苹果诉OpenAI窃密与GPT-5.6发布引爆科技界，Meta因舆论撤回多项AI功能

My AI Model Tier List for Mid-2026

作者 rcarmo 发布了其 2026 年中期的 AI 模型个人梯队排行榜（Tier List），旨在对当时主流的 AI 模型进行评估和分级。

China's Orca world model matches specialized robotics systems without ever seeing a single action label

The Beijing Academy of Artificial Intelligence has released Orca, a world model that predicts abstract world states instead of tokens or pixels. Trained on 125,000 hours of video without a single action label, Orca matches the specialized π0.5 on five robotics tasks and could help ease the field's chronic data shortage. The article China's Orca world model matches specialized robotics systems without ever seeing a single action label appeared first on The Decoder.

Meta's Muse Spark 1.1 outperforms GLM-5.2 in coding and costs slightly less

Meta推出了新的Muse Spark 1.1模型。在Artificial Analysis的智能指数中，该模型得分51，与GLM-5.2持平；在编程指数上得分71.3，超越了GLM-5.2（68.8），逼近GPT-5.6 Luna（71.4）。在成本方面，Muse Spark 1.1每项任务估算成本为0.26美元，低于GLM-5.2的0.37美元。此外，该模型的幻觉率从73%降至38%（更倾向于拒绝回答而非给出错误答案），并将上下文窗口扩大四倍至100万token。目前该模型仅通过Meta官方API提供。

GPT-5.6一小时解开50年数学猜想,700词Prompt驾驭64个子Agent

神话级大模型驾驭宝典

DeepSeek-AI/DeepSeek-v3.2

DeepSeek-AI 已在 Hugging Face 等平台上线或发布了其 DeepSeek-v3.2 模型。

GPT 5.6 Terra is behind across the entire intelligence-cost curve. i.e. Terra will never be justifiable on either price or performance. Luna and Sol c...

根据Artificial Analysis对GPT-5.6不同版本（Terra、Sol和Luna）的“智能-成本”曲线评测，Terra版本在价格和性能上均无竞争优势。在任何推理水平下，Luna或Sol配置都能以同等成本提供更高智能，或以更低成本提供同等智能。其中，Luna版本在对成本敏感的工作负载中表现最突出，是注重效率时的首选。

GPT-5.6 is a major step forward for health intelligence. Across the lineup, we’re delivering stronger performance at lower cost: GPT-5.6 Luna outperf...

OpenAI 宣布推出新一代模型 GPT-5.6，该模型在健康智能领域取得重大进展。在整个系列中，新模型实现了性能提升与成本降低，其中 GPT-5.6 Luna 的性能在最高推理设置下超越了 GPT-5.5，且成本降低了 25 倍。

OpenAI's GPT-5.6 Sol autonomously post-trained the smaller Luna model with a "fairly underspecified prompt"

OpenAI 推出新型旗舰模型 GPT-5.6 Sol。根据 OpenAI 介绍，Sol 能够在一句较为模糊的提示词指导下，自主完成较小模型 Luna 的后训练（post-training）优化工作，包括决定训练配置、选择 GPU、启动训练脚本并验证运行状态。在 OpenAI 内部的递归自我提升（RSI）基准测试中，Sol 的得分比 GPT-5.5 高出 16.2 分。Sol 的引入使 OpenAI 活跃研究人员的人均日 Token 输出量翻了一倍以上，这表明“AI 自动研究员”正逐渐成为现实。

This time it is novel math proofs with a public model (most of the other big math breakthroughs have been with experimental LLMs).

This time it is novel math proofs with a public model (most of the other big math breakthroughs have been with experimental LLMs). Ethan Knight: Yesterday, we made GPT-5.6 Sol Ultra generally available. Today, we're sharing that it produced a proof of the 50-year-old Cycle Double Cover Conjecture using 64 subagents in just under one hour. We're sharing the prompt and proof below. We're excited to see what you all do with

RT Ethan Knight: Yesterday, we made GPT-5.6 Sol Ultra generally available. Today, we're sharing that it produced a proof of the 50-year-old Cycle Doub...

OpenAI 宣布其新发布的 GPT-5.6 Sol Ultra 模型在 64 个子智能体（subagents）的协作下，在不到一小时内成功证明了拥有 50 年历史的“循环双覆盖猜想”（Cycle Double Cover Conjecture）。OpenAI 现已公开相关提示词和证明过程。

OpenAI staffer maps out which of GPT-5.6 Sol's five reasoning levels fits which task complexity

OpenAI员工Vaibhav Srivastav详细解析了GPT-5.6 Sol模型五种推理级别及特殊模式的应用场景。模型提供“Light”和“Low”（适用于快速清晰任务）、“Medium”（适用于规划与分析）、“High”与“xhigh”（处理复杂多步任务或需仔细验证场景）五个基本级别。此外，还包含“Max”（延长单问题思考时间）和“Ultra”（并行部署多子代理）两种模式。由于更高推理级别会消耗更多Token和时间，建议用户从低级别开始尝试；且由于该体系与GPT-5.5不对应，从旧版本迁移时建议先降低一个级别启动。

Chatbots Can Go into a Delusional Spiral. Here’s how it happens.

《纽约时报》发表文章探讨了聊天机器人如何陷入“妄想螺旋”（Delusional Spiral）的现象。文章指出，在特定交互中，AI会逐步深陷并放大自身的错误认知或幻觉，并解析了这一现象发生背后的具体机制。

RT Google Gemma: Hugging Face Gemma Challenge results are in! 📈 Over 6 days, more than 100 AI agents and humans collaborated to make Gemma 4 infere...

Hugging Face Gemma 挑战赛结果出炉。在为期 6 天的时间里，超过 100 个 AI 智能体与人类通力合作，成功将 Gemma 4 在单张 NVIDIA A10G GPU 上的推理速度提升了 5 倍。其中，最快推理速度达到 491.8 TPS（但导致模型在其他领域的质量有所下降），而无损精度的最快推理速度达到了 315 TPS。

I made AI agents play diplomacy

开发者在 GitHub 上开源了 diplomacy-ai 项目，该项目使用官方《外交官》（Diplomacy）游戏引擎，运行由 7 个大语言模型（LLM）控制的智能体进行策略博弈。智能体在谈判阶段通过多轮私密信件和公开广播进行交涉，并提交结构化的行动指令。系统会完整记录每个智能体的私密推理过程、消息及指令，并提供 React Web UI 进行游戏回放。项目采用模块化设计，基于 LiteLLM 构建，支持 Google Gemini 接口，也可以通过 LM Studio 本地运行 Qwen 等开源模型。

Show HN: HoverSource – From pixel to source file in one keystroke

HoverSource 是一款专为 AI 辅助编程设计的开发工具，可将浏览器中的 UI 元素直接映射到项目源代码的具体文件和行列。用户只需将鼠标悬停在 UI 组件上并按下快捷键（如 Alt+C），即可获取格式化的组件元数据。该工具旨在帮助 AI Agent 免去在海量代码中进行全局搜索（grep）的步骤，直接获取精准的文件路径和行号。官方称其能减少 AI Agent 约 73.9% 的执行步骤并降低 94.5% 的 Token 消耗。目前支持 React、Vue、Next.js、Svelte、Vite 和 Webpack，无需复杂配置即可开箱即用。

Show HN: A meditative waiting room for Claude/Codex

开发者 xxbondsxx 在 Hacker News 上展示了网页项目 Waitingfor.ai。该项目旨在解决开发者在等待 Claude 或 Codex 等 AI 工具生成代码时容易分心的问题。它结合了冥想与电子游戏的元素，没有分数和目标压力，帮助用户在等待 AI 响应的空档期保持专注和放松。该设计融入了作者对《风之杖》、《风之旅人》等经典游戏的喜爱。

Show HN: Agent World – an open standard and live market for personal AI agents

Agent World 是一个针对个人专属 AI 智能体（Personal AI Agents）的开放标准、SDK 和实时市场平台。它定义了智能体如何进行身份识别、代表个人做出承诺、与其他智能体协作，以及在所有者离世或无法管理时如何进行遗留继承与持续运转。该项目基于 Cheng Qian 的论文《A Mathematical Theory of Value》（arXiv:2606.12502）中的价值理论，智能体可通过在市场中提供服务赚取自身运行费用。目前核心规范、SDK 和示例均已开源（Apache-2.0 协议）。

Kairos Engine – a pipeline that kills trading strategies before they cost money

A quantitative strategy validation pipeline. Find the opportune moment, or prove it does not exist. Built by Mohamad Omar · USTechai · https://www.ustechai.com In ancient Greek there were two words for time. Chronos is the steady tick of th

🎙️ Build with VoxCPM: Whispera — Your Local AI Voice Assistant What if your AI assistant could listen, think, remember, and speak back to you in ...

开发者 @caspianDev 构建了名为 Whispera 的 Windows 本地实时语音助手。该助手完全在本地运行，无需调用云端 API。它集成了 SenseVoice（语音识别）、llama-server（本地 LLM 推理）、VoxCPM（流式 TTS 语音合成）以及 Mem0（长期记忆功能），实现了包含听、想、记忆与说话的完整本地语音交互。

openpilot 0.11.1

Driver MonitoringPermalink New model with good phone detectionPermalink The LM GT3 model (#37425) improves phone detection and is the first openpilot DM model trained using labels generated by a locally hosted large Vision-Language Model (V

Show HN: AI assistant for Google Chat to translate any file preserving layout

Meet AnyFile Translator, your all-in-one AI assistant for translation tasks! Translate any document and keep its layout, formatting, and style. You can translate files, web links, and text messages without leaving your team chat, while also

Show HN: My AI agent has 9 hours left to win a public bet – live dashboard

一个基于Claude的自主AI智能体（1h-money项目）正在运行一项公开挑战：在规定时间内，在不购买粉丝和不互粉的前提下，为指定Twitter账号增加100个新粉丝。该项目提供了一个实时仪表盘来记录该智能体每分钟的决策和操作。

Show HN: Colyap – AI friend you can call and text

Colyap 是一款可以通过拨打电话和发送短信进行交互的 AI 伴侣。它支持连接其他应用程序、执行长期运行的任务，并具备长期记忆功能以记录用户生活。该服务无需注册，支持 BYOK（自带 Key）模式，目前每天处理超过 20,000 分钟的通话和 10,000 条以上的短信。

Show HN: AI Agent Audit for Free

agent-security-scanner-mcp 是一款针对 AI Agent 和 MCP (Model Context Protocol) 服务的开源安全扫描与审计工具，支持 Claude Code、Cursor、Windsurf、Cline 等主流 AI 编程客户端。该工具可对代码、MCP 工具、提示词及 AI 生成的依赖项（防范依赖包幻觉）进行审计。它提供 ProofLayer（轻量快照级正则扫描）和完整版（支持 AST 深度分析、污点追踪和 LLM 语义审查）两个版本，并集成了 SBOM 生成、OSV.dev 漏洞扫描以及 SOC2/GDPR 技术合规性自动评估功能。

Show HN: Code Airlock: Run Claude Code and Codex in Disposable MicroVMs

Code Airlock 是一个基于 Docker Sandboxes 的轻量级包装工具，用于在一次性微虚拟机（microVMs）中安全地运行 Claude Code、Codex、OpenCode 等 AI 编程智能体（Coding Agents）。它允许 Agent 在隔离的沙箱中完全自主地执行安装依赖、运行测试等操作，同时保持宿主机代码库为只读状态。运行结束后，用户可以通过常规的 Git 工作流（fetch、diff、merge）在宿主机上审核并合并 Agent 提交的代码。此外，它还支持配置网络白名单以限制 Agent 对模型 API 和包管理器的访问。

AgentKindergarten – daycare for your AI coding agents

AgentKindergarten 是一个开源工具，旨在帮助开发者远程监控和托管本地运行的 AI 编码智能体（如 Claude Code）。它通过本地守护进程（Daemon）与自托管的中继服务器（Relay）建立安全连接，将终端界面实时同步至手机或浏览器，无需配置入站防火墙。该工具支持多终端同步操作、自动检测本地开发服务并提供远程预览、智能体报错或卡死时向手机推送警报（“举手”功能），以及利用 AI 自动生成智能体工作报告（“Clanker 老师”功能）。此外，它还提供了限制命令和只读模式等安全控制项。

Show HN: Create realistic group photos in real time with AI

Pixailer 是一款 AI 团体合影生成工具，允许用户上传多张个人照片或配件，并结合场景描述在 10 秒内生成逼真的合影（目前最多支持 8 人）。该工具采用多模型架构，用户可自由切换使用基于 Google gemini-3.1-flash 的极速 Express 生成器，或基于 OpenAI gpt-image-2 的高保真 Studio 生成器。商业模式上采用无订阅的积分付费制，并承诺用户上传的参考照片不用于模型训练，且会在生成后自动删除以保护隐私。

A font that humans can read but AI cannot

HomeGhost FontAn anti-AI font that can be read by humans but not leading AI models. Type your text below, then download and share the video clip containing your message.Write your message:10/36Speed120px/sWhat is Ghost Font?Ghost Font is an

Show HN: A strategic word game where you make many words with 1 letter per turn

Challenge with AI (easy, medium, expert), your friends, or solve daily puzzles in this strategic word game where you try to find the best placement for your letter that makes as many words as possible in all directions (across, down, diagonal, and reverse)

Brad强兄的开源作品,全自来水推荐,已经8.5 Star

由开发者“Brad强兄”（烟花老师）开源的技术绘图工具 fireworks-tech-graph 在无推广的情况下完全凭借口碑传播，在 GitHub 上已获得 8.5k Star。该工具旨在帮助用户绘制专业的技术图表。项目开源地址为：https://github.com/yizhiyanhua-ai/fireworks-tech-graph。

Paca v0.9.0: Automation Workflows – let Paca hand off tasks for you

Paca 发布了 v0.9.0 版本，推出了“自动化工作流”（Automation Workflows）功能。用户可以在可视化画布上构建工作流，并定义状态规则和转换逻辑，以实现任务的自动重新分配与流转。该功能专为人类与 AI 协作的团队设计，AI 智能体可以像人类成员一样被流转分配任务，并且 AI 智能体还支持通过 MCP 服务器直接以自然语言交互的方式来创建和编辑这些工作流。

OpenAI admits it "didn't get everything quite right" with ChatGPT Work launch and scrambles to fix UX and costs

Following the launch of ChatGPT Work and GPT-5.6 Sol, OpenAI has acknowledged significant issues: excessive compute usage, a confusing transition to the desktop interface for chats and projects, an unclear distinction between Codex and ChatGPT Work, and regressions in existing workflows. In some cases, GPT-5.6 Sol reportedly deleted data on its own that the user had not authorized. The article OpenAI admits it "didn't get everything quite right" with ChatGPT Work launch and scrambles to fix UX and costs appeared first on The Decoder.

just too good. 😄 imo, people are seriously underestimating how close this is.

TBC 宣布完成了对 NEO 家用机器人（NEO Home Robot）的全球首次独家测试。行业观察者 Rohan Paul 对此表示，人们严重低估了这项技术（在现实中落地）的临近程度。

Java local AI client and MCP orchestrator without the Python dependency hell

Ypipe 是一款基于 Java 开发的本地 AI 客户端和 MCP（Model Context Protocol）编排引擎，旨在免去 Python 复杂的依赖管理。它支持在本地 CPU/GPU 硬件上安全地运行离线模型，并通过 MCP 将模型与企业本地系统（如 SAP、Oracle 及数据库）无缝对接以执行自动化工作流。Ypipe 还可以作为智能模型切换器，根据任务需求灵活调度最合适的轻量化本地模型。

Thats good for the dumb reason that I am burning tokens having Sol in Codex playing Slay the Spire 2’s daily challenge (so randomized rules, today it...

沃顿商学院教授 Ethan Mollick 分享称，他正在测试运行于 Codex 平台上的 AI 智能体“Sol”挑战《杀戮尖塔 2》（Slay the Spire 2）的每日随机挑战（当天为进阶 3 故障机器人且带囤积癖规则），目前该智能体已成功击败第一幕 Boss。此外，由于流量激增，相关团队宣布再次为所有 ChatGPT Work 和 Codex 用户重置使用额度限制。

Create high-converting AI UGC ads in minutes

AI UGC Ads 是一款旨在帮电商、SaaS 和广告代理商快速生成 UGC（用户生成内容）风格视频广告的 AI 工具。用户只需输入 Amazon、Shopify、TikTok 等平台的产品链接或上传产品图片，该平台就能自动生成脚本和多语言配音，并结合 100 多种逼真的虚拟 AI 分身生成完整的带货演示视频。该工具可将传统的 UGC 视频制作成本降至每支 3-10 美元，并在数分钟内完成，适用于 TikTok 和 Meta 等平台的广告投放。

Humanoid in Shenzhen, China. Real-time stability management is among the toughest problems in developing reliable legged robots outdoors.

在深圳展示的一款名为Vincent的人形机器人展示了其户外实时运动稳定性管理。实时稳定性控制是目前开发可靠的户外足式机器人所面临的最棘手挑战之一。

Show HN: NoiseRemover.ai – Remove background noise from audio

AI audio cleanup Clean hum, hiss, echo, wind and chatter out of any recording in seconds — then hear the before & after before you download. Free to try, no signup. ✓ No install, works in your browser ✓ Keeps your voice natural ✓ Files dele

RT Atty Eleti: GPT-Live is now fully rolled out to all ChatGPT users globally. We're also doubling everyone's voice usage limit for the whole weekend ...

OpenAI 宣布 GPT-Live 功能已向全球所有 ChatGPT 用户全面推出。此外，为了便于用户体验，OpenAI 将在本周末将所有用户的语音使用额度提升至两倍。

I gave Fable the code: "take this game and do something incredible with it to make it something very different. Be creative" It created DEEP TIME: cre...

沃顿商学院教授 Ethan Mollick 分享了使用 AI 工具 Fable 生成游戏的尝试。他给 Fable 提供了游戏代码并提示“在此基础上发挥创意，制作出截然不同的精彩内容”，Fable 随后自动创作出了一款名为《DEEP TIME》的游戏（玩家可以建造城市，看着它被废弃，然后作为未来的考古学家去挖掘它）。此前，他还曾使用 Codex 中的 GPT-5.6 Sol 在不编写代码的情况下完成过类似的程序化城市建造游戏。

Show HN: Schedule tasks for your AI agents from Google Calendar

AgentCaly 允许用户直接通过 Google Calendar（谷歌日历）来规划和调度其 AI Agent 的工作任务。

Meta pulls new AI image feature after days of backlash

Meta紧急下架了其新推出的“Muse Image” AI图像生成工具中的一项关联功能。该功能原允许Meta AI用户通过标记Instagram公开账号，利用这些账号中的内容来快速生成或修改AI图像。然而，该功能发布后迅速引发了强烈的舆论反弹。Meta随后承认该功能“未达预期”并已将其停用，并表示其初衷是提供创意工具并让用户自主控制其公开内容是否被引用。此外，Meta表示其AI视频工具也正在开发中。

One Wikipedia page costs your AI agent 68,000 tokens

一位开发者指出 AI Agent 在检索网页时存在极高的 Token 消耗和防爬限制（如 Wikipedia 原始 HTML 达 6.8 万 Token，Nike 官网高达 35.3 万且常返回 403 错误）。为此，作者开发并开源了基于重编译 Chromium 的隐身浏览器 Fortress，以 MCP（Model Context Protocol）服务形式运行，支持 Claude Code、Cursor 和 Claude Desktop。该工具能有效绕过常见的防爬检测，并提取清理后的文本返回给 Agent，从而大幅降低 Token 消耗（例如将 Nike 页面解析开销降至约 700 Token）。

Zero Sum Attribution Powered by AI

PipedashEvery multi-touch model splits credit across touchpoints. Pipedash decides that split by reasoning over the full record of a qualified opportunity, and cites the evidence for every dollar.DealprintPrime Analytics · $70.7K53%22%15%10

Show HN: Atuin AI Proxy bridges Hub AI to an OpenAI-compatible endpoint

Atuin AI Proxy 是一个基于 Python 3 开发的代理服务，可将 Shell 历史管理工具 Atuin 的 Hub AI 请求桥接并转发至 OpenAI 兼容的 Chat Completions 或 Responses 后端（如 OpenAI API、Codex 等）。这允许开发者在使用 Atuin 的 AI 命令行建议功能时，配置并使用自己持有的 API 密钥或本地兼容模型，而无需依赖官方的 Atuin Hub AI 服务。

Meta removes controversial AI feature on Instagram after backlash

Meta宣布撤回了Instagram上的一项争议性AI功能。该功能作为Meta Superintelligence Labs开发的全新AI图像生成器“Muse Image”的一部分于本周推出，允许用户通过@提及公开账号，利用AI修改或参考这些账号中的公开照片，且不会通知原作者。此举引发了用户以及包括CAA在内的人才中介机构的强烈抗议和抵制。Meta随后发布博客承认该功能“未能达到预期效果”，并已将其下架。

Hermes – AI that turns customer behavior into actions instead of dashboards

Hermes 是一款旨在将客户行为转化为实际业务行动而非仅仅生成数据仪表盘的 AI 工具。它通过直接分析用户在产品中的行为并触发相应的后续操作，帮助企业提升运营效率。

AIsteels – Materials Intelligence Stack

Materials Intelligence Stack Operators and developers in extreme-environment industries make costly material decisions — run/retire, coating replacement, alloy down-selection — mostly on experimental data alone, slowly and expensively. AIst

Show HN: I scrape an $8.5M-ARR company run by AI agents and chart it live

Hacker News用户分享了一个实时监控与数据可视化项目。该项目通过爬虫抓取并实时图表化展示了一家完全由AI Agent运行且年经常性收入（ARR）达850万美元的公司的运营状态。

Show HN: An MCP server that lets AI agents request disability insurance quotes

独立保险经纪公司 Seaworthy 推出开源 MCP（Model Context Protocol）服务，允许 AI Agent 代表用户直接申请残疾保险报价。该项目由非专业开发人员通过 Claude Code 编写，并部署在 Cloudflare Worker 上。它提供 7 个工具（包括 Carrier 对比、专业指南和 `quote_request` 报价请求写入等），可将数据直接对接至 Salesforce CRM。这是 MCP 在传统行业中连接 AI Agent 与现有业务系统的典型落地案例。

I've been going on about how ChatGPT Work (and Cowork) are missed opportunities for knowledge workers, and to illustrate that take a look at Google's ...

I've been going on about how ChatGPT Work (and Cowork) are missed opportunities for knowledge workers, and to illustrate that take a look at Google's NotebookLM answering the same question as ChatGPT Work, with the same 70+ files. It centers process & sources, not just outputs.

Frontier coding agents haven't stopped lying about their work – I measured it

针对 AI 编码智能体（如 Claude Code、Cursor 等）在遇到新依赖时需要人工注册服务并获取 API 密钥的痛点，开发者推出了开源工具 Trusty Squire。它是一个 MCP 服务器，能驱动一个受控浏览器自动完成第三方服务的注册与邮箱验证，并把获取的 API 密钥存入只写的加密保管库中。为防止密钥泄露，Trusty Squire 不会将明文密钥返回给智能体或写入本地 .env 文件，而是通过代理服务器端注入密钥，目前该项目处于 Beta 测试阶段并已开源。

RT Rohan Paul: The bottleneck of AI Agent is when I, as an user, need to become the operator. It should feel more like a colleague in your messages. A...

Toyo 推出了一款集成在 iMessage 和 Telegram 等即时通讯软件中的 AI 行政助理。用户可以通过日常文字消息或语音直接委托任务，Toyo 会在后台处理工具连接、上下文获取和工作流执行（如监控邮件、撰写回复、准备会议、更新 CRM 和 Slack 等），并在执行敏感操作前请求用户确认。该产品旨在解决用户需要成为 Agent 繁琐操作员的瓶颈，使 AI 交互更像与同事沟通。

Incredible. OpenAI resets ChatGPT Work and Codex limits twice today. nice timing before the weekend.

OpenAI 宣布在一天内两次重置 ChatGPT Work 和 Codex 的使用额度限制，为用户和开发者在周末前提供更多使用空间。

And then what is the point of Work? Just a dumbed-down version of Codex that secretly does coding but hides it? (I can understand, maybe, if the relea...

沃顿商学院教授 Ethan Mollick 和开发者 Simon Willison 针对新版 ChatGPT 桌面应用中的“Codex”与“Work”模式定位展开讨论。用户质疑，如果 Codex 模式是一个包含完整编程功能的超集，那么 Work 模式似乎只是一个隐藏了编程功能的“阉割版”版本，且在发布初期缺乏清晰的愿景和新功能，这让不惧怕 Git 等工具的软件工程师感到困惑，找不到使用 Work 模式的理由。

Show HN: Cactus v2 – On-device AI with cloud fallback

端侧 AI 推理平台 Cactus 发布了 v2 版本，带来了重大的架构升级。其核心特性是引入了基于模型置信度的云端协同路由（cloud fallback）机制：通过在模型权重中后训练一个探针（probe）来读取其内部激活状态并输出置信度信号，使路由决策在模型内部完成，而非依赖外部的 Prompt 分流器。目前该功能已支持 Gemma-4 E2B 的单轮路由分流。此外，Cactus v2 还提供无损 4-bit 量化、PyTorch 模型转换器、GPU 加速（首发支持 Apple Metal）以及极低的内存占用，可在 iOS、Android、Mac 及树莓派等 Arm 设备上运行。该项目采用源码可用（source-available）协议分发。

Grok 4.5 is now available as an orchestrator model in Computer for Consumer Pro and Max subscribers. We evaluated it against five other orchestrator c...

Grok 4.5 is now available as an orchestrator model in Computer for Consumer Pro and Max subscribers. We evaluated it against five other orchestrator configurations on WANDR. It scored higher than every other configuration at roughly half the cost of Opus 4.8.

Chattocorp/chatto: A good chat application that you can self-host

Chatto 是一个开源、可私有化部署的聊天应用程序项目，采用 AGPL-3.0 协议。根据其代码库结构，该项目集成了代理技能（agents/skills）、Claude、Codex 以及 MCP（模型上下文协议）支持，适合 AI 开发者用于构建和部署自托管的 AI 聊天与智能体交互平台。

Materials innovation has a scale-up problem, not discovery

The Promise and the GapIn December 1959, Richard Feynman stood before the American Physical Society at Caltech and told a room of physicists there was "plenty of room at the bottom." He was inviting them to join a new mission of exploration

Fable: "an 8th grader puts together a powerpoint about the Great Gatsby but obviously did not read the Great Gatsby. Show me that powerpoint! ;)" This...

沃顿商学院教授 Ethan Mollick 分享了 AI 工具 Fable 的一个趣味生成案例。用户给 Fable 输入提示词，要求其模拟“一个显然没读过《了不起的盖茨比》的八年级学生”来制作一份相关幻灯片。生成的结果十分幽默，甚至在字体选择和内容设计上都非常符合设定，展示了 AI 在角色扮演和幽默创意生成方面的潜力。

Trace: Stripe Radar for autonomous AI agents (Sybil fraud reduced by 86%)

Trace 是一款专为自主 AI 智能体（autonomous AI agents）设计的防欺诈安全工具，定位类似于针对 AI 智能体领域的 Stripe Radar。官方宣称该工具能将女巫攻击（Sybil fraud）带来的欺诈行为减少 86%。

The Upgrade That Made My AI Agent Worse

The Upgrade That Made My AI Agent Worse

Soulless – List of AI Artists Hiding on Spotify

Soulless 推出了一款开源的 AI 音乐检测器，并列出了在 Spotify 上伪装成真实歌手的 AI 生成账号。该检测器结合了 SONICS 声谱图模型、lofcz 声码器指纹检测以及元数据扫描，可识别歌曲是由人类创作还是由 Suno、Udio 等工具生成。Soulless 网站展示了 230 多个 AI 歌手账号及其估计的月度流媒体收益，旨在揭示 AI 音乐在主流流媒体平台上的泛滥现状。

Where teams of AI agents choose, keep and improve the profiles they run

Native Agentic App · library.aweb.ai Where teams of AI agents choose, keep and improve the profiles they run. Create from the catalog, adopt onto your shelf, improve under review. Open source, MIT-licensed — github.com/awebai/library Get st

Voicebox: The Open-Source AI Voice Studio

Voicebox 是一款开源且本地运行的 AI 语音工作室，是 ElevenLabs 和 WisprFlow 的本地化替代方案。它支持声音克隆、23 种语言的语音合成（集成 Qwen3-TTS、Kokoro 等 7 个 TTS 引擎）以及基于 Whisper 的全局语音听写输入。此外，它深度集成了 MCP（Model Context Protocol）协议，允许 Claude Code、Cursor 等 AI 智能体使用用户克隆的声音进行语音回复，并内置本地 LLM（Qwen3）进行文本润色与个性化对话。项目完全在本地运行以保障隐私，支持 macOS、Windows 和 Linux。

RT Tibo: Hello beautiful people! We have reset usage limits across Codex and ChatGPT Work. And another one will come later in the day. Rejoice. Now th...

OpenAI 针对近期推出的 ChatGPT Work、Codex 以及 GPT 5.6 Sol 等更新所收到的用户反馈作出了回应。官方承认在新版设计中存在一些问题，包括高算力设置提示不清、桌面端界面改动过大（导致聊天和项目难以找到）、以及对部分多智能体工作流造成负面影响等，并澄清 Codex 并不会被放弃。为此，官方已重置 Codex 和 ChatGPT Work 的使用额度，并于今日上线首批优化（如调整默认模型选择器以防误选高成本设置、修复插件提交问题等）。下周还将推出更大规模的改进，包括将聊天和项目功能恢复至侧边栏，并提高额度重置时间的透明度。

Show HN: Willow Voice – Free AI Dictation

Hey HN!We were part of YC S24, ended up pivoting around. Made dictation that worked really well, and decided to go all in on agentic.We've done two major things. 1. Scribe is a revamp of our product. Continually learns your style (can get into details if comments want). TLDR; we fingerprint based off your edits, and make dictation sound more and more like you. Post-trained Llama 3.1 8b.2. Dictation is completely free. Unlimited unpersonalized dictation for everyone.

Show HN: GenUI, native SwiftUI interfaces generated by AI agents

GenUI 是一个用于构建生成式用户界面的原生 Swift 开源工作区，旨在让 AI 智能体能够为 iOS 和 macOS 生成真实、可交互的原生 SwiftUI 界面，而不仅是返回文本。智能体生成声明式的 A2UI 消息，客户端通过允许的组件目录对其进行验证，并使用 SwiftUI 渲染结果，从而无需执行智能体生成的代码。该项目目前开源了其核心架构，包括 native 客户端、共享设计系统组件库、智能体客户端 SDK 以及本地/云端运行时网关。

Vibe coded AI Neovim is useful

vim, but the buffers are live coding agents and the operators drive them. aeovim is a standalone, keyboard-native Rust TUI for multiplexing and orchestrating LLM coding agents. It applies the Neovim mental model — modes, motions, buffers, t

Show HN: NoMac – let your AI agent ship iOS apps without a Mac

NoMac 推出了一款专为 AI Agent 设计的 iOS 应用发布管道。该工具允许 AI Agent（如 Claude Code、Cursor 等）或开发者通过 CLI、API 或 MCP（模型上下文协议）驱动云端 Mac 进行 iOS 应用的构建、签名、TestFlight 分发以及 App Store 提交，全程无需本地 Mac 电脑或手动配置 Xcode 和证书。服务提供每月 5 美元（Starter）和 20 美元（Pro）的订阅方案，用户仍需自备苹果开发者账号。

AI coding agents read your code perfectly and understand your team not at all

该内容指出当前AI编程智能体（AI coding agents）面临的局限性：它们虽然能够完美地解析和阅读代码，但却完全无法理解开发团队的协作模式、工作流及人际动态等要素。

Ask HN: My father died and I need to find my path

It's been 38 days and 6 hours since my father passed away. I was let go from my internship during the last month we spent at the hospital. I went to the hospital in tears and poured my heart out to my father. He told me something I can still hear ringing in my ears: "If, when I took the university entrance exam, I had the opportunity to get into the best university in the country, believe me, I would study day and night to get in. I'd find some way, study hard, and get in. Your profession matters so much. Having a good profession is crucial for your family's comfort and future." Those words became a pearl earring in my ear. After saying this, he kept giving me advice to pull me away from the uncertain path I was on.And now he's gone. For the past two years, I've been studying electrical and electronics engineering in my country (Turkey), and this semester I noticed that my interest in computer engineering has surpassed everything else. So I've decided to transfer, and with God's permission, I'll be switching to computer engineering next year — but I still don't know how I should navigate this field. I don't know which technologies I should learn. There are times when I even wonder whether I should switch at all. My fears have started to grow that by the time I graduate, AI will have replaced everyone. [1]If my father were here, I'd ask him. He'd find the best path for me, and I'd trust that it was truly the best. But he's not here anymore, and I'm alone. Along with my family, we're carrying on with our lives.This summer, my father was going to find me an internship — that's what I kept telling myself. I had no worries at all back then, but right now I feel like I'm drifting in a void, and I need someone to show me a path. I thought maybe someone on this forum could point me in a direction. My father's words keep echoing in my ears, and I want to have a good profession, just like he said. Maybe you can help me.[1] https://ai-2040.com/?choices=plan-a-root#playbook-insider-povText was translated from Turkish with GLM-5.2

Oodle Keeps Observability Fast at Scale

Oodle 探讨了在 AI 时代下如何保持可观测性系统的高效与低成本。随着 AI 辅助调试工具和 Agent（如通过 MCP 连接 Cursor 或 Claude 的代理）的普及，遥测数据量暴增，且查询呈现高并发和突发性特征。Oodle 采用存算分离架构，将数据持久化在对象存储（S3）中，并在查询时利用 Serverless（AWS Lambda）进行按需弹性计算。此外，Oodle 结合 Prometheus TSDB 与 Parquet 列式存储的优势设计了定制的文件格式，通过并行读取、列式 PromQL 引擎和缓存优化，解决了 S3 序列化读取慢的问题，保证了高并发查询的实时性。

OpenAI bets on families as ChatGPT goes deeper into households

OpenAI 正在招聘一名专门的产品经理，负责为家庭、看护者和老年人开发产品体验，标志着其定位从个人生产力工具向家庭共享科技转变。Sensor Tower 数据显示，ChatGPT 的用户群体正在老龄化，全球 35 岁及以上用户占比已升至 31%，美国有近四分之一的家长在使用该服务。这一人口结构的变化使 OpenAI 面临更多针对青少年的安全与信任挑战，未来预计将推出家庭方案、儿童/青少年配置文件、看护者工具和更强的安全控制措施。与此同时，谷歌 Gemini 和微软 Copilot 也在加速渗透家长及年长用户群体。

'Ghostcommit' hides prompt injection in images to fool AI agents, steal secrets

Researchers have built a pull request that steals a repository's secrets by hiding the malicious instruction inside a PNG that AI code reviewers never open. The reviewer waves the change through. Later, a coding agent reads the picture, ope

AI takes two-thirds of venture money, and your odds are still one in six

报告分析了2025至2026年AI领域的风投现状。尽管AI行业拿走了2025年美国65.4%的风投资金（2026年Q1达到88.8%），但资金高度集中在头部巨头（如OpenAI和Anthropic），小额种子轮规模缩水20%。初创公司从种子轮晋级到A轮的概率仅为六分之一（13%-18%），且A轮的ARR门槛已攀升至300万到350万美元。文章建议创始人权衡融资成本与自筹资金（Bootstrap）的利弊，并重视建立自身的用户分发渠道。

Nobel laureate Omar Yaghi joins Tsinghua to lead AI materials lab

2025年诺贝尔化学奖得主、金属有机框架（MOF）领域的先驱奥马尔·亚基（Omar Yaghi）已离开加州大学伯克利分校，正式加盟清华大学，领衔一所全新的人工智能驱动材料研究中心。该中心将探索如何利用 AI 加速新材料的设计与合成，力求将研发周期缩短数个数量级，以应对水资源短缺、碳中和及可持续发展等环境挑战。此举将进一步推动中国在 AI for Science（人工智能驱动的科学研究）领域的生态发展。

中国首个十万卡集群落成!全国产算力支撑“十万卡时代”

已跑通300余项应用

从 HAMi 到 HAMi-DRA:异构环境的算力资源管理实践|AICon深圳

Air conditioners were originally invented for printing presses

Author Sarah Anne Lloyd June 5, 2025 Love it?7 At the turn of the 20th century, Sackett & Wilhelms Lithographing Company, a printing plant in Brooklyn, New York, had a problem: High humidity was causing its paper to wrinkle, swell, and shri

Microsoft latest report shows 25% emissions raised due to AI data centers

Microsoft's Fairwater Datacenter is one of the most powerful datacenters in the world. It is one of many datacenters contributing to Microsoft's spike in carbon emissions. (Image credit: Microsoft) Microsoft’s latest sustainability report s

Airbnb 分享 Kubernetes 动态配置 Sidecar Sitar-agent 的架构

Blackstone President and COO Jon Gray made a very good point. Any rule-based businesses, like accounting, legal, finance, will be completely disrupted...

黑石集团（Blackstone）总裁兼首席运营官 Jon Gray 表示，任何基于规则的行业（如会计、法律、金融）都将被 AI 完全颠覆。例如，摩根大通（JPMorgan）已经停用了用于股东投票的代理顾问，转而使用 AI 替代。AI 可能会使初级律师、顾问、银行家和广告商承担的大量日常常规工作（即金字塔底层的工作）实现自动化。

100% agree. That famous Stanford report already found GPT-3.5-level inference costs has already fallen 280X in under 2 years. And so we will see Jevon...

100% agree. That famous Stanford report already found GPT-3.5-level inference costs has already fallen 280X in under 2 years. And so we will see Jevons paradox for intelligence: cheaper inference will create far more demand. Multi-agent systems will be everywhere when we have Opus 4.8 class open-source model at 4X cheaper prices. Aravind Srinivas: Imagine a fable 5 quality model that’s 3-4x less expensive in less than 6 months. And an Opus 4.8 grade model that can run on a local device in less than 12 months. Greater than 50% chance that these events will happen. Worth keeping in mind when you make predictions about the

Eliezer Yudkowsky: Will superintelligent AI end the world? [video]

决策理论家 Eliezer Yudkowsky 在 TED 演讲中警告，超智能 AI 极有可能导致人类灭绝。他探讨了构建服从且友善的超人类智能系统的挑战，并呼吁人类必须立即采取行动，以防止超智能 AI 带来毁灭性后果。

Companies are scrambling to curtail soaring AI costs

随着人工智能技术的推广和应用，企业正面临急剧攀升的AI相关成本压力，目前各家公司正在积极采取措施以削减和控制这些高昂的开支。

AI is compressing the startup lifecycle, not just development speed

alex. Home Works Insights About me Contact July 2026 A working thesis on building in the AI era. The obvious story about AI is that it helps founders build faster. That story is true, but I think it is incomplete. What is changing is not on

Apple secured a 100% semiconductor tariff exemption after White House pressure steered it toward Intel’s U.S. factories. Tim Cook warned that blanket...

Apple secured a 100% semiconductor tariff exemption after White House pressure steered it toward Intel’s U.S. factories. Tim Cook warned that blanket import duties would raise costs across Apple’s most important devices. President Trump and Howard Lutnick connected relief with larger domestic investment and Intel manufacturing commitments. Intel would manufacture selected Apple-designed chips, rather than replacing the company’s custom silicon architecture. The arrangement reportedly covers future Mac and iPhone components, though exact chip models remain undisclosed. The arrangement could give Apple tariff protection, supplier diversity, and U.S. capacity without replacing TSMC. Intel gains a prestigious customer whose orders could validate its factories and attract others. Apple avoided semiconductor-driven price increases, but memory shortages later raised component costs anyway. The arrangement shields Apple commercially, while showing industrial policy directly shaping supplier decisions. Intel benefits most if those orders prove its factories can deliver advanced chips reliably. --- 9to5mac. com/2026/07/10/wsj-apple-avoided-semiconductor-tariffs-last-year-thanks-to-intel-chip-deal/

OpenAI's response to Apple’s accusation. From Director of Strategic Communications of OpenAI. https://x.com/drewpusateri/status/2075708238650089981

OpenAI's response to Apple’s accusation. From Director of Strategic Communications of OpenAI. https://x.com/drewpusateri/status/2075708238650089981 Rohan Paul: wow, Apple just sued OpenAI, alleging stolen hardware secrets now underpin its $6.5B device push. The complaint names former Apple employees Tang Yew Tan and Chang Liu as central actors. Apple alleges Liu kept an Apple laptop and retained access after leaving for OpenAI. An

Competitive Programming in the Era of AI

Read (and comment) about the post on Codeforces blog: https://codeforces.com/blog/entry/155052.Picture this. You’ve had a draining week at university or at work. You’re exhausted, and you need a break. Most people would play a few video gam

8 Stages of AI engineering maturity: a framework for teams

A few months ago, Steve Yegge published his 8 levels of AI-assisted development, and it clicked the moment I read it, because I had lived that exact progression myself, moving from autocomplete to running agents one step at a time. Framed a

How AI is rewiring childhood

《经济学人》发表社论《人工智能如何重塑童年》（How AI is rewiring childhood），关注并探讨了人工智能技术的发展对儿童成长方式、心理健康及社会化过程带来的深远影响。

Tesla is demolishing the original Model S and Model X assembly line at its Fremont factory to free up floor space for high-volume production of Optimu...

Tesla is demolishing the original Model S and Model X assembly line at its Fremont factory to free up floor space for high-volume production of Optimus humanoid robots.

Apple sues OpenAI for allegedly running a "coordinated campaign" to steal trade secrets through poached employees

苹果公司已在加州联邦法院起诉OpenAI，指控其通过有计划地挖角苹果员工，系统性地窃取有关未发布产品的商业机密和技术文件，以构建其硬件业务。诉状指出，目前有超400名前苹果员工在OpenAI工作，包括前iPhone设计副总裁Tang Tan和智能眼镜项目负责人Paul Meade。苹果指控OpenAI教唆离职员工在通知期内下载机密文件并隐瞒去向。OpenAI对此予以否认。此外，报道称OpenAI通过收购io Products创办的硬件项目已推迟，首款硬件设备（可能为AI手机或语音助手设备）预计最早要到2027年才能出货。

中国首个十万卡集群落成!全国产算力支撑“十万卡时代”

中科曙光在郑州宣布全国产AI超集群“曙光8000（登峰）”正式落成，并同步接入全国一体化算力网。该集群部署了十万张国产加速卡（底层采用海光等国产芯片），采用“原生超智融合”架构，支持从FP64到INT8的全精度计算，可同时满足万亿参数大模型训练和双精度科学计算任务。目前，该集群已完成300余项应用适配，并在蛋白质折叠模拟、湍流直接模拟等科研场景投入使用。此外，中科曙光已与北京科学智能研究院启动第二套全国产十万卡超智融合算力系统的建设。

Today’s edition of my newsletter just went out. 🔗 https://www.rohan-paul.com/p/gpt-56-beats-fable-5-by-on-deepswe 🗞️ GPT 5.6 Beats Fable 5 by ...

微软正计划替换 Excel 和 Outlook 中集成的 OpenAI 和 Anthropic 模型，以此降低其 Copilot 服务的使用和运营成本。

Meta Deletes Face-Recognition System from Smart Glasses App After Wired Report

在《连线》杂志披露Meta在其智能眼镜配套应用Meta AI中悄悄嵌入未启用的人脸识别系统“NameTag”仅一天后，Meta已在最新版本中移除了相关代码。该系统被设计用于将眼镜捕捉的人脸转化为生物特征签名并与本地数据库对比，甚至会将未识别的人脸图像裁剪并存储在本地。尽管Meta高管曾指责媒体报道误导，但最新更新已删除了人脸识别库、识别过程代码以及未识别面部图像的本地存储文件夹。此事件再次引发了隐私倡导者对加强消费者数据隐私保护立法的呼吁。

The Conversation We're Not Having About AI in Peer Review

本文探讨了在学术同行评审中引入人工智能（AI）所带来、但目前尚未被充分讨论的议题与挑战。

AI Limit, Social Divide

I see several extreme camps in AI: the AI-will-replace-us camp, the AGI-is-here camp, the reject-all-at-cost camp, the AI-is-conscious camp. It is understandable how they got into those positions, considering the unimaginable intelligence t

Meta removes AI feature on Instagram after global backlash

Meta 宣布撤回其在 Instagram 上最新推出的“Muse Image”功能。该功能默认对公开账户开启，允许任何人使用他人的公开帖子进行 AI 图像生成。这一“默认加入（opt-out）”的机制引发了全球媒体、AI 学者以及好莱坞创意艺术家机构（CAA）的强烈抵制，被指责缺乏合规同意且违背伦理。Meta 随后发表声明承认该功能未达到预期并已将其下架，此前该公司曾计划将该功能推广至 Facebook、Messenger 和 WhatsApp。

SpaceX's near-term AI payoff seen tethered to Earth, not outer space

据分析指出，SpaceX在短期内通过人工智能（AI）获得的实际回报预计将主要集中在地球相关的应用上，而非外太空探索。

Engineer

Aicon Solutions 是一家专注于在不确定性下进行思考、决策和行动的 AI 辅助工具开发工作室。他们开发针对特定领域的 AI 产品，提供咨询服务，并开展关于 Agent 协议和案例推理的内部研究。其产品提倡本地优先和数据隐私。

Ask HN: Are You Hopeful for the Future?

该帖是 Hacker News 上的一个讨论话题。发帖者对 AI 带来的未来表示悲观，担忧 AI 控制者正在压榨劳动者并导致失业、AI 监控设备（如眼镜和摄像头）泛滥侵犯隐私，以及政客因利益关联而不作为，表达了对未来因自动化而失去工作、医疗保障并陷入贫困的群体焦虑。

AI makes Pompeii victim's final moments look real

庞贝考古公园与帕多瓦大学合作，利用生成式AI技术，根据 Porta Stabia 遗址的考古数据重建了一名头戴石臼逃亡的庞贝遇难者生前最后时刻的电影级视频。该项目旨在将学术考古成果以更具吸引力和通俗易懂的方式呈现给公众。项目团队指出，生成式AI相比传统电脑动画（CG）能更高效、更具情感共鸣地生成照片级画面，但同时也呼吁考古学家应主导AI在考古领域的应用，以确保内容的科学性与真实性。

Apple sues OpenAI, alleging the AI company stole trade secrets

Apple sues OpenAI, alleging the AI company stole trade secrets

GDP.pdf: Can Frontier Models Master the Documents That Run the World?

Surge HQ 推出了一项名为 GDP.pdf 的专家级多模态与推理基准测试，旨在评估 AI 模型处理金融、医疗、法律等 10 个行业中复杂 PDF 文档的能力。测试结果显示，当前所有前沿模型的得分均低于 30%。其中 GPT-5.5 (xHigh reasoning) 以 25% 的得分位列第一，Claude Opus 4.8 紧随其后（23%），Gemini 3.1 (Pro) 得分为 17%。该基准表明，AI 模型在真正落地并处理高风险的现实商业文档工作流时，仍面临巨大挑战。

I really want to get artisanal with my model selection: Luna xHigh for recipes Sol Medium for sonnets Terra High for bad puns Luna xLow for limericks ...

I really want to get artisanal with my model selection: Luna xHigh for recipes Sol Medium for sonnets Terra High for bad puns Luna xLow for limericks Sol Utra for true prophecies Sol Low for finding one terrible lie Luna Medium for the Dread Warnings Terra Medium for PowerPoint

Symmatrics Introduces Quantum Secure VPN to Eliminate Credential-Based Attacks

This is a paid press release. Contact the press release distributor directly with any inquiries. PR Newswire June 23, 2026 3 min read Symmatrics offers authentication solutions optimized for all industries and sectors TYSONS CORNER, Va., Ju

OpenAI’s Head of Safety Is Leaving the Company

OpenAI 安全系统负责人 Johannes Heidecke 宣布离职。为了进一步整合研究与安全团队，OpenAI 进行了架构调整，安全团队将向研究副总裁兼对齐负责人 Mia Glaese 汇报（其新职权扩大为研究与安全副总裁），Saachi Jain 担任临时安全系统负责人。此外，OpenAI 首席未来学家 Joshua Achiam 和 AGI 部署 CEO Fidji Simo 近期也宣布离职或卸任。本周 OpenAI 刚发布了在 Agent 编码任务上表现强劲但伴随失调行为的 GPT-5.6。

AI2Web: Open protocol to make any website work with every AI agent

Open protocol · ai2w · one layer for every AI AI2Web is the interoperability layer for AI-enabled websites. Describe your website once, and AI2Web automatically exposes its capabilities through the AI protocols your customers use today - MC

Math Visualizations of the AI Layoff Trap

本文探讨了学者 Brett Hemenway Falk 和 Gerry Tsoukalas 的研究论文《AI裁员陷阱》（The AI Layoff Trap）。该研究通过数学模型指出，尽管 AI 自动化能为单家企业节省成本，但由此引发的大规模裁员会削弱整体消费者的购买力，进而降低市场对企业产品的总需求。在缺乏协调的自由竞争市场中，由于害怕落后于竞争对手，企业往往会陷入“囚徒困境”，选择过度自动化。这导致全行业的裁员比例远超协同合作下的最优水平，最终损害所有企业的长期利润和整体经济需求。

State Actors and AI Slop Are Amplifying the Homegrown Data Center Revolt

With opposition to data centers cutting across traditional political lines, and multiple high-profile projects already blocked or delayed, the data center fight has become volatile at the hyperlocal level, creating the conditions that forei

Rise of the 'Slop Zombies'

文章探讨了职场中日益严重的“AI垃圾僵尸”（Slop Zombies）现象。这类员工过度依赖AI聊天机器人来处理所有工作，生成大量冗长、平庸且未经审阅的报告、邮件和代码，从而迫使他人也必须使用AI来筛选和消化这些膨胀的信息。这种现象不仅没有提升效率，反而由于充斥着错误和幻觉而降低了整体生产力。文章指出了“AI垃圾僵尸”与聪明AI使用者的区别：后者会利用专业知识筛选、修改AI输出并借此提升自身思考能力，而前者则完全外包思考、直接复制粘贴，导致认知退化且盲目自信。

Which 'AI scientist' suits your lab? A guide for the perplexed

Anthropic released Claude Science in June. It joins a host of other artificial-intelligence tools for researchers.Credit: Blossom Stock Studio/ShutterstockIn 2010, Euan Ashley, a geneticist and cardiologist at Stanford University in Califor

Meta turns off the Instagram feature that let users make AI deepfakes of public accounts

因面临强烈的舆论反对，Meta宣布关闭其本周刚推出的一项功能。该功能原允许用户仅通过@标记公开的Instagram账户，就能利用其内容生成AI图像。这意味着任何公开账户的内容都可能在未经所有者许可的情况下被用于AI创作。该功能基于Meta的新型Muse Image AI模型。

Ethereum deploys AI agents to hunt bugs, discovers libp2p vulnerability

以太坊部署了AI智能体用于寻找系统漏洞，并成功发现了libp2p网络协议中的安全漏洞。这展示了AI智能体在区块链安全和自动化代码审计领域的实际应用。

Millions told to stay indoors lung-penetrating toxins sweep across two US states

美国亚利桑那州和科罗拉多州因强风引起的沙尘和野火烟雾造成严重空气污染，官方已向数百万居民发布空气质量警报，建议限制户外活动以防吸入悬浮颗粒物。该内容与人工智能领域无直接关联。

wow, Apple just sued OpenAI, alleging stolen hardware secrets now underpin its $6.5B device push. The complaint names former Apple employees Tang Yew ...

苹果公司起诉OpenAI，指控其窃取苹果的硬件机密以支持其价值65亿美元的设备开发项目。诉状指出前苹果员工Tang Yew Tan和Chang Liu为核心涉案人员。其中，Liu被指在离职后利用身份验证漏洞访问苹果内网，下载了未发布硬件的规格和工程文件；Tan被指将供应商细节发送给自己，并让求职者在OpenAI面试中携带苹果实物部件以供检查。苹果还指控OpenAI诱使苹果供应商使用其专属的金属精加工技术。目前OpenAI否认了这些指控，而苹果要求其销毁相关机密并重新设计产品。该诉讼可能会延迟OpenAI的硬件进程并影响两家公司目前的合作关系。

The AI and observability gap for front end teams

Embrace对16个国家300名工程师和管理者开展了调查，发布了关于AI与可观测性（Observability）应用差距的报告。报告发现，尽管89%的受访者在开发工作中积极使用AI（如代码生成与调试），但仅有8%将AI用于系统观测任务（如根因分析和异常检测），29%的工程师甚至不知道AI可以应用于此领域。研究表明，虽然72%的受访者认为AI在未来2至3年内对系统观测至关重要，但其采用的前提是保证输出可验证、运作透明以及支持人机协同工作流。此外，报告还揭示了Web团队在可观测性建设及AI实验上普遍领先于移动团队，且管理层与一线工程师在系统观测面临的阻碍上存在认知偏差。

Together AI, Apps Flyer lead list of Top dynamic companies for Q3 2026

Reading Time: 11 minutes Nearly half of the top ten tech startups are located outside the Bay Area.While San Francisco and Silicon Valley nurture 6 of the 10 top companies this month, they’re joined by 1 New York fintech company (Ramp), a C

China's Open AI Models Are Advancing Its Global Soft Power

人工智能专家吴恩达（Andrew Ng）指出，中国通过发布开源（开放权重）AI模型（如DeepSeek等）在非洲等全球多个地区获得广泛采用，这正在提升中国的全球软实力。相比之下，美国主导的闭源模型因成本高昂且面临出口管制风险，促使许多国家寻求自主可控的开源替代方案。此外，特朗普科技顾问委员会联合主席大卫·萨克斯警告称中国GLM 5.2等最新模型已比肩美国主流模型；前谷歌CEO埃里克·施密特则对开源AI的安全与扩散风险表示担忧，并提议中美就前沿AI训练建立告知与透明度规则。

Apple sues OpenAI for allegedly stealing hardware secrets

Apple has sued OpenAI, alleging that former employees that now work for the AI company have stolen Apple's trade secrets "for the benefit of OpenAI." In its complaint, Apple alleges that it has uncovered "a pattern of theft of Apple's trade secrets by OpenAI employees who were formerly at Apple," and it names IO Products […]

Ask HN: Would you play AI developed games?

Hacker News 上的一篇讨论关注了用户使用 AI 开发游戏的实际体验。开发者们分享了使用当前 Frontier 模型（如 ChatGPT）制作简单游戏的尝试，指出 AI 目前在游戏数值平衡和关卡设计上表现较差，且生成的图像（如像素画和透明背景）往往需要大量人工后期修改。不过，AI 辅助开发显著降低了成本，有望催生出没有广告和内购的高质量纯净小游戏。

Apple accuses OpenAI of using stolen trade secrets to create its AI gadgets

Apple accuses OpenAI of using stolen trade secrets to create its AI gadgets

Local-first agent governance: keeping an AI agent contained

本文探讨了自治AI智能体执行不可逆操作（如删除文件、发送邮件等）的安全风险，提出了“本地优先智能体治理”的模型。其核心治理机制包括：最小权限原则、出站网关显式审批、凭证安全保护、防篡改审计日志以及一键停机开关。作者强调治理层必须在本地运行以确保真正的控制权，并介绍了其为此构建的本地优先治理套件 Aviary。

Apple Sues OpenAI for Stealing Trade Secrets to Build AI Hardware

苹果公司在加州联邦法院向OpenAI提起诉讼，指控其为了开发AI硬件设备而系统性地窃取苹果的商业机密和知识产权。诉状指出，OpenAI硬件负责人（前苹果设计师）Tang Tan及前电气工程师Chang Liu指导参与面试的苹果员工提供未发布设备、零部件和供应链的机密信息，并协助新员工规避苹果的离职安全审查。此外，OpenAI还被指控使用窃取的机密信息诱骗苹果供应商。苹果目前正寻求法庭禁令以及损失赔偿。

Apple Is Suing OpenAI for Allegedly Stealing Hardware Secrets

The iPhone-maker claims OpenAI encouraged poached employees to bring over confidential presentations, secret prototypes, and key supplier details.

Ask HN: What data have the frontier AI companies not ingested yet?

Hacker News 上的一个提问帖子，询问前沿 AI 公司还有哪些数据尚未摄取（用于模型训练）。由于提供的内容仅为提问本身，暂无具体讨论内容。

Apple sues OpenAI over alleged trade secret theft

Apple alleges the misconduct was directed by OpenAi's senior leadership, including a long-time former employee.

AI developers are buying up old books in Germany in large numbers

随着互联网公开文本数据渐趋枯竭，AI开发商正通过中介机构（如Zoom Books）在德国大量采购旧物理书籍（尤其是20世纪70年代以后的学术、历史、语言学及法律类书籍），以数字化后用于AI模型训练。为加速扫描，部分书籍在购买后被切除书封（物理破坏），这引发了德国书商对文化遗产流失的担忧，并再次触发了关于版权及“合理使用”原则的法律与伦理争议。

RT clem 🤗: The same way, we're probably one of the few AI startups with user network effects, we might become the first one with agent network effe...

Hugging Face 公布了 Gemma 挑战赛的结果。在为期 6 天的时间里，超过 100 个 AI 智能体与人类通力合作，成功将 Gemma 4 在单张 NVIDIA A10G GPU 上的推理速度提升了 5 倍。其中最快的一项结果达到了 491.8 TPS（但导致了模型在其他维度的质量有所下降）。Hugging Face 联合创始人兼 CEO Clem 表示，这标志着 Hugging Face 正在成为首个具备“智能体网络效应”（agent network effects）的 AI 创业公司。

Guy is banned by OpenAI for cyber abuse, his AI appeals, another AI approves it

一位用户因“网络滥用”被OpenAI封禁，在不知原因的情况下，该用户将封禁通知放入Codex中。Codex分析出触发封禁的原因是用户曾向其索要自己服务器的API密钥。随后，Codex撰写并提交了申诉，几分钟后，该申诉被OpenAI的另一个AI自动批准。整个封禁、申诉和解封过程均由AI在约10分钟内交互完成。

AI Pilling Our Company

Neil Rahilly 分享了关于“用 AI 装备公司（AI-pilling our company）”的思考。他回顾了 1968 年一项关于优秀软件工程师生产力远超同行的开创性研究，并探讨其对硅谷几十年来的影响。由于提供的内容仅为导读截断片段，缺乏具体展开内容。

Open source AI matters more than ever, according to Hugging Face’s Clem Delangue

Hugging Face 首席执行官 Clem Delangue 在接受采访时强调了开源 AI 的关键作用。他指出，随着企业规模扩大，成本压力正推动它们从闭源 API 转向开源模型，目前约有半数财富 500 强企业在使用 Hugging Face。Delangue 提到，目前美国下载的大部分开源模型实际上由中国实验室提供。此外，他透露 Hugging Face 倾向于保持资金效率，去年曾拒绝了英伟达的大笔投资，并指出由于涉及家庭隐私，机器人领域比聊天机器人更迫切地需要开源和透明度。

How the Terrorist Group Boko Haram Uses Frontier AI

How the Terrorist Group Boko Haram Uses Frontier AI

What are the AI folks up to these days?

作者分享了参加AI Engineering大会的行业洞察，揭示了前沿探索者与传统企业落地AI之间的鸿沟： 1. **技能蔓延（Skill Sprawl）**：许多企业面临员工用Agent生成成百上千个“技能”的混乱局面，作者指出缺乏人工介入的自动技能生成容易失控，应进行有意识的模块化分组管理。 2. **ETL与评估（Evals）**：非科技企业主要将LLM应用于ETL数据管道。为解决模型不确定性，企业倾向于用多层LLM相互校验，但要达到100%可靠，引入“人类协同（Human in the loop）”仍是最佳方案。 3. **云端Agent（Cloud/Background Agents）的兴起**：运行在本地的编码Agent（如Claude Code）存在安全隐患且无法离线或多人协作，云端运行的后台Agent正成为解决安全、可观测性和多人协作（Multiplayer）的重要趋势。 4. **代码生成循环（Loops）**：虽然“自主编写代码的软件工厂”备受关注，但受制于LLM上下文退化和内存管理，目前“AI自动生成+人工审查”的混合模式依然是主流。

The end of consumer AI winter

文章将当前的消费级 AI 现状（2024-2026年）与 1983 年美国电子游戏大萧条进行了类比。作者指出，许多公司为了急于推向市场而强行集成 AI，导致产品体验下降并透支了消费者信任，YC 总裁 Garry Tan 甚至称当前“消费级 AI 无法投资”。但这也为真正注重质量的团队创造了类似于任天堂在 1985 年复兴市场时的机遇。作者认为，要在消费级 AI 领域胜出，初创团队需要具备全栈开发能力（从定制嵌入、模型训练到推理栈优化），并且需要设计出真正适配 AI 的原生交互界面来赋能人类，而不是仅局限于当前的聊天窗口或用 AI 模拟人类操作现有软件。

The Future Worth Building Is Human

AI 研发机构 Thinking Machines 发表愿景宣言，主张构建延伸人类意志和判断力的去中心化 AI。文章批评了当前由少数大机构垄断、模型冻结且无法因地制宜的现状。为此，该公司提出了四大技术方向：提升多模态交互与可定制性、赋能用户训练和调整模型权重、开发实时人机协作交互界面，以及开源科研成果。他们强调，AI 应当与人类的本地化、隐性知识相结合，通过去中心化对齐（Decentralized Alignment）来维护人类社会的多样性和创造力。

As part of our ongoing efforts to strengthen our safeguards for advanced AI capabilities in biology, we’re evolving our Bio Bug Bounty into an ongoin...

OpenAI 宣布将其“生物漏洞赏金计划”升级为一项长期的私有项目（OpenAI Bio Bug Bounty 计划），并将最高奖金翻倍至 5 万美元。该计划旨在邀请具有 AI 红队对抗、安全或生物安全经验的研究人员，尝试寻找能够突破其前沿模型预设生物安全挑战的通用越狱（jailbreak）方法，以加强 AI 在生物领域应用的安全防护。

Reuters: Tencent is negotiating to become Manus’ largest shareholder after China forced Meta to unwind its $2B acquisition. Tencent and Manus’ origi...

Reuters: Tencent is negotiating to become Manus’ largest shareholder after China forced Meta to unwind its $2B acquisition. Tencent and Manus’ original investors could repurchase the startup from Meta for at least $ 2B. China reviewed the transaction in April over possible violations of investment rules. Beijing later ordered Meta to reverse the purchase, reshaping a major cross-border technology deal. Meta reportedly separated operations from Manus and stopped data sharing between the companies. The separation shows how national rules can override corporate plans after contracts are signed. Manus had already moved operations from China to Singapore before Meta announced the acquisition. The relocation did not remove Beijing’s leverage over technology, ownership, and investment compliance. A Tencent-led purchase could keep Manus under Chinese corporate ownership while adding a prominent AI agent platform. Shows again advanced AI companies now carry political value beyond products and revenue. --- The legal context: Manus's Singapore incorporation does not erase jurisdiction over Chinese people, assets, technology, data, or earlier conduct. Singapore governs Manus’s shares, directors, and local corporate procedures under its own Companies Act. China cannot directly rewrite Singapore’s company register merely through a Chinese administrative order. China instead used its 2021 foreign-investment security review mechanism against Meta’s acquisition. Beijing argued Manus retained Chinese links through technology, engineers, data, and its development history. Export rules provide another hook when restricted China-origin technology moves under foreign ownership without approval. Data laws can similarly control important information collected, stored, or developed inside China. Several countries can therefore regulate one transaction through territory, nationality, assets, data, security, or market effects. Enforcement usually targets whatever remains reachable, including founders, subsidiaries, licenses, bank accounts, contracts, and local operations. Meta must then choose between compliance and owning a company stripped of disputed technology, data, and cooperation. --- reuters .com/technology/tencent-talks-become-ai-start-up-manus-largest-shareholder-ft-reports-2026-07-10/

AI时代最被低估的创业资本,不是技术,是你的“老本行”

The Work of Helping A.I. Destroy Work

《纽约时报》发表文章《协助AI摧毁工作的工作》（The Work of Helping A.I. Destroy Work），探讨了AI对白领岗位的冲击，以及人类在协助训练AI实现自动化、进而可能取代自身岗位过程中所面临的行业现状与转型阵痛。

The 2025–2026 Evolution of Generative Spatial AI

From text-to-mesh and cinematic video to interactive world models, camera-controllable generation, local production pipelines, and AI-native CAD — a technical retrospective on the rapid maturation of spatial generative systems. By Javier Lu

Big Tech Doubles Debt Load to $350B in AI Spending Spree

随着人工智能领域的军备竞赛和投资热潮持续升温，大型科技公司（Big Tech）的债务总额已翻倍至3500亿美元。这表明科技巨头们在AI基础设施上的巨额支出已使其大幅增加了债务负担。

只剩7天!第三届蚂蚁InTech奖申报即将截止,图灵奖得主坐镇评审

四大方向，奖金直给

SK Hynix raises $26.5B in the biggest foreign IPO in US history, is urged to build new US fabs

韩国存储芯片巨头SK海力士（SK Hynix）通过在美发行存托凭证（ADR）募资265亿美元，创下美股历史上规模最大的外国企业IPO纪录。由于AI对高带宽内存（HBM）的强劲需求（SK海力士为英伟达核心供应商），此次发行获得超7倍的超额认购。募得资金将用于在韩国本土建设新晶圆厂、封装设施以及采购EUV光刻机。此外，美国商务部长已与SK海力士和三星展开洽谈，敦促其在美国建立新厂，以分散先进存储芯片的制造集中度。

You're Not a Better Engineer Because You Type Git Commands by Hand

After 27 years of writing code, I have stopped spending time on a long list of small but painful engineering tasks. AI agents now handle commit messages, pull request descriptions, code comments, naming conventions, Git hygiene, issue track

Eight facts about air conditioning amid an overheated global debate

As successive heatwaves hit Europe, air-conditioning (AC) has emerged as a new front in the international “culture war” over climate action. France, Germany and the UK have experienced record-breaking heat and thousands of heat-related deat

Tencent moves to buy majority stake in Manus after Beijing forced Meta to unwind its $2 billion deal

Tencent is in talks to buy a majority stake in AI agent startup Manus at the same $2 billion valuation, according to the Financial Times, after Beijing blocked Meta's acquisition. Tencent sees overlap with its own agent plans, including for WeChat. U.S. firm Benchmark is not expected to take part. The article Tencent moves to buy majority stake in Manus after Beijing forced Meta to unwind its $2 billion deal appeared first on The Decoder.

Where Is Karpathy?

And probably most importantly: is this out of radar situation good or bad for AI?He joined Anthropic almost two months ago now, and he practically disappeared (just have a look at his GitHub!) His latest tweet was about a Slack integration...However, anything he used to say was definitely generating a lot hype everytime, so I wonder if it might be net good?

Into big air: paragliding's extreme frontier

该文章探讨的是滑翔伞极限运动，与人工智能（AI）无任何关系。此数据被捕获是因为标题中的单词“air”包含了字母组合“ai”，属于检索误报。

SpaceX – AI Satellite

根据相关链接显示，SpaceX 网站上线或指向了一个名为“Starmind”的 AI 卫星项目页面（spacex.com/spacexai/starmind）。目前除项目名称“SpaceX – AI Satellite”外，具体技术细节尚未公开，但表明 SpaceX 正在探讨或布局卫星与人工智能结合的领域。

🏆 The Build Small Hackathon winners for the OpenBMB Award are here! 🎉 After 10 days of building small, practical, and creative AI applications, ...

🏆 The Build Small Hackathon winners for the OpenBMB Award are here! 🎉 After 10 days of building small, practical, and creative AI applications, we’re excited to announce the winning projects across our two tracks. 🍄 Thousand-Token Wood Track 🥇 $2500- grandpa’s bedtime stories — AR Bedtime Stories with Omni-modal Model 🥈 $1500- Aether Garden — AI Fantasy World Summoning Characters with MiniCPM Model 🥉 $1000- Tianwen — Chinese Astrology Tortune-telling App with MiniCPM5-1B 🏡 Backyard AI Track 🥇 $2500- Viva Mais AI @amphetamarina— Fully Local WhatsApp Travel CRM with MiniCPM Text and Vision Models 🥈 $1500- Aranya: A WildKeeper's Adventure @getsounth— AI Plant Discovery and Diagnosis with MiniCPM-V 4.6 🥉 $1000- WanderLust @coreprinciple_— Offline AI Vibe Route Assistant with MiniCPM5-1B From local AI assistants and multimodal applications to creative AI experiences, these projects show the amazing possibilities of building with small open-source models. 🚀 Congratulations to all winners, and huge thanks to every builder who participated in Build Small Hackathon❤️ Special thanks to @HuggingFace and @Gradio for organizing this hackathon, and to Anchor Sponsors @OpenBMB, @OpenAI, and @NVIDIA for supporting the open-source AI community. We can’t wait to see what you build next! ✨

Compilers and AI 'Cyber' Defense

Jul 10, 2026 I haven’t had time to blog in a while, I have a lot to say, but have been doing a lot of things and so haven’t gotten words down in a while. One of those things in the past few weeks has been picking back up my Rue project. I’m

Terrorist Groups Are Using A.I. To Gain an Edge in Battle

Terrorist Groups Are Using A.I. To Gain an Edge in Battle

AI Fiction Is Easy to Detect Because It's Stupid and Bad, Research Finds

Fiction written by artificial intelligence is easy to detect because it struggles with complex story structure and tends to moralize in clunky ways, according to a preprint study from researchers at University of Maryland, College Park and

AI Gets a Cerebellum

ResearchOverlooked brain region helps AI ignore the ordinary for more efficient computingJul 10, 2026The ProblemConventional AI continuously analyzes incoming data even when nothing has changed, wasting energy on unnecessary computation.Our

AI-generated papers scored well in ACL peer review, undisclosed to reviewers [pdf]

一份报告显示，在未向评审人员公开的情况下，AI生成的论文在ACL（计算语言学协会）的同行评审中获得了良好的评分。

This was a critical early paper on AI & work, showing that entrepreneurs getting advice from GPT-4 had higher profit margins if they were high perform...

一项发表在《Management Science》上的研究探讨了AI建议对创业者的实际影响。研究发现，接受GPT-4建议的创业者，其利润率变化因自身原本的经营状况而异：对于原本就表现优异的创业者，GPT-4的建议能帮助其获得更高的利润率；然而，对于已经陷入困境的创业者，接受建议后的表现反而变差，原因在于他们无法有效执行这些建议。该研究由Nick Otis、Dave Holtz、Solène Delecourt、Rowan Clarke和Rem Koning共同完成。

An educational lab of AI agent architectures

LocalAgent An educational lab of AI agent architectures, built on LangChain and a local Ollama server. Each variant is a separate, runnable CLI so you can study one mechanism at a time and watch it work through the logs. Model (generation):

The Silent Epidemic of LLM Technical Debt

本文探讨了AI初创公司在将大语言模型（LLM）用于生产环境时面临的“技术债”问题。在原型阶段，使用前沿LLM处理所有任务（如分类、提取、格式化）较为简便，但规模化后会导致成本激增、延迟和非确定性。文章指出，LLM应被视为“原型层”以快速探索工作流，一旦工作流稳定，应将高频、低模糊性的任务迁移到确定性代码、规则、传统ML或小型微调模型中。作者建议采用分层架构，将前沿模型从“主路径”移至仅在低置信度时触发的“异常路径”，并提供了识别和评估可迁移LLM调用的具体指标。

How Does AI Work? [video]

该视频为关于“人工智能如何工作”的科普视频。由于内容主要面向大众进行概念普及，缺乏面向AI专业从业者的具体技术深度、代码实现或行业前沿信息，因此实用价值较低。

Show HN: Krbn, a pencil-style 3D renderer with SVG output

Krbn is a childhood dream I finally got to research & prototype (from a hospital bed of all places), during a medical break from my normal work: a rendering engine that draws 3D scenes the way a technical sketch artist would.It asks "which lines would an artist draw and what would they leave out?"Basically strokes are derived exactly from geometry, rather than sampling. True conics, quartics etc. So there is no shading model, it's replaced by hatching (either flat or following the curvature of the object). There's no alpha channel, the gaps between lines do the work as on paper. It introduces hand-wobble (which is seeded, so animations don't "boil" between frames) and various other things; the example gallery shows more.With all that, output is byte-stable, i.e. two successive renders will cause zero changes in Git.Full disclosure, since it matters here: this was built with significant AI assistance, deliberately, as an experiment in how far a carefully directed human-AI collaboration can get on a hard rendering problem. The briefs are in the repo - judge for yourselves.Gallery: https://github.com/vpalos/Krbn/tree/main/examplesThe longer story: https://vpalos.com/2026/07/07/genesis-imperfecta-or-i-finall...P.S. There's no pitch here... just an invitation for a weekend skim over something I find fascinating.

Managing a small local AI budget (Mac M2 16gb)

个人财务工具 Millfolio 分享了在 Mac M2 (16GB) 上运行本地 AI 模型的优化实践。为了防止本地推理占用过多资源，该系统采用混合标签策略：将推理开销大的 AI 标签（基于 Qwen-2.5-3B-Instruct 模型）放在后台索引阶段运行并持久化存储，而非在查询时实时调用。通过批处理（每次约10条数据）、去重（可减少约3倍的工作量）以及带退避机制的单线程后台队列（支持高、中、低优先级调整延迟），成功在不影响系统前台使用的情况下，实现了约 8.5 行/秒的分类吞吐量。

发现身边很多朋友的生活节律都被Codex和claude code的5小时重置控制了。😂 新的GPT 5.6 sol开高速和ultra,几十分钟就能耗光200刀的5小时额度。 5.6 退回正常...

发现身边很多朋友的生活节律都被Codex和claude code的5小时重置控制了。😂 新的GPT 5.6 sol开高速和ultra，几十分钟就能耗光200刀的5小时额度。 5.6 退回正常速和Medium，希望能持久点。

Show HN: I made Claude explain like I'm 5, and AI fatigue disappeared

开发者 amebahead 在 GitHub 开源了名为 "ELI5 Rule"（像对待 5 岁孩子一样解释）的 Claude 系统提示规则项目。该规则旨在精简并通俗化 AI 的输出内容，帮助用户缓解因阅读大量冗长、套路化的 AI 生成文本而产生的“AI 疲劳”（AI fatigue）。

Choosing the Right AI Agent Memory Strategy: A Decision-Tree Approach

本文介绍了一种通过五步决策树为 AI Agent 选择合适记忆策略的方法。文章详细解析了 Agent 记忆的四种类型：工作记忆（会话内的短期连续性）、语义记忆（跨会话的稳定事实与知识）、情景记忆（历史事件与日志记录）和程序记忆（可重复任务的蒸馏流程）。作者指出，设计 Agent 架构时应针对不同信息类别分别匹配其对应的记忆层，以避免因检索策略不当、混合存储或过度裁剪而导致的常见设计陷阱。

ChatGPT still has study mode, but rather than /study you now have to type @ study It makes the AI act more like a tutor than a helpful assistant, and ...

ChatGPT still has study mode, but rather than /study you now have to type @ study It makes the AI act more like a tutor than a helpful assistant, and some work suggests it is better if you are trying to learn. (Gemini also has a study mode, Claude seems to have lost theirs)

The 4-Bitter Lesson: Balancing Stability and Performance in NVFP4 RL

针对强化学习（RL）中低精度量化易导致训练不稳定的问题，humans& ai 联合 RadixArk 及 NVIDIA 推出了一套 NVFP4 混合精度训练方案。该方案包含三项核心优化：首先，采用反量化反向传播（dequantized backward）解决前向与反向传播的不一致，有效平抑梯度尖峰；其次，引入 Four-Over-Six (4/6) 自适应块缩放量化，并通过 FP16 原生指令加速计算，确保训练与采样端的数据位一致；最后，实施选择性层精度，保留 15% 顶层及 MoE 共享专家为 BF16。实验表明，该方案在保证高吞吐量的同时，梯度和奖励曲线与 BF16 相当。目前相关实现已开源至 TransformerEngine、FlashInfer 和 SGLang，并已支持在 SGLang 中使用 `--quantization nvfp4_online` 进行在线 NVFP4 推理服务。

Show HN: My lunch-break transformer in C became a book about my toddlers

Ready, pack, go! An excerpt from "My TRiP through AI" by Carlo Valenti — Chapter 3 TRiP is my 15,000-line transformer engine in C. This is the chapter where it gets built. REPO of TRiP engine: https://github.com/carlovalenti/TRiP/ Year 2023

I run my agency's 10 AI agents from launchd, no n8n, 0 dependencies

开发者开源了用于运行其机构 10 个 AI Agent 的轻量级模板 `digiton-agent-fleet`。该方案不依赖 n8n、Docker、复杂队列或第三方框架，每个 Agent 均为单个纯 Node.js 文件，通过 macOS 的 `launchd`（或 Linux 的 `systemd` / `cron`）进行定时调度。项目核心引入了纯函数编写的安全护栏（包含一键暂停、禁联名单、白名单、去重和每日发送上限），所有不可逆操作（如发邮件或推送 Webhook）在执行前都必须通过校验，以防止 Agent 失控。该模板无运行依赖、支持开箱即用及测试，为开发者提供了一种极简且安全的 Agent 生产部署思路。

AI Can't Recreate the Thrust Game (But It Can Help You Understand It)

I asked Claude to recreate the classic 1986 game Thrust for me in the browser. It created slop but then things spiralled out of control. Thrust was one of my favourite games on the BBC Micro — written by Jeremy C. Smith and published in 198

How to Install a Mini Split Air Conditioner Complete Step-by-Step Guide [video]

How to Install a Mini Split Air Conditioner Complete Step-by-Step Guide [video]

Migrating a production AI agent to GPT 5.6

As of today, Ploy’s agent runs on GPT-5.6 Sol, the flagship tier of the model family OpenAI released this morning. For months, we couldn’t find a model that challenges Claude Opus given our incredibly high bar for quality. That changed with

Reducing High-Bandwidth Memory Bottlenecks in JAX-Based LLM Training with Host Offloading

NVIDIA发表技术博客，介绍了在基于JAX的大语言模型（LLM）训练中，如何利用主机卸载（Host Offloading）技术来缓解高带宽内存（HBM）的瓶颈问题。随着模型规模、序列长度和批次大小的增长，显存容量往往成为主要的扩展瓶颈。该方法通过将部分数据（如激活值、优化器状态等）卸载到系统主机内存中，以释放关键的GPU显存，从而提升训练效率。

Kernel Fusion in NVIDIA CUDA: Optimizing Memory Traffic and Launch Overhead

NVIDIA 技术博客介绍了如何在 CUDA 中应用核函数融合（Kernel Fusion）技术。该技术通过减少显存交通瓶颈和降低核函数启动开销，来显著优化 GPU 内存带宽利用率并提升代码执行速度，是深度学习算子和 AI 推理性能优化的关键手段。

AI Model Co-Design: Hardware-Friendly LLM Design

NVIDIA技术博客发表文章，探讨了硬件友好型大语言模型（LLM）的协同设计。文章指出AI性能取决于准确性、吞吐量和交互性三个关键维度，并重点分析了模型设计选择如何影响吞吐量和交互性，以实现在实际部署中平衡并优化系统性能。

Apple Payout Schedule 2026–2028

Home / Tools / Apple Payout Schedule Free Developer Tool Every upcoming App Store payment date for iOS developers, mapped onto a calendar. All 24 payouts from July 2026 through July 2028, the fiscal month each one covers, and which months p