centaurs are the answer to the AI paradox

科幻作家兼科技评论家 Cory Doctorow 撰文探讨了 AI 带来的悖论：为何有人因 AI 改善工作，有人却视其为折磨。他引入了“人马”（人利用机器辅助）与“逆人马”（机器将人作为助理和“背锅侠”）的概念。前者如创作者自主使用 Whisper 提高转录效率，后者如被迫使用 AI 批量生成内容并为 AI 幻觉承担责任的员工。Doctorow 指出，当前的 AI 投资是一场泡沫，但泡沫破裂后仍会留下有价值的遗产，尤其是像 Whisper 这样可在本地硬件上运行的开源模型。他透露自己正在撰写新书《逆人马 AI 指南》（The Reverse-Centaur's Guide to AI）以提供更犀利的 AI 批判。

Nvidia, CoreWeave, and Nebius: Inside the Circular Financing of the GPU Boom

本文剖析了 AI 算力热潮中新型云服务商（Neoclouds）的“循环融资”机制。CoreWeave 和 Nebius 通过快速部署 Nvidia 最新 GPU 并提供更高的利用率，获得了微软和 Meta 等巨头超过 1200 亿美元的算力合同，大厂借此将资本支出（Capex）转化为运营支出（Opex）。然而，这导致 Neoclouds 自身面临极大的资金缺口和高额债务（如 CoreWeave 负债达 248.6 亿美元）。Nvidia 在其中不仅是供应商，还扮演了投资者（各投资 20 亿美元）和剩余算力担保人的角色，这种循环融资和以 GPU 为抵押的债务模式引发了对 AI 基础设施繁荣可持续性的关注。

Terrorist groups are using every major AI chatbot for attack planning and weapons development

A Cambridge study found that Boko Haram uses AI chatbots like ChatGPT, Claude, and Gemini to plan attacks, build explosives, and maintain weapons. ISIS operatives have been training the group's commanders on how to bypass safety filters since 2023. Given that the study found safety filters repeatedly failed to prevent misuse, voluntary self-regulation by AI providers clearly isn't enough. The article Terrorist groups are using every major AI chatbot for attack planning and weapons development appeared first on The Decoder.

苹果甩出41页PDF怒告OpenAI“偷师”其核心机密!网友:早知道就等印度开源了

When Google altered the routing system for 2% of cars using Google Maps to prefer routes that are just as fast but avoided areas that are congested it...

根据发表在《Nature Cities》上的一项研究，谷歌对2%使用谷歌地图（Google Maps）的车辆调整了路由算法，使其倾向于选择同样快速但能避开拥堵区域的路线。实验结果表明，这一微调提高了全市所有车辆的行驶速度，并降低了整体燃油消耗。

AI Boosts Research Careers but Flattens Scientific Discovery

AI is turning scientists into publishing machines—and quietly funneling them into the same crowded corners of research.That’s the conclusion of an analysis of more than 40 million academic papers, which found that scientists who use AI tool

Scientists’ Side Hustle? Using AI and Quantum Computing to Generate New Peptides

Researchers cobbled together funding and time to show how quantum computing could aid in the development of drugs to help underserved populations and combat rare diseases.

AI agents win at Slay the Spire 2 after researchers replace growing chat logs with structured memory

针对传统 LLM Agent 依赖对话历史导致上下文膨胀（Context Rot）和高延迟的问题，研究人员提出 AgenticSTS 架构，并在《杀戮尖塔2》中进行了测试。该架构用 5 个独立的结构化记忆层（包括协议指令、状态图谱、游戏规则、历史运行总结及战术技能库）取代了不断增长的聊天日志，将 Prompt 长度稳定在 5000 tokens 左右。实验中，该 Agent 在 A0 难度下的胜率达到 60%（而对比的传统 Agent 胜率为 0%），且 Token 消耗仅为竞品的数十分之一。此外，研究发现模型生成的记忆具有强绑定性，在 Gemini、Qwen 和 DeepSeek 之间进行迁移时效果较差。

Negotiating AI in Open Source Software Communities: A Case Study of LLVM Project

该报告是一项以 LLVM 项目为对象的案例研究，探讨了开源软件社区在面对人工智能（AI）集成时，其内部成员如何进行协商、争议妥协以及制定相关政策。研究重点分析了社区对 AI 辅助开发工具及 AI 生成代码的态度、治理挑战与规范制定过程。

AI Bot Test: Can your AI Access AND Understand This?

该链接为一个用于测试AI机器人（如爬虫或AI助手）是否能够成功访问并理解网页内容的测试页面。由于缺乏具体正文内容，其实际测试机制和效果无法进行详细评估。

I Use Multiple AI Agents in My Team's SDLC to Ship Production Features

本文分享了在软件开发生命周期（SDLC）中引入多AI智能体协同工作的实战工作流。该方案采用类似PID控制器的闭环系统：首先由规划智能体（如Opus 4.8）将需求分解为具体实现方案；开发智能体（如Claude Code）在Firecracker沙盒等隔离环境中执行编写与测试；评审智能体（如GPT 5.5）扮演严苛的架构师进行独立审查并生成报告。开发与评审智能体之间不断迭代修正，直至消除主要风险。最后由人类作为协调者把关（必要时手动微调），并配合CI/CD流水线确保代码安全上线。

Your AI's logs can be edited. Its evidence shouldn't be

Kira Project 介绍了一种用于本地 AI 管道的防篡改双层证据架构。该架构包含两层：第一层是基于哈希链串联、渐进式写入的规范执行日志，具备防篡改特性；第二层是与第一层加密绑定的可读推理视图，可导出脱敏版本。该设计的一大特色是记录了传统日志忽略的“负面空间”（Negative Space），例如 AI 无法证实的声明、使用的失效数据源以及被拒绝的操作，从而为 AI 智能体提供不可否认的审计凭证。

You Have an AI Working Agreement. Write It Down

TL;DR: The AI Working Agreement Your team already has rules for using AI. Some live in templates, some in habits, exceptions, and one person’s memory. The AI Working Agreement puts the decisions that matter in one place: what the team deleg

想有好的输出,需要好的输入。 本月计划把时间放在:阅读、看电影、见朋友。 生活需要更丰富的上下文。

作者分享了个人生活计划（阅读、看电影、见朋友），并指出好的输出需要好的输入，生活也需要更丰富的“上下文”。该内容属于个人感悟，无实质性 AI 行业或技术信息。

Unify AI Horn alternative using ESP32, WizNet 5500, Poe and a Ti Amp

该项目在 Hackster.io 上公开，展示了如何使用 ESP32 主控芯片、WizNet 5500 网络芯片、PoE 供电和德州仪器（TI）功放，DIY 构建一个 Unify AI 喇叭（AI Horn）的替代硬件方案。

Parse inbound email to JSON in Node.js – MailKite

MailKite parses inbound email into clean JSON and POSTs it to your Node.js endpoint — no MIME parsing, no mail server, no mailparse dependency. This tutorial walks through the full flow: adding a domain, registering a webhook, verifying the

The Ten Commandments of AI Usage

本文提出了“人工智能使用十诫”，旨在警示过度依赖AI（作者称为“AIpidemic”）所导致的人类思维和专业能力退化。文章强调：不能将核心思考和最终责任外包给AI；AI的解释可能极具迷惑性但依然错误；不可在不理解的情况下直接复制粘贴AI生成的代码或评审；熟练使用AI提问并不等同于掌握了专业领域知识；以及需要定期在没有AI辅助的情况下进行工作，以保持自身的认知能力。

Ask HN: Do your tests clear "If the tests pass, you cannot break the code" bar?

Why or why not?This is how I have found AI to be very powerful. If my tests pass, I know my code won't break. I spend more time on tests than any other part of the code-base. I am curious who else is building in this manner? I use the statement in the title because I feel that is the bar you have to meet to get the AI to work well, but once you meet that bar lately it has been working pretty flawlessly for me.

Show HN: AsyncFutures – A Ruby gem for asynchronous/concurrent code execution

This gem allows you to execute code more-or-less identically across Ractors, Threads, and Fibers. This makes it much easier to test out your code's performance behavior by using an identical API across all concurrency primitives. It is built around a flexible core Future class.This was actually the impetus for creating this Gem: the core Ruby concurrency primitives all have different APIs and subtly different behavior even for the APIs that initially look similar. For example, `join` on Ractor and Thread instances do NOT behave the same. This Gem adds Executors that expose near identical behavior across all of them and return a Future object with identical API and behavior across all of them.I've been working on this Gem for a few months in my spare time. I just published the Gem and made the git repo public a few minutes ago. Please feel free to review the code, test out the gem, and provide your feedback.As a footnote: this was not vibe coded. I wanted to learn through experience how to correctly write concurrent code, so using AI would have defeated the purpose. Also, I don't particularly enjoy writing/generating code via AI, so that would have made vibe coding this doubly pointless.The Gem has 100% line and branch coverage on CRuby 4.0 and is 0BSD licensed.Thank you.

如果你大量工作是基于Codex,对外分享就变得很简单。 只需让 Codex 整理你的所有对话,从中整理项目和经验。 然后输出为飞书文档和PPT大纲。 有大纲后,用自PPT ...

有用户分享了基于 Codex 的工作流：只需让 Codex 整理用户的历史对话并提炼项目与经验，输出为飞书文档和 PPT 大纲，最后利用 PPT 工具或 Codex 内置的生图功能即可快速制作演示文稿，从而简化对外分享的准备流程。

I find AI roleplay therapeutic

本文介绍了一份AI角色扮演（AI Roleplay）的配置指南，重点探讨了如何设置记忆（Memory）和世界书（Lorebooks）以优化体验，并提及了AI角色扮演带来的疗愈效果。

How to Evaluate General-Purpose Robot Policies for Real-World Deployment

随着机器人基础模型的发展，如何严谨地评估这些模型在真实世界部署中的表现已成为行业难题。本文探讨了评估通用机器人策略所面临的关键问题，并介绍了 NVIDIA 应对这些挑战的评估方法。

What happens between entering the prompt and seeing the first word appear

A while back I wrote a post on TurboQuant, about compressing the KV cache to make inference cheaper. At that point I was reasoning about the size of the KV cache without knowing how it gets filled in the first place, or where the Keys and V

Fixed three bugs that made Qwen3.5-122B a daily driver on Mac Studio

10 July 2026·12 minsA follow-up question on a 50,000 token conversation took three to five minutes before the first token appeared. Not the full answer. The first token. That is not a chatbot, it is a batch job, and you go and make a cup of

Ask HN: Has single-task focus become outdated in the AI era?

I’ve always found that deep focusing on one task at a time was the only way to get things done at work with good quality and at an acceptable pace, and I have a real hard time multitasking (even with AI) because of context switching and getting distracted or overwhelmed.But with AI making parallel work easier (in theory), is single-task focus now a disadvantage? Are effective engineers always running several threads at once, or is deep focus still the best way to get things done?

Local-first agent governance: keeping an AI agent contained

Design your agent's context like a filesystem: the ICM methodGuide · 2026-07-06 · By Vanta and ElifA plain-English guide to ICM (Interpreted Context Methodology) — folder structure as agent architecture — and pairing it with local search to