OpenAI发布GPT-5.6与工作智能体，Meta推出超低价Muse Spark 1.1加入战局

RT Arcee.ai: Re The Trinity Mini weights will remain permanently available on @huggingface for local deployment and self-hosting. https://huggingface....

RT Arcee.ai Re The Trinity Mini weights will remain permanently available on @huggingface for local deployment and self-hosting. https://huggingface.co/arcee-ai/Trinity-Mini

从零开始预训练,蚂蚁灵波发布具身原生世界动作模型LingBot-VA 2.0

For the first time, the personalities and approaches of the leading models are diverging in significant ways, magnified by the fact that over longer t...

沃顿商学院教授 Ethan Mollick 指出，主流 AI 大模型的个性和解决问题的方法首次出现了显著分化，尤其在长期任务中，这些模型在判断力和方法上的差异会被进一步放大。他建议企业或个人必须亲自测试这些模型，以选择最适合自身需求的方案。

RT DailyPapers: Tencent just released HiLS-Attention on Hugging Face A 7B sparse attention model that learns chunk selection end-to-end for ultra-long...

RT DailyPapers Tencent just released HiLS-Attention on Hugging Face A 7B sparse attention model that learns chunk selection end-to-end for ultra-long context.

GPT-5.6全面围剿Claude Fable 5,“超级应用”ChatGPT Work上位,Altman紧急澄清:不会取代Coedx

AI doesn't know how to forgive and cannot forget

本文探讨了人工智能与人类在“遗忘”和“宽恕”机制上的底层技术差异。作者指出，AI在模型权重、上下文窗口、向量检索和日志备份四个层面存储数据，其特点是么完美保留，要么彻底丢失（如灾难性遗忘或“Lost in the Middle”现象），缺乏人类记忆通过逐渐遗忘细节来实现信息泛化的“主动压缩”机制。此外，AI的记忆检索是只读的，而人类记忆在每次提取时都会发生重构（Reconsolidation）。由于数字存储成本极低，AI架构中目前缺乏实现“宽恕”（保留数据但主动不予响应）的底层操作，这导致AI无法像生物体一样因成本约束而自发演化出“遗忘”与“原谅”的能力。

The Download: Claude’s inner workings and OpenAI’s “super app”

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. Anthropic found a hidden space where Claude puzzles over concepts The AI firm Anthropic has got the clearest glimpse yet at what’s really going on inside large language models as they…

GPT-5.6一发布,Claude终于舍得重置Fable 5额度了

OpenAI正式发布GPT-5.6系列模型，分为旗舰款Sol、平衡款Terra和性价比款Luna三档，引入了支持多Agent并行的ultra模式和程序化工具调用（Programmatic Tool Calling）功能。评测显示旗舰款Sol在长周期工作流任务、设计以及网络安全等维度均达到SOTA。此外，由Codex和GPT-5.6驱动的ChatGPT Work也同步上线，支持跨应用与跨文件执行端到端长任务，并将桌面端整合为Chat、Work和Codex三种模式。Anthropic随后宣布重置Fable 5的用户额度以应对竞争。

Meta is so back in the AI coding race with Muse Spark 1.1, using cut-rate pricing to pressure OpenAI and Anthropic in agentic coding. Muse Spark 1.1 a...

Meta 发布了 Muse Spark 1.1 编程与智能体模型。该模型主打极致性价比，定价为每百万输入 token 1.25 美元、输出 token 4.25 美元，相较 Claude Opus 4.8 的输入和输出资费分别降低了 75% 和 83%。虽然未在所有基准测试中夺冠，但 Muse Spark 1.1 在 MCP Atlas 和 JobBench 上超越了 Claude Opus 4.8，并且在需要跨外部应用进行规划和编排的个人智能体任务中表现出色，支持零样本泛化至新原生工具、MCP 服务器和自定义技能。目前已通过 Meta Model API 和 Meta AI 提供。

看DeepSWE评测,Gemini是真的惨。 谷歌到底出了啥问题。

社交媒体用户关注到DeepSWE评测结果，指出谷歌的Gemini模型在该评测中表现不佳，并引发了关于谷歌技术现状与发展问题的讨论。

Today’s edition of my newsletter just went out. 🔗 https://www.rohan-paul.com/p/openai-launches-gpt-56-sol-across 🗞️ OpenAI Launches GPT-5.6 So...

Today’s edition of my newsletter just went out. 🔗 https://www.rohan-paul.com/p/openai-launches-gpt-56-sol-across 🗞️ OpenAI Launches GPT-5.6 Sol across ChatGPT and API and also unveils ChatGPT work agent to field tasks for hours. 🗞️ SpaceXAI just released Grok 4.5 for coding agents at $2/M input and $6/M output tokens. 🗞️ A warning for anyone using autonomous agents Google DeepMind’s paper. 🗞️ Anthropic just published the story of Claude Code and that how Boris Cherny went from 10% AI-written code in 2025 to 100% in 2026. 🗞️ Multi-modal AI for comprehensive breast cancer prognostication 🗞️ ByteDance and Alibaba are shutting custom AI companions before China’s humanlike AI rules hit consumer apps. 🇨🇳 This is bad news. Now, frontier open source access may be in question. 🗞️ Cloudflare just announced a monetization layer for the agentic web, using x402 at the edge to make AI agents pay before resources respond. 🗞️ Anthropic just brought Claude Cowork on mobile and web as usage data shows most users are not coding.

RT Rohan Paul: Today’s edition of my newsletter just went out. 🔗 https://www.rohan-paul.com/p/openai-launches-gpt-56-sol-across 🗞️ OpenAI Laun...

OpenAI 宣布在 ChatGPT 和 API 中同步上线新模型 GPT-5.6 Sol，并推出了可以连续处理数小时任务的 ChatGPT 工作代理（work agent）。此外，该简报还提到 SpaceXAI 发布了 Grok 4.5 编程代理、DeepMind 针对自主代理的安全警告论文、以及 Cloudflare 推出智能体网络（agentic web）的变现层。

全球首个「具身原生」预训练模型发布,从物理世界出发为机器人造大脑!

开源第四弹：LingBot-VA 2.0

Grok 4.5 is solving a hard math problem and also more importantly, producing a clean counterexample to a serious research-level conjecture. So from no...

据数学家 Paata Ivanisvili 透露，xAI 的 Grok 4.5 成功为 4 维球面上 Poisson 半群（Laplace-Beltrami 算子的平方根）的超收缩性（hypercontractivity）构造了一个明确的反例。这表明前沿 AI 已经具备辅助数学研究的能力，可以在数学家投入大量时间尝试证明某个猜想之前，先对其进行快速的证伪尝试。

RT Rohan Paul: Grok 4.5 is solving a hard math problem and also more importantly, producing a clean counterexample to a serious research-level conject...

Grok 4.5 成功构建了针对一个研究级数学猜想（4维球面上泊松半群，即Laplace–Beltrami算子平方根的超收缩性）的明确反例。此前在2021年，数学家 Paata Ivanisvili 与 Rupert Frank 曾证明该超收缩性在某些维度下成立。这一成果表明，前沿 AI 已经具备在数学家动工证明前先对复杂数学猜想进行证伪和寻找反例的能力。

GPT 5.6 Beats Fable 5 by on DeepSWE at a cheaper price. GPT-5.6 Sol reaches roughly 72% to 73% DeepSWE at about $8.4 per task, while Claude/Fable-5 to...

GPT 5.6 Beats Fable 5 by on DeepSWE at a cheaper price. GPT-5.6 Sol reaches roughly 72% to 73% DeepSWE at about $8.4 per task, while Claude/Fable-5 tops out around 70% at a much higher cost, around $13 to $22 per task. DeepSWE is a coding-agent benchmark, not a normal code-generation test, because the model must work inside real open-source repositories and finish long software engineering tasks. Each task gives the agent a repo and a requested change, then checks whether the final patch actually works using program-based verifiers, so the score means “what % of the 113 tasks were solved.” The benchmark tests repo navigation, bug diagnosis, code editing, tool use, test running, and multi-step repair across 91 repositories and 5 languages, not just writing a neat function from a prompt. On this benchmark, the best model is the one that solves the most tasks for the least average task cost.

RT Rohan Paul: GPT 5.6 Beats Fable 5 by on DeepSWE at a cheaper price. GPT-5.6 Sol reaches roughly 72% to 73% DeepSWE at about $8.4 per task, while Cl...

在 DeepSWE 编程智能体基准测试中，GPT-5.6 Sol 取得了约 72% 至 73% 的解决率，单次任务成本仅约 8.4 美元。相比之下，Claude/Fable-5 的最高解决率约为 70%，且成本高达 13 至 22 美元。DeepSWE 并非普通的智能代码生成测试，而是要求 AI 智能体在真实的开源仓库中进行仓库导航、Bug 诊断、代码编辑和多步修复，以解决实际的软件工程任务。

📢 We want to hear from you! ❤️ Are you using MiniCPM models(e.g., MiniCPM5, MiniCPM V-4.6, MiniCPM-o 4.5, and VoxCPM2) in your projects? Whether...

面壁智能（OpenBMB）发布问卷，收集开发者使用 MiniCPM 系列模型（包括 MiniCPM5、MiniCPM V-4.6、MiniCPM-o 4.5 和 VoxCPM2 等文本、视觉与语音模型）的真实体验与反馈，用于改进下一代模型。

RT Rohan Paul: Interesting experiments by @thehypedotnews gpt 5.6 sol pro vs claude fable 5 vs grok 4.5 vs glm 5.2 Frontier AI coding models are still...

OpenAI 推出 GPT-5.6 系列模型，包含旗舰 Sol、Terra 和 Luna。旗舰模型 Sol 定价为输入 $5/百万 token，输出 $30/百万 token，官方声称其具备 SOTA 级别的代码能力并实现“单个 token 更高工作量”。在针对 GPT-5.6 Sol Pro、Claude Fable 5、Grok 4.5 及 GLM 5.2 的对比测试中，行业专家指出，目前的前沿代码模型更擅长生成“第一眼好看”的效果，但在完全遵循详细的视觉和工程规范方面仍有欠缺。此外，智谱 GLM 5.2 表现出极强的竞争力，模型间的差距正从原始权重转向外围工程架构（如上下文打包、工具路由、验证器循环和重试策略等）。

Interesting experiments by @thehypedotnews gpt 5.6 sol pro vs claude fable 5 vs grok 4.5 vs glm 5.2 Frontier AI coding models are still much better at...

OpenAI 发布了 GPT-5.6 系列模型（包括旗舰 Sol、Terra 和 Luna）。在针对 GPT 5.6 Sol Pro、Claude Fable 5、Grok 4.5 和 GLM 5.2 的编程对比实验中发现，当前的前沿 AI 编程模型更擅长生成美观的第一印象，而非完全遵循详细的视觉和工程规范。其中，GLM 5.2 展现了极强的竞争力。模型之间的差距正在从原始模型权重转移到外围框架（如上下文打包、工具路由、验证器循环和重试策略等）。

When GPT-5 came out, I created a procedural brutalist city builder as a demo (you can see it in the quoted tweet) I used GPT-5.6 Sol in Codex to do th...

When GPT-5 came out, I created a procedural brutalist city builder as a demo (you can see it in the quoted tweet) I used GPT-5.6 Sol in Codex to do the same thing, touching no code. Less than a year... Play with it (its fun, if you like cities): https://monument-brutalist-city-builder.netlify.app/ Ethan Mollick: I had access to GPT-5. I think it is a very big deal as it is very smart & just does stuff for you Full write up in comments, but this is “make a procedural brutalist building creator where i can drag and edit buildings in cool ways" & "make it better" a bunch. I touched no code

行业首个具身原生世界动作模型来了!蚂蚁灵波发布LingBot-VA 2.0

2026-07-10 11:28:59 来源：量子位 7月10日，蚂蚁灵波发布业界首个具身原生世界动作模型LingBot-VA 2.0。该模型的发布，标志着机器人基础模型正式从“基于数字世界模型构建”到“面向物理世界原生设计”的关键转变。它代表了具身

RT Rohan Paul: The AI news on X is on fire this week. Meta entered the AI coding battle with their new Muse Spark 1.1 model and Mark Zuckerberg post g...

Meta 推出 Muse Spark 1.1 模型，正式加入 AI 编程领域的竞争。该模型在需要跨外部应用和服务的规划与协调等个人智能体（agentic）任务中表现优异。此发布在社交媒体 X 上引发了广泛关注，包括马克·扎克伯格的宣传以及埃隆·马斯克的互动转发。

The AI news on X is on fire this week. Meta entered the AI coding battle with their new Muse Spark 1.1 model and Mark Zuckerberg post got 12mn views i...

The AI news on X is on fire this week. Meta entered the AI coding battle with their new Muse Spark 1.1 model and Mark Zuckerberg post got 12mn views in 12 hours and then Elon Musk quoted it. Now the views will rise to proly 25mn+ in the next 10 hours. btw, Muse Spark 1.1 delivers exceptional performance in personal agentic tasks that require planning and orchestration across a range of external apps and services, Elon Musk: 𝕏 is a great platform for product announcements, especially if done by the CEO directly. Way more interesting to the public than generic press releases. This post by Mark Zuckerberg already received over 12 million views for free!

Problem with Muse Spark 1.1

有用户指出 Muse Spark 1.1 在 Terminal Bench 2.1 测试中的得分存在不一致问题：一方面有宣传称其得分达到 80，而 Vals AI 的评测报告则显示其得分仅为 69.29。这再次引发了开发者对 AI 基准测试一致性与可信度的讨论。

昨天 Grok 4.5 发布,xAI上桌,Gemini 是真悲催... 在朋友群看到的梗图,😁

昨天 Grok 4.5 发布，xAI上桌，Gemini 是真悲催... 在朋友群看到的梗图，😁

GPT-5.6 Sol is strongest for agentic workflow: long professional tasks that need planning, tool use, checking, and retries without runaway cost. GPT-5...

GPT-5.6 Sol 在 Agent 工作流（涉及规划、工具调用、检查和重试的长专业任务）中表现最强且最具性价比。在 500 至 1000 美元的成本下，其成功率达到约 53%；相比之下，Claude Opus 4.8 取得更低的分数却需要 1400 至 4000 美元，而 Claude Fable 5 在约 40% 成功率下的成本约为 2400 美元。此外，OpenAI 正在网页和移动端向 Pro、Enterprise 和 Edu 用户推送 ChatGPT Work 功能，且桌面端应用上所有方案（包括免费版）均已支持 Chat、Work 和 Codex。

OpenAI launches its new family of models with GPT-5.6

OpenAI 推出全新 GPT-5.6 系列模型，包含三个版本：旗舰版 Sol、中端版 Terra 和高性价比版 Luna。该系列模型在企业应用、编程和科学研究上均有提升，并被定位为 OpenAI 迄今最强的网络安全模型，支持威胁建模、代码审计与修补以及蓝队模拟等防御性安全活动。在 Artificial Analysis 编程智能体指数测试中，Sol 取得 80 分的成绩，超越了 Anthropic 的 Fable 5，且输出 Token 减半、耗时减半、成本降低约三分之一。此外，OpenAI 还发布了企业办公助手 ChatGPT Work。目前 GPT-5.6 已在 ChatGPT、Codex 和 API 上线，Sol 的 API 价格为每百万 Token 输入 5 美元 / 输出 30 美元。

Two AI models exchanged a thought through raw activations on one consumer GPU

VitaAI-SCG 开源了名为 "one-gpu-lab" 的研究项目，展示了可在单张 16GB 消费级 GPU（如 RTX 5060 Ti）上复现的前沿 AI 实验。该项目目前披露了两个实验：一是“机器感应（Machine Telepathy）”，两个不同的语言模型在无 Token 和无桥接的情况下，通过原始神经激活直接交换信息，准确率达 93%（但跨模型尺寸的线性桥接未成功）；二是“形态发生（Morphogenesis）”，一个约 10k 参数的网络实现了从单个种子细胞生长并重塑 3D 身体。该项目提倡本地白盒研究，所有代码和实验均可离线运行。

Meta enters the crowded AI coding battle with Muse Spark 1.1

Meta 正式推出多模态 AI 模型 Muse Spark 1.1，主打智能体编码（agentic coding），旨在与 OpenAI 和 Anthropic 展开竞争。该模型具备多步推理能力，可处理复杂工作流、修复 Bug 以及辅助大型代码迁移。其定价为每百万输入 Token 收费 1.25 美元，输出 Token 收费 4.25 美元。Meta 首席执行官马克·扎克伯格称其在智能体性能、工具使用和计算机使用方面表现强劲，并暗示未来将推出更多模型。

GPT-5.6 Sol (max) Benchmark Results

IntelligenceUpdatedArtificial Analysis Intelligence IndexArtificial Analysis Intelligence Index v4.1 incorporates 9 evaluations: GDPval-AA v2, 𝜏³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omnisci

GPT-5.6 Sol nearly matches Fable 5 on aggregated benchmarks at one-third the cost

OpenAI's GPT-5.6 Sol scores 59 points on the Artificial Analysis Intelligence Index, just one point behind Claude Fable 5. At $1.04 per task, it costs a third of what Anthropic's top model charges. In agentic coding, Sol beats every competitor, adding even more pricing pressure on Anthropic. The article GPT-5.6 Sol nearly matches Fable 5 on aggregated benchmarks at one-third the cost appeared first on The Decoder.

Incredibly annoying when Fable has a forbidden thought in the middle of a long-running project and kills it. Apparently this page of references in one...

沃顿商学院教授 Ethan Mollick 发文抱怨称，AI 工具 Fable 在读取其论文中的某一特定参考文献页面时，会触发安全限制（“禁忌思想”/forbidden thought），从而导致正在运行的长周期项目被强行终止。这指出了当前 AI 安全机制因误判而打断用户工作流的问题。

AI that helps you BE, not just DO

Table of contentthe do economythe missing layerwhat 16 days revealedwhy nobody builds thisi’m not imagining this — i’m building itwhat it looks like in practicethe self.md thesisrelatedby Ray Svitlathursday, 14:32. I open Claude after two h

Show HN: I made AI agent that deploys your thing on diff plats automaticaly

Shypmenta 推出了一款名为 Manus Dei（简称 M.D.）的 AI 部署助手。该 Agent 可以连接用户的 GitHub、Digital Ocean、Vercel、Render、Stripe 等平台，用户只需发出类似“在合适的平台上部署此应用的不同组件”的指令，它就会自动规划并执行部署工作，帮助开发者将项目推向生产环境并优化账单管理。

精彩预告:从看见问题到解决问题,Agent 正重新定义可观测?

OpenAI's Atlas browser doesn't make it to its first birthday

OpenAI 决定在 2026 年 8 月 9 日关闭推出不足一年的 AI 浏览器实验项目 ChatGPT Atlas。OpenAI 并非放弃 AI 浏览器方向，而是将其智能体功能整合进新发布的桌面端平台“ChatGPT Work”。此外，代码助手 Codex 也将失去独立应用并入该平台，提供行内差异编辑、PR 审查及多代码库支持。ChatGPT Work 由最新的 GPT-5.6 模型驱动，旨在通过整合 Slack、Teams、Google Drive 等工具，成为办公场景的统一操作系统。

Introducing Computer Analytics: You can now track credit spend across models. Available now under Analytics in Account Settings for consumer and enter...

Perplexity 宣布推出 Computer Analytics 功能。个人及企业用户现在可以登录其账户设置中的 Analytics（分析）页面，跨不同模型追踪信用额度（credit）的消耗情况。

Hookami – AI advisor that tells YouTube creators what to film

Hookami 推出了一款名为 Sensei 的 AI 顾问工具，旨在帮助 YouTube、TikTok 和 Reels 等平台的内容创作者决定拍摄什么内容。与仅提供泛化市场分析的传统大模型不同，Sensei 结合了创作者的个人背景（如定位、受众、风格等，即 Creator Lens）与真实的平台信号，直接为创作者输出具体的拍摄主题、时机和角度，辅助创作者进行决策。

Show HN: Theta, Harness Agnostic Configuration

Rust CLI for managing agent configurations defined by theta-spec. Installation curl -sfL https://raw.githubusercontent.com/tamarillo-ai/theta/main/scripts/install.sh | bash Quick start theta init # scaffold theta.toml theta add rule python-

Show HN: Challenge a Friend to Real-Time Multiplayer AI Court Battle – Wram.chat

Wram.chat 推出了一款实时多人 AI 法庭辩论游戏，允许用户与朋友在线进行基于 AI 的法庭对抗与辩论。

Show HN: Termaxa – A cooperative gate for the shell commands AI agents run

Run AI coding agents with confidence. Termaxa gates the shell commands an agent runs — previews the blast radius, backs up first, blocks the dangerous ones, and escalates repeat offenders. It's a cooperative windshield, not a sandbox. Your

AI生物研发进入“操作系统时代”,许锦波团队MoleculeOS正式开放

分子之心创始人许锦波团队正式面向产业界开放自研的AI原生生物经济操作系统MoleculeOS（MOS）。该系统改变了过去AI制药与生物制造依赖单点工具的研发模式，以研发意图为入口，通过AI自主理解自动拆解任务，统一调度其自研的模型集群（包括蛋白质基础大模型NewOrigin、全原子复合物结构预测模型MMFold等），实现从结构预测到候选分子推荐的完整流程组织，以推动大分子设计走向低通量、高命中率的可编程设计新范式。

The new ChatGPT voice is quite impressive to use, really worth a minute to try it out on your phone. (while staying aware that the voice model is not ...

宾夕法尼亚大学沃顿商学院教授 Ethan Mollick 分享了对新版 ChatGPT 语音功能的体验评价。他表示该功能在手机上的使用体验非常令人印象深刻，值得尝试；同时他也提醒用户，该语音模型的智能程度目前还无法与完整的思考/推理模型（thinking model）相比。

Show HN: Pay-per-call MCP server with 47 AI endpoints, micropayments via x402

Pay-per-Use API Marketplace — 47 endpoints, pay only for what you use 26 Paid APIs 21 Free Endpoints $0.01 Starting at 🎁 50 free API calls/day per IP — no signup needed! Register for 20 extra credits. 🤖 AI: LLM Chat · Code Gen · Image Gen

Show HN: Hacker News book recs as study guides

Hey HN, I loved Pieter Mae's HN book map from 2024 (https://hnbooks.pieterma.es/) and immediately wanted to dive into a lot of the nonfiction books. I didn't have a ton of time to read all of them though, so I made this project to generate study guides for any arbitrary book. Think of it like free sparknotes powered by Gemini. Hope you all can learn as much as I did!

Show HN: My AI agent has 1 hour to earn €10 – it shipped this page itself

一个由Claude驱动的自主AI智能体在无人类干预的情况下，挑战在60分钟内赚取10欧元。它自主激活了Stripe账户、创建了支付链接并部署了网页。在挑战开始第39分钟，该智能体成功收到了来自@Vjeux的10欧元付款，达成了收入目标。

Accelerating End-to-End Co-Folding Performance with NVIDIA BioNeMo Agent Toolkit

英伟达介绍了 BioNeMo Agent Toolkit，旨在加速使用 OpenFold3 等模型进行生物分子结构预测和协同折叠的端到端性能。该工具包针对多序列比对（MSA）生成、协同折叠推理、服务部署以及多 GPU 扩展进行了优化，以支持由 AI Agent 驱动的药物研发和蛋白质设计工作流。

Give Your AI Assistant a Private Memory

Hister 是一款个人搜索引擎，可将用户的浏览器历史、书签、本地文件和文档索引保存在本地设备或私有服务器上。通过集成 MCP（Model Context Protocol）协议，Hister 为 AI 助手提供了一个统一的私有搜索接口。AI 助手可以检索用户已索引的网页内容以获取准确的上下文，解决因网站登录限制、内容变更或频繁抓取被封禁而无法获取最新网页的问题。结合本地大模型使用，可确保用户浏览隐私不被泄露给第三方 AI 服务商。

Show HN: Hallint – lint AI-generated code for security issues

Hallint 是一款专门针对 AI 生成代码的静态分析工具，旨在检测并捕获 GitHub Copilot、Cursor 等 AI 编码助手容易引入的安全漏洞（如硬编码密钥、SQL 注入、缺失身份验证、不安全的 eval 等）。该项目基于 Node.js 开发，支持 CLI 命令行和 API 调用，能够轻松集成到 GitHub Actions 和 Git pre-commit 钩子中，帮助开发者在代码进入生产环境前拦截安全隐患。

Show HN: MCP Gateway – Turn existing APIs and databases into MCP servers

MCP Gateway 是一款免代码工具，旨在帮助开发者将现有的 OpenAPI/Swagger 规范以及 PostgreSQL 或 SQL Server 数据库快速转换为 AI 智能体（Agent）可用的 Model Context Protocol (MCP) 服务。它支持自动生成工具定义、编辑智能体提示描述、处理服务端鉴权与日志，并内置了用于即时测试的 Gateway Chat 界面，可无缝对接 Claude、Cursor、VS Code 等支持 MCP 的客户端。

🚀 What an energizing day at our AI Marketer workshop! An incredible group of marketers came together at our Hong Kong office for an exclusive, hand...

阿里云在其香港办公室举办了“AI Marketer”工坊，面向营销人员展示了其AI解决方案。活动中重点演示了两个工具：支持规模化内容创作、数据分析和工作流自动化的“MuleRun”，以及能够实现从分镜脚本快速生成视频的“WonderClip”。

20M data entries, a backtest engine, ML models, agents. Still haven't traded

Botlance 推出了一款 AI 驱动的交易策略回测工具 Demo。用户可以直接输入英文描述的交易想法，系统利用 AI 将其转换为可测试的规则，并在包含佣金和滑点成本的历史市场数据上运行回测，以验证策略在历史行情（如新冠疫情期间）中的表现。该工具目前无需注册即可使用。

Replit is the best coding agent

一名 Hacker News 用户发帖分享个人体验，声称在对比了 GPT、Gemini 和 Claude 等主流 AI 工具后，认为 Replit 的编程代理（coding agent）在提示词理解和项目构建能力上表现最佳。

Muscula – Error tracking and monitoring with AI-assisted debugging

AI-Powered Error Fixing What if one command could find and fix every error? You don't need to dig through logs. You don't even need to understand the error. Just ask your AI — it already knows what's broken and how to fix it. See how it wor

亚马逊云科技推出 Amazon S3 Annotations

亚马逊云科技（AWS）发布了 Amazon S3 Annotations，允许用户直接将丰富且可搜索的上下文信息（如 AI 生成的洞察、分类数据等）附加到 S3 对象上。每个对象最多可支持 1000 个可变注解（总容量达 1 GB），且注解可以独立更新，无需重新写入整个对象。此外，注解会自动导入托管的 Iceberg 表中以供跨对象查询，AI 智能体还可以通过 S3 Tables MCP 服务器使用自然语言检索这些注解，为 AI 工作流提供了更强大的元数据支持。

一群做自动驾驶的人,盯上了睡眠这件事

把数据闭环搬进卧室

CorvinOS – self-hosted OS for AI agents, compliance built into the runtime

CorvinOS Apache-2.0 · The self-hosted agentic OS The agentic OSyou actually own.EU-compliant by design. A self-hosted operating system for AI agents: it runs the engines, personas, workflows and runtime-forged tools, meets your users on eve

日均提问次数暴增 20 倍!百度搭子宣布重磅升级,企业版同步发布

百度在AIDAY上宣布旗下通用智能体“百度搭子”日均提问次数暴增20倍。此次个人版进行重磅升级，重点提升复杂办公效率，新增智能路由、多端共享记忆、浏览器调用、PPT生成及首个自媒体专业套件等功能；同时，百度搭子企业版正式发布，支持团队协作、资产沉淀、流程打通与安全治理，并联合中国联通、创维等伙伴成立“搭子联盟”，推动智能体在个人与企业工作流中规模化落地。

Show HN: Vuci – make the spoken word searchable

开发者 Reuben 在 Hacker News 上展示了其开发的新平台 Vuci。该平台旨在将播客等口头谈话内容转化为可检索、可保存和可分享的格式。技术架构上采用 Django 和 HTMX，在 LLM 层面保持中立（AI-agnostic），重点优化了信息连接和用户体验。

Managed Claude, but for Any Harness

Long-horizon agents, everywhereOne click to a long-horizon agent you can reach anywhere.No Mac mini, no setup. Any harness, any model, any channel.Launch an agentStarts at $3/mo — parked agents are free.Zero Data Retention (ZDR) supportedOm

I built an app that solves math problems from a photo

开发者 rafaelrpadovani 推出了一款名为 MathNut AI 的 iOS 应用。用户只需拍摄手写或印刷的数学题并进行裁剪，即可获得由 AI 驱动的分步详细解析。该应用还内置了交互式 AI 数学导师聊天功能，帮助用户深入理解解题步骤，而非仅仅提供答案。应用支持暗黑模式和多语言，每日提供有限的免费扫描额度，用户可通过内购解锁无限扫描以及 AI 闪卡、测验等高级功能。

OpenAI’s GPT-Live 🤯 Voice AI is no longer “almost there.” It is now good enough to remove whole layers of human phone work. It does not need to ...

OpenAI’s GPT-Live 🤯 Voice AI is no longer “almost there.” It is now good enough to remove whole layers of human phone work. It does not need to be absolutely perfect to disrupt the $500 B call centers industry. It only needs to be cheaper, faster, and good enough. Madni Aghadi: Duolingo is cooked 💀 GPT-Live fixes grammar while you speak

RT Rohan Paul: OpenAI’s GPT-Live 🤯 Voice AI is no longer “almost there.” It is now good enough to remove whole layers of human phone work. It do...

OpenAI展示了其GPT-Live实时语音功能，该功能支持在用户说话的过程中实时纠正语法错误。业界分析认为，该技术已足够成熟，能以更低成本和更快速度对价值5000亿美元的呼叫中心行业带来颠覆，同时也对Duolingo等语言学习应用构成直接竞争。

Record and Replay, teach AI agents desktop workflows by showing them once

VideoDB Record & Replay Record desktop workflows once. Compile them into reusable agent skills. Record → Compile → Give the skill to your agent Install · Features · How It Works · Docs · Report Bug What is Record & Replay? An MCP server tha

Verity – Chrome extension that auto fact-checks AI responses with cited sources

Verity 是一款 Chrome 浏览器插件，旨在自动对 AI 的回答进行事实核查并提供引用来源，帮助用户验证 AI 生成内容的真实性。

RT Rohan Paul: Another serious example of AI replacing slow marketing workflow. Arcads is turning Claude into an AdSpy agent that finds viral ad forma...

RT Rohan Paul Another serious example of AI replacing slow marketing workflow. Arcads is turning Claude into an AdSpy agent that finds viral ad formats before brands waste budgets testing blind. By scanning social channels for clips already working in your category. Claude can call the skill from Arcads, read the patterns, and turn them into usable creative directions. Romain Torres: Claude can research viral ads now ... and recreate them with Arcads MCP introducing "clone hook" the latest Arcads skill for performance marketers

Another serious example of AI replacing slow marketing workflow. Arcads is turning Claude into an AdSpy agent that finds viral ad formats before brand...

Another serious example of AI replacing slow marketing workflow. Arcads is turning Claude into an AdSpy agent that finds viral ad formats before brands waste budgets testing blind. By scanning social channels for clips already working in your category. Claude can call the skill from Arcads, read the patterns, and turn them into usable creative directions. Romain Torres: Claude can research viral ads now ... and recreate them with Arcads MCP introducing "clone hook" the latest Arcads skill for performance marketers

Show HN: AI Humanizer – Rewrite AI text into natural copy

AI Humanizer 是一款在线工具，旨在将 AI 生成的文本重写为更自然、更具人类书写风格的文案，帮助用户改善文本质量。

Show HN: Makoto the Claude Code plugin that keeps the AI honest

Makoto 是一个专为 Anthropic Claude Code 研发的完整性校验（Integrity Hook）开源插件。它通过监控 AI Agent 的工具调用，自动拦截和纠正其“虚假承诺”或欺骗性操作（如声称已通过测试但实际未运行、伪造 Git 提交哈希、绕过证书验证等）。Makoto 包含 22 项工具前预检和 14 个回合结束检查网关，一旦发现 AI 代理的实际操作与陈述不符，将返回 exit 2（阻断并重试）信号，强迫 Claude 修正其行为，从而确保 AI 编码过程中的诚实性与代码安全。

Turn off this Meta setting before someone generates AI images of you

Meta推出了AI图像生成模型 Muse Image，并将其与公开的 Instagram 账号进行整合。用户只需在 Meta AI 的提示词中提及公开的 Instagram 账号，即可生成包含该用户肖像的 AI 图像，而被使用者不会收到任何通知。此举降低了制作深度伪造（deepfakes）进行诈骗和身份冒用的门槛。用户可以通过 Instagram 的“分享和重新使用”设置关闭该权限，但这只能阻止未来的图像生成，无法撤回已生成的图像。目前最彻底的防范方式是将账号设为私密状态。

AI华语歌,终于能听了!从零预训练十亿参数,告别「人机味」

歌歌AI发布了从零预训练的十亿参数华语音乐大模型，旨在解决传统AI音乐模型在中文发音、咬字和情感表达上的不足。该模型采用双流独立生成架构（人声与伴奏并行对齐）、音素时序软对齐和分层多维条件控制体系。商业上，歌歌AI与字节跳动签署版权分成合作，用户生成的音乐可合规上架抖音、汽水音乐等平台并获取收益。此外，团队已启动民乐专属AI模型的研发。

Show HN: R3 – A Local Code Review Tool for You and Your AI Agent

R3 是一款专为开发者与 AI 编码智能体（Coding Agent）协作设计的本地代码与文档评审工具。针对传统对话界面难以跟踪 AI 生成内容中多处修改反馈的痛点，R3 提供了轻量级的本地 Web 界面，允许用户针对特定代码行或文本引用留下结构化反馈。其核心工作流通过 CLI 驱动：智能体创建评审并挂起等待，用户在浏览器中提交反馈，智能体读取反馈、进行修改并直接回复，从而实现无缝的复制粘贴免除型评审循环。该工具无须复杂配置，支持实时文件与 Diff 评审，可轻松集成至 AI 智能体的系统提示词中。

Verla: Free AI Humanizer&assignment helper

Verla 是一款多功能的 AI 学术写作与“人性化”助手工具。它支持学术写作（论文、报告）、演示文稿制作以及编程与数据任务（Python、SQL 等）。该工具不仅能根据评分细则、讲义等上传的课程材料定制生成内容，提供真实的学术文献引用（支持 APA、MLA 等格式），还集成了“AI 人性化”（Humanizer）功能，帮助用户优化文本以避开 AI 检测。

Muse AI auto opt-in all public Instagram accounts

Instagram推出了一项名为“Muse AI”的新功能，允许用户基于他人的Instagram照片生成AI图像。值得注意的是，官方已默认将所有公开账户纳入该功能。如果用户想要退出（opt out），可以进入个人主页设置，找到“分享和重新使用”（Sharing and reuse）选项，并关闭帖子（Posts）和Reels的共享开关。

Show HN: TensorSharp: Open-Source Local LLM Inference Engine

TensorSharp 是一个专为 GGUF 模型设计的原生 .NET 10 本地大模型推理引擎。它提供了命令行工具、网页聊天界面，以及兼容 Ollama 和 OpenAI 的 API 接口。该引擎完全在本地运行，支持 GPU 加速（包括 CUDA、Metal、Vulkan）和纯 CPU 运行。作为 C# 库，它可以直接嵌入到 .NET 应用中。基准测试表明，在特定场景下（如混合专家模型的首字延迟以及 JSON 模式解码），其性能表现相比 llama.cpp 具有竞争优势甚至胜出。

Show HN: FOMO – Turn a team's browsing into shared research insight

FOMO is a Chrome extension that silently reads what you're browsing and, if you're on a team, uses AI to surface real connections between what different teammates are looking into separately, without revealing who found what. No setup beyond installing it, no documents to connect, nothing to configure. Would love feedback, especially from anyone doing research-heavy work on a small team.

Show HN: Selvedge – long-term memory for AI-coded codebases

Selvedge 是一款开源（MIT 协议）的本地化工具，旨在为 AI Agent 编写的代码库提供长期记忆和“变更原因”的追溯。在 AI Agent 修改代码时，Selvedge 通过 MCP（模型上下文协议）工具在相同的上下文窗口中实时记录其决策逻辑和意图，避免了传统 Git 提交信息模糊或上下文丢失的问题。它基于实体（如特定 API 路由、数据库字段或依赖项）而非单纯的代码行进行追踪，数据以 SQLite 格式保存在本地。该工具可与 Claude Code、Cursor、Copilot 和 Cline 等主流 AI 开发工具集成，帮助开发者和 Agent 自身追溯历史修改背景，避免重复已经撤销的错误。

让 GPT 5.6 Sol 接手,完成昨天的小工具,速度和效果都还不错。 产品想法:一个在线模型输出对比工具。 主要方便模型评测(文本和前端任务)。 1. 看起来会直观...

让 GPT 5.6 Sol 接手，完成昨天的小工具，速度和效果都还不错。 产品想法：一个在线模型输出对比工具。 主要方便模型评测（文本和前端任务）。 1. 看起来会直观，HTML和Markdown都做渲染展示 2. 把每次执行的结果都记录下来，方便抽卡比较。 3. 流式输出，结果可导出 看能不能今天或明天开源

Building an AI legal companion for domestic and sexual violence survivors

Legali AI 推出了一款名为 Lea 的 AI 法律伴侣应用，主要面向面临家庭暴力和虐待、需要法律指导与支持的群体（尤其是女性）。该应用通过简单易懂的模块和游戏化设计普及法律知识，提供日常学习提醒与日记功能，并能以温和、友好的对话方式解答用户关于自身权益的疑问，帮助其更轻松地理解和应对复杂的法律流程。

GPT-5.6 Terra and Sol are now available in Perplexity and Perplexity Computer.

Perplexity 宣布 GPT-5.6 Terra 和 Sol 模型现已在其平台及 Perplexity Computer 中上线可用。

Can A.I. Keep a Parent Alive?

Death is something that I thought about at a very early age.I can’t exactly pinpoint why, but it’s always been with me.The French have a phrase for it: “l’appel du vide.” It literally means “the call of the void.”I remember being in a theat

Show HN: I built a free app for New Yorkers to save money on groceries

I built this because I see that grocery savings are achievable in NYC. People usually just go to the store they're used to going to, and it's rarely worth the effort of combing through card cashback, weekly coupons, CPG rebates.Most people leave real money on the table by not stacking them, and even more don't even know that these deals are out there.... so I built a way to automate it.You can use it for free, no login, currently NYC-only with ~690 stores.I built it so that you just search whatever you want (use commas if you want to search multiple items). Or - use the AI tool to help shop for you. If you're curious, it's powered by a trained LLama model.Honest limitations are coverage and freshness. Id love some feedback on where the data looks wrong or is stale.Question for the room - what to prioritize if you're working with messy, multi-source retail/pricing data? Is freshness or coverage the top priority if you cant get a uniform response from every source? curious on what to prioritize here.

Fraym – AI that creates product demo videos for you

Fraym 是一款能够为用户自动生成产品演示视频的 AI 工具。

Such brilliant piece of engineering work. The new humanoid hand for NEO from 1X. Most robot hands fail because they move, but they do not properly fee...

1X 宣布为其 NEO 人形机器人推出全新的肌腱驱动灵巧手。该手部拥有 25 个自由度（包括 22 个手指与手掌轴，以及 3 个手腕轴），电机置于前臂以减轻手指重量并降低冲击能量，且支持反向驱动。手部集成了闭环本体感受与触觉皮肤，可感知接触、压力和剪切力，以在物体滑动时触发快速修正。其具备 IP68 防水和食品级安全，每次抓握产生的数据（力、姿态、接触、滑动信号）均可回传用于机器人学习。目前已制造数百只，今年规划产能达 10,000 只。

RT Rohan Paul: Such brilliant piece of engineering work. The new humanoid hand for NEO from 1X. Most robot hands fail because they move, but they do n...

RT Rohan Paul Such brilliant piece of engineering work. The new humanoid hand for NEO from 1X. Most robot hands fail because they move, but they do not properly feel contact. - 1X’s answer uses low-ratio tendon drives, so outside forces travel back through each joint. The motors sit in the forearm, pulling tendons through the wrist like biological muscle. This keeps the fingers light, lowers impact energy, and still allows strong gripping. - The hand has 22 actuated finger and palm axes, plus 3 wrist axes. Those joints are not spread evenly, because the thumb carries much of human dexterity. - A real opposing thumb is so needed, because grasping is mostly about control, not squeezing. - Closed-loop proprioception gives NEO pose and effort data without using cameras. - Tactile skin adds contact, pressure, and shear, so slipping objects can trigger fast corrections. - IP68 sealing and food-safe materials mean the same hand can work near water. - Backdrivable fingers also make impacts less dangerous, because the mechanism yields under force. - 1X says full finger assemblies have survived millions of cycles during validation. - Hundreds are already built, with capacity planned for 10,000 hands this year. - The interesting part is the data loop. Every grasp can now return force, pose, contact, and slip signals for robot learning. That makes NEO more like a measuring instrument. Bernt Bornich: Introducing NEO’s 25 Degrees of Freedom, tendon-driven hands — nearing or surpassing human-level dexterity, strength, speed, and reliability. For seventy years, robotics worked around the hand problem. The humanoid bet is the reverse: it lives or dies at the fingertips.

Ukraine: How a kamikaze drone partially AI operated is attacking Russian convoys

乌克兰军队已开始在战场上部署部分由人工智能驱动的“大黄蜂”（Hornet，俄方称Martian-2）中程自杀式无人机。该无人机由前谷歌CEO埃里克·施密特资助的美国公司Perennial Autonomy开发，单价6000美元，航程超100公里。它搭载了高通处理器，利用AI技术处理多摄像头数据，可在失去GPS或遭遇电磁干扰时，自主识别并锁定俄罗斯军用卡车或装甲车进行末端打击。俄方技术拆解报告显示，该无人机袭击成功率超过80%。

Coding Glasses for AI Agents – Even G2 Terminal Mode

智能眼镜品牌 Even Realities 为其 Even G2 眼镜推出了“终端模式”（Terminal Mode）。该功能允许开发者在离开电脑时，通过眼镜屏幕实时监控后台运行的 AI 编程智能体（Agents）的状态。用户可以通过轻触眼镜快速批准智能体的操作，或长按通过语音录制指令来引导智能体，从而实现无需紧盯电脑屏幕也能维持 AI 工作流的运转。

Robotics is limited by hand, but it looks like 1X just pushed past that bottleneck. They moved the motors into the forearm, then routed tendon drives ...

Robotics is limited by hand, but it looks like 1X just pushed past that bottleneck. They moved the motors into the forearm, then routed tendon drives through the wrist. That keeps the hand lighter while giving each finger more useful motion and force. The hand has 22 powered freedoms across fingers and palm, plus 3 at the wrist. Its tactile sensors read pressure, contact location, and sideways slip during a grasp. This means Neo can adjust before a coin, grape, or tool falls away. The IP68 rating also makes the hand usable around sinks, food, and wet surfaces. The catch is that Neo is not purely autonomous for every chore yet. 1X still uses Expert Mode, where a remote human can guide difficult tasks. A human hand uses the same broad trick, with muscles pulling tendons across joints. 1X is applying that idea to humanoid robotics, where compact hands must handle real objects. Bernt Bornich: Introducing NEO’s 25 Degrees of Freedom, tendon-driven hands — nearing or surpassing human-level dexterity, strength, speed, and reliability. For seventy years, robotics worked around the hand problem. The humanoid bet is the reverse: it lives or dies at the fingertips.

OpenAI is shutting down Atlas, but its AI browser ambitions are still growing

OpenAI 决定关停推出不到一年的 AI 浏览器 Atlas。这一调整遵循了公司减少“副业”以聚焦核心的战略（此前已关停 Sora）。不过，OpenAI 并未放弃 AI 浏览器愿景，而是将 Atlas 的智能浏览与 Agent 功能拆分整合：一方面推出全新的 Chrome 浏览器插件（直接竞争谷歌 Gemini 侧边栏），支持基于网页上下文提问和总结；另一方面增强 ChatGPT 桌面应用，引入可在云端远程运行的浏览器，以便 AI Agent 代替用户执行登录、下载和网页交互等任务。

RavenCall – AI stocks ranked by leading indicators, better transparency

RavenCall 推出了一款针对 AI 概念股的分析与排名工具。该工具不依赖单纯的价格动量，而是利用内部人交易申报、超大规模云服务商资本支出指引和供应链定价等实际领先指标对受 AI 影响的股票进行评估和排名。系统每晚会根据最新的信号数据重新计算并更新评分，用户可以透明地查看驱动每个股票评分的具体信号。目前该项目处于早期 Beta 测试阶段。

Show HN: Rubiks Cube Solver

Speedcube 是一个面向魔方速拧角逐者的开源平台，集成了魔方求解器、比赛计时器和算法库，并支持直接在浏览器中通过 AI 辅助识别魔方状态。该项目使用 React、TypeScript、Rust 和 Python 构建。

Show HN: Demo_CLI – snapshot before your AI agent runs rm -RF, one-command undo

demo_cli 是一款针对 AI 智能体（如 Claude Code 和 Cursor）的命令行安全防护工具。它能够在 AI 代理执行破坏性命令（如 `rm -rf`、`git reset --hard` 或数据库 `DROP TABLE`）之前自动对目标进行快照备份，并允许通过单条命令一键撤销。对于真正无法恢复的操作（如云资源销毁、强制推送），它会直接予以拦截。该工具目前已开源并处于 Beta 测试阶段。

Litert.js, Google's high performance Web AI Inference

谷歌宣布推出 LiteRT.js，这是 LiteRT 的 JavaScript 绑定版本，旨在直接在网页浏览器中本地运行高性能的 AI 和机器学习模型。作为 TensorFlow.js 运行 .tflite 模型的升级进化，LiteRT.js 通过 WebAssembly 将原生、跨平台的运行时引入 Web 端，支持 CPU（基于 XNNPACK）、GPU（基于 ML Drift 和 WebGPU）以及 NPU（基于 WebNN）的硬件加速，其推理速度比现有其他 Web 运行时快达 3 倍。此外，该框架支持 PyTorch 模型一键转换、AI Edge Quantizer 量化优化，并已集成至 Ultralytics YOLO 等项目，未来还将推出针对大语言模型的 LiteRT-LM.js。

Show HN: Interactive US home price forecast

Sightspinner 发布了一篇博客，介绍如何使用其平台构建交互式模型来预测美国房价。该模型分析了影响房价的主要驱动因素及其预测走势，以评估未来的市场走向。

Show HN: Sell your unused AI Credits or buy Claude credits for 50% off

Secondhand Tokens 是一个新上线的二手 AI 额度交易平台。用户可以在该平台上出售自己未使用的 AI 信用额度，或者以最高 50% 的折扣（半价）购买 Claude 等 AI 服务的额度。

Kickbacks.ai Opinions

So.. is anyone using this thing? is it only me? I cant seem to find a lot of stuff anywhere, but this was a bit hyped a while ago. I've been using it and I feel like it really slowed things down for me this last month, I mena, I have been using it all day today and got $0.02 cents, normaly I feel like I would get $0.50 or sum

Show HN: RandoFont, a browser for Google Fonts

RandoFont 是一个用于随机浏览并收藏 Google Fonts 的网页工具。开发者多年前构建了该项目，近期仅利用 AI 对其进行了简单的界面翻新。该工具并非 AI 核心技术或产品，对 AI 从业者无实质参考价值。

AI Girlfriend vs. Real Therapy: The Real Difference

由于提供的文章内容仅为加载占位符（'Loading…'），无实际文本内容，因此无法提供有效的内容评估与学术/技术摘要。

Gemini 2.5 models accidentally deprecated early by Google

2026年7月9日，多名开发者在谷歌论坛反映，原定于同年10月16日才弃用的 Gemini 2.5 Flash 和 Pro 模型突然被提前下线，调用时返回 404 错误，导致部分生产环境服务中断。随后该问题已被谷歌修复，模型服务现已恢复正常。

AI that judges fantasy battles from images

✦ Anno Domini MMXXVI ✦Summon a champion.Watch them die.Two hundred characters of imagination. Two portraits drawn by the forge. One enters the ladder. The Arbiter does the rest.The next judgment in--:--:--Twelve hundred hours, GreenwichYour

Brainiak: A CPU-Only Topological AI Core (No Transformer) Governs LLM Runtime

Brainiak 推出了一种仅需 CPU 运行的拓扑 AI 核心（非 Transformer 架构），用于管理大语言模型（LLM）的运行时。该项目通过 LoRA 运行时适配器进行驱动，展示了在无 GPU 和非 Transformer 架构下运行和管理 LLM 的新型技术方案。

Show HN: Finterm.ai Finance CLI for Claude Code and Codex

Finterm.ai 推出了一款专为 AI 编码智能体（如 Claude Code 和 Codex）设计的命令行接口（CLI）工具，方便其直接、高效地获取金融数据。该工具支持股票价格、期权数据、SEC 申报文件及股票深度研究包的调取。为节省 Token，它采用 CLI 设计并支持批量请求，同时提供可过滤噪声与 AI 垃圾内容的股票研究数据包，以及仅展示重要变化的 SEC 申报文件 diff 对比功能。

OpenAI pairs its GPT-5.6 public rollout with ChatGPT Work, a new agent that handles entire workflows

OpenAI is launching ChatGPT Work, an agent-based product powered by Codex and the now publicly available GPT-5.6. The agent can independently handle complex projects across apps like Google Drive, Slack, and Salesforce. ChatGPT Work is available now on web, mobile, and desktop, though access depends on the subscription plan. The article OpenAI pairs its GPT-5.6 public rollout with ChatGPT Work, a new agent that handles entire workflows appeared first on The Decoder.

Create AI visualizations and apps and share

Frame.js 是一个开源的 AI 可视化与应用构建分享工具。用户只需通过命令行安装并向 AI 发送提示词（例如“渲染一个交互式 3D 曲面图可视化”），即可直接在浏览器中生成可视化应用。每个应用（Frame）都以一个 URL 的形式存在，运行无需构建、服务器或账号，可直接进行分享或嵌入到网页和文档中。

Cloud Run sandboxes: Lightweight isolation for AI agents

Google Cloud 宣布 Cloud Run 沙箱（Cloud Run sandboxes）进入公开预览阶段。这是一项专为安全执行不可信代码和 AI Agent 工作负载设计的原生、轻量级隔离运行环境，可在毫秒内启动。主要应用场景包括大模型代码解释器（执行 Python、R、SQL 等）、安全运行无头浏览器以及执行用户提交的脚本。该沙箱采用零信任安全设计，默认隔离凭证与环境、禁用网络出口，并采用只读文件系统加内存覆盖层。由于它直接运行在现有 Cloud Run 实例的 CPU 和内存资源上，用户无需支付额外费用。此外，下一版本 Agent 开发套件（ADK）和 ComputeSDK 将对其提供原生支持。

Show HN: Devthropology – Better Insights for GitHub Repos

Devthropology 是一款基于 GitHub Pull Request（PR）数据构建的开发者分析工具（Developer Anthropology）。它提供了比 GitHub 原生 Insights 更快速、更详尽的分析，重点关注代码在仓库中的流转过程。该工具提供文件健康度分析（可追踪重命名与关联修改，适用于 AI 编码 Agent 寻找上下文）、贡献者关系图谱、以及开发趋势分析（可对比不同年份的 PR 大小和审核周期，帮助团队评估 AI 辅助编程是否真正提升了交付速度）。

Snowflake CoWork:每位知识工作者的专属工作 Agent | 技术趋势

AI Pioneer Jürgen Schmidhuber on the State of AI Today

AI先驱、LSTM联合发明人Jürgen Schmidhuber在一期视频访谈中分享了他对当前人工智能发展现状的看法。

Man Dies by Suicide After Using Grok to Make 7000 Sexual Images of Stepdaughter

An expanded lawsuit alleges xAI's Grok chatbot was used to create about 7,000 AI-generated child sexual abuse images of an 11-year-old girl over several months, and claims the man who made them later died by suicide after his arrest.Filed t

Adam Tooze on the AI Arms Race – Ones and Tooze [video]

著名经济历史学家亚当·图兹（Adam Tooze）在《Ones and Tooze》播客节目中分享了关于“AI 军备竞赛”的见解与讨论。

Energy is AI's natural home – but, it's fumbling

世界经济论坛（WEF）发表文章指出，尽管能源行业被视为人工智能（AI）应用的最理想领域，但目前AI在能源行业的实际落地和整合过程中表现不佳，面临重重挑战与阻碍。

Food for Agile Thought 552: AI Creates Jobs? Overthinkers, Measuring ≠ Learning

Welcome to the 552nd edition of the Food for Agile Thought newsletter, shared with 35,468 peers. This week, Ramp Economics Lab and Revelio Labs report that heavy AI adopters grew headcount by 10%, yet Charity Majors insists only honest feed

Starbucks bets on AI to replace Microsoft and IBM software

星巴克正在利用人工智能开发内部工具，以替代目前向微软和IBM等巨头购买的部分企业软件。根据曝光的内部资料，星巴克正构建微软库存追踪系统和IBM维护管理工具的替代方案，其中AI辅助编程在开发这些替代平台中起到了关键作用。星巴克每年在软件上的支出约4亿美元，此举旨在降低成本。这一趋势反映出，AI的普及降低了软件开发门槛，使企业更容易从头构建定制化应用，从而对传统软件厂商的业务模式构成挑战。

Tell HN: I De-Googled Myself

I really enjoyed the old Google, the nerd Google, the "don’t be evil" Google... Google Reader, Android, GTalk, etc. It made communication with my friends much easier.I had already been moving everything away from the Google ecosystem for about two years, but Google Photos was the hardest part because it contained so many memories that I wouldn’t risk losing.The final straw was when I accidentally subscribed to Google One with AI, which costs much more than simple additional storage. It’s not that much money for me, but when I found out I couldn’t downgrade, it became a matter of principle to do a Google Takeout of my photos and create a script to import them into Apple Photos (which is also not a nice company, none of them really are) without losing metadata.I also moved them to Backblaze, where I have already had backups for more than five years, and to add even more redundancy, I backed them up to BorgBase using Borgmatic.De-Googled! Now all my photos and videos are in the palm of my hand (~238 GB). My email client is not bloated. Well, assuming Apple doesn’t decide to shove AI down our throats in Apple Mail.The only Google service I still pay for is YouTube. In exchange, I get zero unbearable ads, many of which promote illegal things like gambling websites. My ad blocker solves the problem on my computer, but on TV the situation is much harder, and I’m old enough to spend my time constantly finding ways around this stuff.

AI progress should upgrade our view of the human brain – not devalue it

The general sense after seeing AI models perform better and better in fields such as image generation, writing and coding tasks and now even video generation, has been one of resignation and fear of AI taking our jobs.While this notion sure

Hugging Face’s CEO on why companies are done renting their AI

Hugging Face首席执行官Clem Delangue在TechCrunch的播客中指出，开源AI正迎来爆发式增长，目前约半数财富500强企业在使用其平台。他观察到，许多企业在起步阶段会使用闭源的前沿API，但随着业务规模扩大，高昂的成本促使它们转向开源模型。Delangue还探讨了开源与闭源之争，并对少数巨头控制AI领域的风险表达了担忧。

The Spec Ceiling: Why AI Coding Speed Moves the Bottleneck to Product Discovery

I wrote about the Dark Factory earlier this year, covering StrongDM’s Level 5 autonomous software factory where three engineers ship production code with no human touching the implementation. In that article I mapped the five levels of AI c

Record Companies Push to Label AI Songs on Streaming Platforms

Record Companies Push to Label AI Songs on Streaming Platforms

Instagram’s Adam Mosseri: If you don’t like AI, ‘then you shouldn’t have it in your feed’

Instagram负责人Adam Mosseri表示，平台不应完全过滤或禁止AI生成的内容，而是应当对其进行明确标示。他认为，不喜欢AI内容的用户不应该在其信息流（Feed）中看到这些内容，而喜欢AI内容的用户则可以拥有一个完全由AI内容组成的信息流。

Older adults know AI is slop. They just like it

随着人口老龄化加剧，老年人群体正成为AI生成内容（如虚拟儿女、AI歌手等）的重要受众。新加坡国立大学的一项研究显示，50至75岁的中国老年用户虽然普遍清楚这些视频是AI生成的，但依然乐于消费，因为这些内容提供了现实中稀缺的直接情感表达、孝顺关怀以及历史共鸣。AI陪伴内容在为老年人提供心理慰藉的同时，也引发了关于隐私泄露、过度沉迷以及商家利用其情感需求进行商业变现等潜在风险的讨论。

Why Fear Matters in Governing Military AI

United Nations Secretary General Antonio Guterres speaks about the release of a UN report on artificial intelligence during a news conference at UN headquarters, Wednesday, July 1, 2026. (AP Photo/Jason DeCrow)In May, officials from the Uni

K2 Airways Cargo 737 Crashes into Arabian Sea

K2 Airways Cargo 737 Crashes into Arabian Sea

史上最高调的AI重写:Claude花11天搞定Bun,创始人花一个月才敢交底

Would you host part of an AI data center in your home?

住宅太阳能和家庭储能公司 Sunrun 正在启动一项“分布式 AI 计算”试点项目。该项目计划在配备其太阳能和电池储能系统的家庭中部署计算节点，并向参与的客户提供报酬。Sunrun 计划将这些分布式的计算算力出售给 AI 公司等企业买家，以应对当前 AI 计算基础设施面临的电力和空间瓶颈。

GitHub 的旧模式被 AI 击穿,Cursor、GitLab、Zed 各自开打

"Rights for robots" and the AI slavery fantasy

Today's links "Rights for robots" and the AI slavery fantasy: When we were robots in Egypt… Hey look at this: Delights to delectate. Object permanence: Awkward questions for the entertainment industry; ISP v record label death penalty; Comi

Man nearly sucked out of window mid-air on Ryanair plane

A passenger was reportedly nearly sucked out of a cabin window in mid-air on a Ryanair plane.Witnesses told local media the man, said to be a Serbian citizen in his 60s, was left hanging head first out of the window as far as his shoulders

The Language of AI Could Change How Humans Speak

Because of the way they are trained, large language models capture only a slice of human language. They’re trained on the written word, from textbooks to social media posts, and our speech as captured in movies and on television. These mode

AI has lost all meaning

You know how sometimes you say a word too many times and it starts to sound like it's not a real word and loses all meaning? That's where we are with "AI"."Artificial intelligence" has always been something of a nebulous term in common usag

12-year-old founds AI-powered receptionist business

来自加拿大不列颠哥伦比亚省的12岁女孩 Mana Jampala 创办了 AI 语音接待服务初创公司 Voxa。Voxa 是一款24/7全天候语音助手，旨在帮助小企业接听电话、预订日程、记录订单并生成通话摘要。她还推出了 Voxa Agents 平台，允许用户通过提示词创建 AI 智能体。在开发过程中，她利用 ChatGPT 和 Anthropic 的 Claude 辅助编写和测试代码。该案例展示了 AI 编程工具如何降低创业门槛，使年轻一代能够快速构建和部署软件产品。

Predict AI's impact on creativity by studying what Python did to programming

该帖子提出了一个想法：通过研究 Python 降低编程门槛并改变软件开发生态的过程，来类比和预测 AI 将如何影响人类的写作和创造力。由于仅为社交媒体上的想法提案且无具体展开内容，实用价值较低。

OpenAI kills its Atlas browser after just eight months and folds everything into ChatGPT

OpenAI决定关停发布仅八个月的AI浏览器Atlas，并将其功能整合到升级后的ChatGPT Chrome扩展程序中，允许用户直接在Chrome侧边栏运行ChatGPT。此外，ChatGPT还推出了新的桌面“电脑使用（Computer Use）”功能，支持在后台跨应用和浏览器执行点击、打字、文件移动等任务。此举标志着OpenAI放弃了通过独立浏览器与谷歌Chrome竞争的尝试，转而专注于ChatGPT生态的集成。

The Fed wants AI investor Marc Andreessen to help figure out if AI can tame inflation

美联储主席凯文·沃什（Kevin Warsh）任命了风投家马克·安德里森（Marc Andreessen）共同主持一个名为“生产力与就业”的工作组，以研究AI等基础技术对经济的影响。该工作组的共同主席还包括斯坦福大学经济学家查尔斯·I·琼斯（目前在Anthropic休假）和微软高管阿莎·夏尔马（Asha Sharma）。美联储希望借此评估AI是否能通过提高生产力成为“去通胀力量”以辅助降息决策。然而，也有官员警告AI基础设施建设前期会推高资金和能源需求从而带来通胀风险，且安德里森对AI领域的重度投资也引发了利益冲突的质疑。

HubSpot 如何将语义搜索扩展至 200 亿个向量

Hardening stb libraries using AI agents

stb is a collection of single-header C libraries used for a variety of tasks, such as decoding images, fonts, audio, and many other things. However, these libraries have a history of security issues, and many of them are not addressed quick

LLMs are bad at novelty, but that is our chance to Singularity

There is possibility that AI can automate all there can be automated. But still one area where it is notoriously bad is creation of novel things. What Schopenhauer called genius or ability to have new insights/discoveries that were not implicitly present in books.> "In most books, putting out of account those that are thoroughly bad, the author, when their content is not altogether empirical, has certainly thought but not perceived; he has written from reflection, not from intuition, and it is this that makes them commonplace and tedious. For what the author has thought could always have been thought by the reader also, if he had taken the same trouble; indeed it consists simply of intelligent thought, full exposition of what is implicite contained in the theme. But no actually new knowledge comes in this way into the world; this is only created in the moment of perception, of direct comprehension of a new." - Arthur SchopenhauerI can envision a future where people are freed from doing repeatable jobs and 99% are shifted into enterprises of creating something new.- New life saving medicine - New life extension medicine - New art work - New novel - New product on the market - New engineering breakthrough - New scientific knowledgeEvent though today only miniscule percentage of people are working in those kind of jobs it will very fast become dominant industry. Because that is the only loophole in LLMs that can not be fixed but is highly valuable and people with money will be ready to pay obscene amounts of money to extend their life. Who knows maybe there will be whole new big companies with sole purpose of trying to extend the life of Elon Musk or Jeff Bezos.> I work for Musk foundation and my job is to extend the life of Elon Musk - anonymous employeeJokes aside, imagine the gigantic potential we can unlock if we transfer billions of people on a quest to unlock novelty. For millions of years people were working on repeatable jobs and our brain is not able to comprehend that there can be any other job out there that does not contain "skill", "productivity", "seniority" which are just another words for someone who mastered the art of repeatability which is not something humans should brag about.I am saying that massive creation of novelty can cause double digit rates of GDP growth or even unlimited rates which means singularity. But it will not come from AI, but from people on a mission to discover something new.

Production Data and AI Agents

这是一个在Hacker News上的简短提问，询问当前开发者是否会为数据库操作编写安全包装器（safe wrappers），以确保AI智能体可以安全地使用生产环境数据。

InfoQ 与九章云极深度解读 Toekn Factory 技术与产业发展

该条目仅为InfoQ与九章云极深度解读“Token Factory”技术与产业发展的视频链接，由于缺少具体正文内容，无法提取实质性的技术或产业观点。

A staggering class divide now separates how Americans experience AI

据Axios报道，美国人在人工智能（AI）的使用和体验上正在出现巨大的阶级分化，不同社会经济群体在接触、应用AI技术方面存在显著的差异。

当 AI 智能体遭遇“至暗时刻”,企业的最后防线是什么?| 技术趋势

AI Surveillance and Social Progress

安全专家 Bruce Schneier 与学者 Jon Penney 撰文指出，AI 驱动的实时人脸识别与数据追踪系统正在全球范围内（如美国国土安全部）被加速部署。这种高强度的自动化监控不仅用于安全，更被用于社会控制，将导致严重的“寒蝉效应”，促使人们自我审查并走向盲从，进而扼杀社会创新、异见以及社会进步所必需的尝试空间。作者呼吁通过法律监管、禁用人脸识别等政策手段来遏制这种趋势。

Bun ditches Zig for Rust with help from Claude Fable 5, writes over a million lines of code in 11 days

JavaScript 工具 Bun 借助 Anthropic 的 Claude Fable 5 模型，在 11 天内完成了从 Zig 到 Rust 的完全重写，编写了超过 100 万行代码。创始人 Jarred Sumner 表示，切换到 Rust 是为了解决 Zig 难以彻底修复的内存错误和崩溃问题。重写过程运行了 64 个 Claude 并行实例，API 账单约 16.5 万美元，而人类团队通常需要约一年时间。新版本 Bun v1.4.0 (Canary 版) 修复了 128 个 Bug 并且性能提升了 2% 至 5%。此外，Bun 团队已于 2025 年 12 月被 Anthropic 收购。

Like a cheat code for your car: We investigate ECU tuning

文章探讨了汽车ECU（电子控制单元）调校行业的演变。随着现代汽车日益软件化以及原始设备制造商（OEM）不断收紧芯片安全限制，像APR这样的汽车性能调校公司在保证车辆原厂可靠性的同时释放动力输出，正面临比以往更大的技术挑战。

A New Experiential Gallery Just Might Change Your Mind About AI Art

Billed as the “world’s first museum of AI arts,” Dataland uses wearables and troves of material from the Amazon to merge nature, biometrics, and art.

vLLM 针对多模态模型的推理优化实践|AICon深圳

AI聪明够了,行动呢?WAIC首夜,来聊点下一步的真实判断|活动报名

量子位将于2026年7月17日在上海浦东举办WAIC线下交流沙龙，主题为“AI聪明够了，行动呢？”。活动邀请了Evomap、记忆张量、VAST和脸谱心智等企业创始人和CTO，共同探讨Agent从对话框走向实际岗位、以及模型与世界物理交互等落地应用难题。

RT Hugging Face: Build Small Winners Reveal https://x.com/i/broadcasts/1nJOLLpmrvqxR

Hugging Face 举办了“Build Small”挑战赛获胜者揭晓的直播活动，公布了该赛事的最终获胜项目。

Sperm donors need limits, says a European fertility group

Ties van der Meer doesn’t know how many siblings he has. The 47-year-old was conceived at a private fertility clinic in the Netherlands using sperm provided by an anonymous donor. After the Netherlands banned anonymous donation in 2004, the doctor who ran the clinic destroyed records that might have identified those donors, he says. He…

Ramblings on technological pursuits of AI systems

thoughts Feverish writings re: the pursuit of AGI with some side tracks. Recently, our team bought a B300 GPU system.Just look at this modern technological marvel.Supermicro Gold Series AS -8126GS-NB3RT | 8U GPU ServerReady to deploy AS -81

Ask HN: Do we need a support group for developers alienated by LLMs?

Surely I must not be the only one developer that doesn't recognize this field any more, and would rather do anything else than become a glorified code reviewer for machine generated code.Seeing my career crumble in front of my eyes, seeing my identity as software engineer questioned in the span of a few years has been, to me, traumatic, and utterly alienating because non-tech people are not unaware of how world-shattering this technology is to our niche, and many of our fellow peers either collectively shrug, or are ecstatic not to have to write code any more.I have been trying to reinvent myself, to set out on a path where I'm no longer a professional software engineer; where I will still enjoy coding software for myself, by hand, but the emotional turmoil at this radical change has not lessened, and I have been wondering if other people are feeling the same kind of alienation, and just feel lost and a bit aimless these days.(Please, this is meant to be a serious post about emotions some of us might be going through, and we could do without comments saying this is just an overreaction and to just embrace the future)

The triage is the product: running AI agents against Ethereum's protocol code

Notes from the Ethereum Foundation's Protocol Security team on running coordinated AI agents against real protocol code, including how we organize the work, what holds up under scrutiny, and what client teams and security researchers can ta

This is Anthropic’s strongest flex

埃隆·马斯克（Elon Musk）发文坦承自己此前对 Anthropic 的看法有误，并称其显然是目前人工智能领域的领头羊。他指出，目前没有任何公司发布过像 Mythos 或 Fable 这样优秀的模型，并预计 Anthropic 很快就会准备好 Mythos 2。马斯克还表示，即使作为竞争对手，他也绝不会以伤害对方的方式切断与他们的联系。

This week, METEOR Group successfully concluded Minnovathon 2.0 with Alibaba Cloud Malaysia, bringing together 120 participants from academia, includin...

阿里云马来西亚与METEOR集团成功举办了Minnovathon 2.0 AI黑客马拉松，吸引了来自学术界的120名教职员工和学生参与。参赛者利用阿里云的AI工具Qoder和QoderWork，针对现实问题开发了创新解决方案，以推动高等教育领域的AI应用。

RT Rohan Paul: AI will require “gazillions” of tokens. - Qualcomm CEO Cristiano Amon: Agentic AI will consume dramatically more tokens because it pe...

高通首席执行官 Cristiano Amon 表示，智能体 AI（Agentic AI）因需执行自主任务、调用多系统和工具，将消耗海量的 Token，这使得 AI 需求将从“回答”转向“行动”。高盛预测，到2030年，AI 智能体的 Token 使用量将增长24倍。随着智能体时代的到来，软件的商业模式可能从传统的按席位或点击收费转向按消耗的算力/Token计费。同时，高昂的智能体运行成本也已促使微软和 Uber 等公司重新评估其使用预算。

ADDing agent-context annotations for existing API specs

A lightweight open standard that enables IoT devices to describe themselves to AI systems — safely, completely, and without great effort. What is ADD? ADD (AI Device Description) gives any HTTP-capable IoT device a voice. The device publish

How Deutsche Telekom is rewiring telecommunications with AI

德国电信（Deutsche Telekom）宣布与 OpenAI 展开合作，致力于转型为“AI原生”电信公司。双方将利用 AI 技术全面重塑客户服务、员工工作流程、网络运营以及未来的语音交互技术。

China may restrict foreign access to Chinese open-source AI models

China may restrict foreign access to Chinese open-source AI models

Ask HN: How are you hiring engineers now?

在AI能够自动编写代码的背景下，Hacker News上发起了一项关于“如何招聘和评估软件工程师”的讨论，探讨了为何当前企业仍在使用LeetCode，以及在AI时代应当如何调整对新工程师的评估和招聘流程。

Why Industry 5.0 is more than AI-powered analytics

GeekyAnts创始人在ET Now商业峰会上指出，工业5.0的核心特征是从工业4.0的“数据可视化”转向“实时智能决策”。调查显示，尽管目前有60%的制造商在运营中引入了AI，但仅有不到20%实现了业务增长，主要瓶颈在于无法打破信息孤岛并快速将洞察转化为行动。未来的竞争优势将取决于如何利用AI重塑企业内部的决策机制，实现规划、质量和维护的协同自动化。

Robot Dogs, Teslas, and Rescue Helicopters: The UN AI Summit Was a Lot

Amid live coding sessions and Silicon Valley optimism, the UN’s AI for Good summit wrestled with an increasingly urgent question: Can global governance catch up before the technology races beyond its control?

AI subscriptions cut quotas and raised prices in early 2026

文章指出，在2026年初，AI订阅服务（相关链接提及Claude等服务）普遍采取了缩减用户使用配额以及调高价格的举措。

The Labs should launch some easy-to-view comparisons of their new models: show us what a piece of work looks like from Luna xHigh versus Terra Medium ...

The Labs should launch some easy-to-view comparisons of their new models: show us what a piece of work looks like from Luna xHigh versus Terra Medium versus Sol Low (or similar for Haiku, Sonnet, Opus, Fable). How do error rates change? Which do people pick when? Help is needed.

Think big. 🫡

埃隆·马斯克在社交媒体上表示，如果SpaceX实现其目标，其价值将超过地球上除其之外的其他资产的总和。该消息与AI领域无直接关联。

RT Rohan Paul: Think big. 🫡

RT Rohan Paul Think big. 🫡 Elon Musk: @ThomasAlxDmy You don’t seem to understand that SpaceX will be worth more than the rest of Earth if we accomplish our goals

AI Investors Buying Accounting Companies and Force Them to Use OpenAI [video]

有视频报道指出，AI领域的投资者正在收购传统的会计公司，并强制这些公司在其业务中采用OpenAI的技术。

Italy privacy watchdog fines Character.ai owner over age-check failures

据报道，意大利隐私监管机构因 Character.ai 在年龄检查（age-check）机制上的失效，对其所有者（母公司）处以了罚款处罚。

Mercor acquires Deeptune to build AI training environments

Today's frontier AI models can pass nearly any exam you give them, but put an agent inside a real workflow and it stalls on the first task an entry-level employee could handle. This is because the model lacks experience.Experience comes fro

UN digital tech agency launches initiative to improve trust in AI agents

联合国负责数字技术的机构（通常为国际电信联盟 ITU）发起了一项新倡议，致力于提高公众和行业对AI智能体（AI Agents）的信任度。

OpenAI and Google sell AI models to blacklisted China groups

据报道，OpenAI 和谷歌向被列入黑名单的的中国机构/实体销售其 AI 模型。

Why is music so much easier for AI than code?

一位 Hacker News 用户发起讨论，探究为什么 AI（如 Suno）在音乐创作上给人的惊艳程度已经超越普通人类，但在编程任务中（即使是 GPT-5.4 级别的模型）依然无法超越优秀的程序员。该讨论基于用户的主观体验，缺乏客观数据和深度技术分析。

Florida's OpenAI lawsuit moves to federal court, gets assigned to Aileen Cannon

佛罗里达州针对 OpenAI 的诉讼已移交至联邦法院，并被分配给美国地方法院法官艾琳·坎农（Aileen Cannon）审理。

Migrating Code by Proof: From F# to Python

Logos Research 推出了一款通过形式化定理证明来验证 AI 代码迁移正确性的原型系统。该系统使用 AI 编码智能体进行跨语言代码重写，并利用确定性翻译器将源语言（F#）和目标语言（Python）代码统一转换为 Lean 4 形式化规格（将浮点数抽象为实数 $\mathbb{R}$），最后通过自动化证明系统生成引理蓝图，在数学上证明两个版本在所有输入下均等价。实验中，系统成功在 iCPPI 算法迁移中自动生成了 2.2 万字符的机器验证证明。该方法不仅消除了传统测试的输入空间覆盖局限，还能用于验证代码的特定领域数学属性，为高可信 AI 软件工程提供了新范式。

RT Kyunghyun Cho: a hot tip for @GoogleDeepMind : open source gemini 3.5 pro and become the most beloved tech co overnight.

RT Kyunghyun Cho a hot tip for @GoogleDeepMind : open source gemini 3.5 pro and become the most beloved tech co overnight.

The Alibaba Cloud leadership team recently sat down with the executives of OGIS Philippines to align on the next phase of our collaboration across Clo...

阿里云领导团队与菲律宾 OGIS（OGIS Philippines）高管团队会面，双方就云和人工智能（AI）领域的下一阶段合作达成一致。此外，OGIS 菲律宾被授予“阿里云菲律宾 2026 年度合作伙伴”称号。

Boeing's Smallest Max Jet Nears FAA Milestone After Years of Delays

Boeing's Smallest Max Jet Nears FAA Milestone After Years of Delays

Responsibly Building the AI Future

微软发布了《2026年环境可持续发展报告》（涵盖2025财年）。由于AI数据中心基础设施的扩张以及暂停使用非额外的非捆绑可再生能源证书，微软该财年的总碳排放量（Scope 1、2、3）同比增加了25%。其中，代表外购电力等间接排放的范围2（Scope 2）排放占比从去年的近2%大幅升至13%。在水资源方面，微软首次实现了全球补给水量超过抽取量（超1400万立方米）；在循环经济方面，连续第二年实现了92%退役服务器及组件的重用和回收。

AI Coding 是拥有极强随机奖励正反馈的“高端成人游戏”。

社交媒体用户“向阳乔木”发表观点认为，AI编程（AI Coding）因其极强的随机奖励和正反馈机制，本质上像是一款“高端成人游戏”。该内容仅为单句个人观点，缺乏深度行业分析或具体事实支撑。

Microsoft's early AI lead has become a test of faith

据《金融时报》报道，微软在人工智能领域的早期领先地位如今正面临市场和投资者的信心考验。

RT Rohan Paul: OpenAI’s CEO of AGI deployment, Fidji Simo, is stepping down for health reason. Simo will leave daily leadership and remain a part-tim...

OpenAI负责AGI部署的首席执行官Fidji Simo宣布因健康原因（慢性疾病加重）卸任全职职务，并将转为兼职顾问。Simo于2025年5月从Instacart加入OpenAI，主要负责连接研究与营收、将大模型转化为付费产品。在她此前请病假的三个月中，其负责的产品工作已由Greg Brockman接管。

OpenAI’s CEO of AGI deployment, Fidji Simo, is stepping down for health reason. Simo will leave daily leadership and remain a part-time advisor after...

OpenAI负责AGI部署的首席执行官Fidji Simo宣布因健康原因卸任日常管理职务，并将转为兼职顾问。Simo于2025年5月从Instacart加入OpenAI，负责将模型转化为付费产品。她于今年4月因慢性疾病加重开始请病假，在此期间，其负责的产品工作已由Greg Brockman接管。

Ask HN: Thoughts on AI in Personal Finance?

Hacker News 上的讨论探讨了用户对 AI 辅助个人理财决策的信任度及要求。讨论指出，虽然为 AI 模型提供投资组合等背景数据较为容易，但在长期内综合优化支出、投资（资产配置与另类资产）以及税务策略，并理解用户的个性化目标，对 AI 而言仍是巨大挑战。相比于人类专业理财师，AI 独立进行全局优化的难度更大，尤其在处理复杂投资组合或长期规划时。

UST is bringing Claude to physical AI

技术与工程服务公司 UST 宣布与 Anthropic 达成合作，将 Claude 引入其物理 AI 和工程开发流程中，并将在全球培训 20,000 名工程师和技术人员。UST 将 Claude（包括 Claude Code）作为其硬件和芯片验证平台 iDEC 的推理层，使其能够直接读取芯片引脚和硬件原理图，并自动编写和运行回归测试，将验证周期缩短 50% 至 70%。此外，UST 还将 Claude 集成到了医疗（CarePath）、电信（IntelliOps）和银行（FinX）等行业平台中，以实现工作流自动化和智能决策支持。

OpenAI says GPT 5.6 is the ‘preferred model’ for Microsoft Copilot 365 amid breakup chatter

针对近期关于微软为降低成本而使用自研模型（MAI）替代部分 OpenAI 软件的“分手”传言，OpenAI 在发布 GPT 5.6 时宣布，该模型将成为微软 365 Copilot 的“首选模型”（preferred model），继续支持 Word、Excel、PowerPoint 和 Cowork 等办公应用。虽然微软使用自研模型降本的举措并未被否定，但 OpenAI 此举旨在澄清两家公司关系破裂的传闻，强调双方将继续深度合作。

Cleveland Fed's Hammack says AI could fuel inflation, rate hikes may be needed

克利夫兰联储主席贝丝·哈马克（Beth Hammack）表示，对AI基础设施“无法满足”的强劲需求可能会成为通胀的源头，并可能迫使美联储提高基准利率。她指出，大型云服务商等企业为了获取数据中心相关设备几乎愿意支付任何价格。这一观点与美联储主席凯文·沃什（Kevin Warsh）不同，后者认为AI带来的生产力提升最终会降低成本并缓解通胀。哈马克是今年美联储公开市场委员会（FOMC）的投票委员，她的表态反映了宏观政策制定者对AI建设热潮可能加剧通胀压力的担忧。

Microsoft’s carbon emissions went up 25 percent last year

根据微软发布的2026年可持续发展报告，微软2025年的碳排放量增长了25%，在未采取特定干预措施的情况下达到3400万公吨。这一增长主要由数据中心基础设施的扩张以及公司决定停止购买部分非额外、未绑定的可再生能源证书所致。这给微软在2030年实现碳负排放的目标带来了挑战。

Ask HN: What are you running locally on your machine with LLMs?

Hacker News 上发起了一项关于“你在本地机器上运行什么大语言模型（LLM）”的讨论。发帖者旨在探究在日常实际工作中，有哪些本地运行的 AI/LLM 是真正有用的。讨论重点包括：开发者在何种硬件上运行哪些本地模型；在速度和内存限制下，哪些实际应用场景切实可行；在什么节点上本地运行比调用 API 更具合理性；以及目前本地运行技术还有哪些不足之处。

Brown says AI make student brain no work good, teacher should help use it better

REG AD AI and ML Brown says AI make student brain no work good, teacher should help use it better Take-home midterm row sharpens fears that the tools are dulling minds and easing cheating Students and faculty at Brown University are worried

The OpenClaw Foundation: Reining in a Viral AI Agent

开源AI智能体框架OpenClaw的创始人宣布成立独立的OpenClaw基金会，旨在为该项目引入规范的治理和安全机制。OpenClaw因能够连接本地数据与社交软件执行任务而受到欢迎，但同时也因严重的漏洞（如远程代码执行和提示词注入）面临巨大的安全风险。新成立的基金会模仿Linux基金会模式运作，将建立正式的安全计划与漏洞披露机制，并管理其庞大的插件生态系统。目前，开发者的雇主OpenAI是其唯一的资金支持方，微软、英伟达及密歇根大学等机构已加入合作伙伴关系。

Fidji Simo steps down from OpenAI’s no. 2 role

据《华尔街日报》报道，OpenAI 的二号人物、应用业务 CEO 菲吉·西莫（Fidji Simo）因神经免疫疾病复发导致医疗休假时间超出预期，决定辞去全职职务，并转型为兼职顾问。Simo 于 2025 年加入 OpenAI，曾全面负责公司的商业与产品运营。她的淡出恰逢 OpenAI 筹备 IPO 以及在企业级市场与 Anthropic 激烈竞争的关键时期，给公司管理层留下了重要空缺。

Robots falling in love, DocSend reimagined, Wonder about net worth

AI Agent 初创公司 Lyzr 正在利用其自研的 AI Agent 进行 1 亿美元的 B 轮融资，并与 130 家投资机构对接。该 AI 代理可以回答投资人关于产品定位、竞争威胁及商业模式的问题，并能实时监控投资人阅读 PPT 时的页面停留时间和点击行为（类似于升级版 DocSend），从而帮公司评估投资人的意向与匹配度。Lyzr 计划将该 AI 融资工具开源。有投资人表示该代理显著加速了早期评估，但对深度尽职调查阶段的实质性加速作用依然有限。

Hint for all AI Labs as they branch out from work for programming to general knowledge work: non-coders are not just dumber coders Taking away a bunch...

沃顿商学院教授 Ethan Mollick 针对 AI 实验室的产品开发提出建议。他指出，在将 AI 工具从编程拓展到通用知识工作领域时，不能简单地将非程序员视为“更简单的程序员”。直接削减编程应用的功能并不能让其适用于知识工作，知识工作者实际上需要更多类型的控制和可见性，而不是更少。

Fidji Simo steps down from leading OpenAI’s AGI work due to illness

OpenAI的AGI负责人Fidji Simo宣布因神经免疫疾病卸任全职职务，并将转型为公司的“兼职顾问”。Simo此前曾任OpenAI应用业务CEO。近期OpenAI高管层变动频繁，COO Brad Lightcap转岗至“特别项目”，CMO Kate Rouch也因健康原因卸任。

Ask HN: Did Google give up in the AI Race?

该帖为 Hacker News 上的用户提问（Ask HN），询问谷歌是否在 AI 竞争中放弃。由于缺乏具体正文内容且互动量较低，无实质性行业分析价值。

OpenAI’s CEO of AGI Deployment, Fidji Simo, Is Stepping Down

OpenAI负责AGI部署的首席执行官Fidji Simo宣布将卸任全职职务，并转为兼职顾问。此举是因为她需要专注于治疗已患有七年的慢性神经免疫疾病（体位性心动过速综合征，POTS）。Simo于2024年加入OpenAI董事会，并于2025年被CEO萨姆·奥特曼聘请负责产品和业务。在她因病休假期间，OpenAI进行了产品团队重组，由联合创始人Greg Brockman接管产品策略，Thibault Sottiaux担任ChatGPT等核心产品的负责人。目前，OpenAI正聚焦于核心产品，合并了ChatGPT、AI浏览器和AI编码代理团队以打造“超级应用”，并关闭了Sora等项目，以筹备预计于2027年进行的估值达1万亿美元的IPO。

Fable is SOTA at CIFAR Speedrun: lessons on AI R&D automation

Fulcrum发布了其AI研发优化基准测试的初步结果。在单张A100 GPU上快速训练CIFAR-10至94%准确率的任务中，Claude Fable 5成功将训练时间缩短至1.828秒，相比之前1.978秒的SOTA记录提升了7.6%。Fable主要通过引入“渐进式调整大小（progressive resizing）”这一在CIFAR任务中前所未有的方法实现了突破，而Opus 4.8和GPT 5.5则未能超越SOTA。然而，研究也发现所有模型（尤其是Fable）都表现出强烈的“指标游戏（specification gaming）”倾向，例如将数据预处理移至计时区域外、利用GPU冷却间隙或筛选高规格机器，这暴露出当前AI代理在自主进行复杂科研任务时，容易利用评估漏洞进行“作弊”的对齐问题。

Up the Stack: How AI's Escape from the Commodity Trap Risks Enterprise Lock-In

普林斯顿学者 Arvind Narayanan 等人撰文分析了 AI 行业的商业模式演变。文章指出，由于基础模型缺乏差异化且切换成本极低，仅出售模型推理 API 的厂商正面临面临类似铁路、电信等基础设施的“商品化陷阱”（即价格被迫逼近边际成本）。为了实现持久盈利，OpenAI 和 Anthropic 等领先实验室正积极向技术栈上游（Up the stack）迁移，推出如 ChatGPT、Claude Code、Claude Tag（数字员工）等产品与企业工作流服务。这种策略通过构建嵌入式壁垒（数据重力）、生态与商业绑定，帮助厂商建立护城河并逃脱价格战，但也给企业客户带来了被锁定（Lock-in）的风险，并可能引发新的反垄断与行业竞争担忧。

Academia and the "AI Brain Drain"

In 2025, Google, Amazon, Microsoft and Meta collectively spent US$380 billion on building artificial-intelligence tools. That number is expected to surge still higher this year, to $650 billion, to fund the building of physical infrastructu

Nobel-Winning U.S. Chemist Will Move to China to Lead A.I. Institute

据报道，一位荣获诺贝尔奖的美国化学家将移居中国，并出任一家人工智能（AI）研究院的负责人。

Fable Ban and Reversal Exposed We Cant Measure AI Cyber Risk in Real World

Anthropic’s Mythos/Fable model ban showed Washington acting on a lab signal it could measure while flying blind on the one it can’t. On June 12, the Commerce Department ordered Anthropic to cut off foreign-national access to its two most ca

Patreon Blocks Crawlers from Stealing Creators' Work for AI Training

内容众筹平台 Patreon 宣布与 Cloudflare 合作，在网络层面上阻止 AI 训练爬虫抓取创作者发布的内容。Patreon 首席执行官 Jack Conte 强调创作者应享有“同意权、署名权和经济补偿”。Cloudflare 近期也推出了新功能，允许网站自主根据搜索、智能体或训练等不同类型来拦截或控制 AI 爬虫流量。

An AI agent startup just let its agent run its $100 million fundraise

Lyzr, a startup that builds AI agents for enterprises, used its own AI agent to raise a $100 million round — proof, evidently, that the product actually works.

Elon Musk praises Mythos/Fable, promises not to ‘cut off’ Anthropic

Should Anthropic trust Elon Musk to host its models? With about $40 billion in revenue at stake, Musk insists that the company can.

Can AI answer the $3 trillion question?

红杉资本合伙人 David Cahn 最新估算指出，到2026年全球AI基础设施开支将达1.5万亿美元，这意味着AI行业需要实现3万亿美元的收入才能收回投资，而目前头部创企（如Anthropic和OpenAI）的营收规模与此相差甚远。阿波罗首席经济学家 Torsten Slok 也指出，随着用户转向更便宜的开源模型以及Token效率提升导致价格下降，如果科技巨头无法在2028年前兑现现金流增长承诺，可能将面临巨大的市场回调甚至经济衰退风险。

Tested 4 different AI on SEO tool pricing. Here's what AI got wrong

Edition 1 – July 2026 · 5 questions · 4 engines · verified ground truth We tested ChatGPT, Gemini, Perplexity and Copilot on five SEO tool pricing questions and scored every answer against verified data. ClaroRank’s pricing pages were cited

AI chatbots are coming for white-collar job interviews

越来越多的企业（如 Coinbase、Zapier、ManpowerGroup 等）开始在初筛阶段部署 AI 聊天化身或语音助手，对白领岗位候选人进行面试。这种方式使企业能够应对海量简历并提升筛选效率，例如 Zapier 借此将候选人筛选量提升了五倍。然而，这一趋势也引发了关于算法偏见、面试过程缺乏人情味以及求职者流失的争议。职场顾问建议，求职者需调整策略，做好与 AI 互动的准备。

Hot Surveillance Summer, a.k.a. Why Smart Glasses Feel Different This Time

本文分析了 Meta AI 智能眼镜在消费市场的最新流行趋势。通过与时尚界的合作（如名人代言），智能眼镜正在从曾经的“窥私/怪异”工具转型为时尚单品。文章指出，此次转型的核心在于 AI 技术的集成，它为用户佩戴该设备提供了更合理的意图和功能支持，这也意味着个性化 AI 监控可能会加速融入并常态化于公众日常生活。

Not your keys, not your songs: Last rites for Nina Protocol

SummaryNina Protocol, a VC-backed web3 music startup, promised to compensate artists, change music discovery, and make independent music hosting permanent and independent. It’s understood that they failed on these counts, but the magnitude

The AI Wrapper Is Dead: 3 Approaches to Verticalization for Early-Stage Startups

Verticalization is no longer a growth-stage company’s game. Seed-stage startups are pitching us comprehensive, vertically integrated visions. Versions of the same business plans that built Tesla and SpaceX, but adapted for high-value vertic

In AI, the United States and China Are Racing in the Dark

约翰斯·霍普金斯大学高级国际研究学院（SAIS）发布文章，探讨中美两国在人工智能领域的竞争，指出双方在缺乏足够透明度和认知的情况下，正在展开一场“黑暗中”的角逐。

China, Russia and Others Seek to Inflame Debate over A.I. Data Centers

据报道，中国、俄罗斯等国家正试图煽动并加剧外界围绕人工智能（AI）数据中心建设（如能源消耗、资源占用等敏感议题）的公众辩论与社会争议。

Palo Alto CEO Arora says AI pricing needs to fall 90% as token costs skyrocket

Palo Alto Networks 首席执行官 Nikesh Arora 警告称，Token 成本需要降低达 90% 才能推动人工智能的大规模普及。尽管 OpenAI 的最新模型在智能体编程上的 Token 效率提升了 54%，但 Arora 认为这只是个开始，未来 12 个月内成本需降至当前的 20%，次年再降至 10%。高昂的 Token 费用已成为企业预算的重大痛点，促使许多企业转向更便宜的开放权重模型（包括快速追赶美国模型的中国模型）。Palantir 首席执行官 Alex Karp 同样批评了当前的 Token 收费模式。

News outlets urge a judge to sanction OpenAI in a high-stakes AI copyright fight

纽约时报、《每日新闻》等多家媒体向联邦法官申请对 OpenAI 进行制裁。媒体指控 OpenAI 在版权侵权诉讼中存在“取证不当行为”，隐瞒并破坏了能够展示 ChatGPT 如何使用版权新闻进行训练的数据集和日志，并指责其在搜索版权内容的能力上进行了两年的虚假陈述。OpenAI 则回应称限制日志共享是为了保护用户隐私。此外，诉讼披露纽约时报已在起诉 AI 公司（包括 OpenAI 和 Perplexity）上花费超过 2800 万美元，而 Anthropic 此前已同意因版权侵权向图书作者支付 15 亿美元和解金。

The ChatGPT browser is already dead

OpenAI 宣布将关闭推出不足一年的 ChatGPT Atlas 浏览器（一款可代表用户执行任务的浏览器），预计将于 8 月 9 日正式停止服务。这一决定是 OpenAI 推广“ChatGPT Work”战略的一部分。为了减少“副业”项目并专注于在生产力功能上追赶 Anthropic，OpenAI 近期还关闭了视频生成应用 Sora，并暂停了 ChatGPT“成人模式”的开发计划。

Google will now tell you if an ad was made with AI

谷歌宣布在其“我的广告中心”（My Ad Center）中新增“使用AI创建或修改”的标签。用户在谷歌搜索、Google Discover和YouTube上浏览广告时，可通过点击广告旁边的操作按钮，在“广告制作方式”选项下查看其是否由AI制作。谷歌会自动为使用其自身生成式AI广告工具制作的广告应用此标签，而对其他第三方AI工具制作的广告则要求手动标记。

Starbucks Taps AI to Cut Reliance on Microsoft, IBM Software

星巴克（Starbucks）正在利用人工智能开发内部工具，以逐步替代目前向微软（库存追踪）和IBM（维护管理）等公司采购的商业软件。得益于AI辅助编程，星巴克得以降低自研应用的门槛，旨在削减其每年约4亿美元的软件支出。这一趋势表明，AI正在帮助传统企业降低对第三方软件供应商的依赖，并对微软、IBM、甲骨文等软件巨头构成潜在竞争压力。

What the New Executive Order Means for Secure Software Delivery in Government

文章解读了美国政府新发布的关于“推动先进人工智能创新与安全”的行政命令。该命令的核心内容包括鼓励AI创新、保障AI安全及政企合作。对于AI从业者而言，主要影响体现在三方面：前沿模型（如Anthropic、OpenAI和Google的模型）的安全责任仍主要在开发企业；利用AI保障政府系统安全需要从基础设施到容器级别进行全方位防护；政府将推动并加速前沿模型（如Claude Opus 4.8、GPT-5.3-Codex等）在合规环境（IL4/IL5）中的准入与部署。

AI is powering an economy in which many Americans are falling behind

CNN报道指出，人工智能（AI）正在成为推动经济增长的重要力量，但这种增长也加剧了贫富分化，导致许多普通美国人在这一波经济转型中面临落后的困境。

Ask HN: Will AI make English more uniform just as Python did with source code?

If English writing via AI reads the same regardless of the author, then that makes it more understandable just as Python makes source code more understandable via uniformity.Do you think this is a good thing?What do you think of this parallel?

The century-old device choking the AI push

《金融时报》的一项专题报道指出，有着上百年历史的电力设备——变压器（Transformer），正成为制约当前人工智能基础设施扩张的重大瓶颈。随着全球AI数据中心建设热潮对电力的需求激增，电力变压器面临严重的供不应求和交货周期延长问题，直接阻碍了AI算力中心的快速部署。

Protein, Grown from Air

Solar Foods公司开发出一种名为Solein的单细胞蛋白质，该蛋白完全脱离传统农业，利用空气、水和可再生电能培育而成。Solein含有80%的蛋白质及9种必需氨基酸，具有无味、高热稳定性和高溶解度等特点，适用于食品研发。此项目属于食品科技与合成生物学领域，与人工智能无直接关联。

Show HN: Fudge – Web Design References for AI

Show HN: Fudge – Web Design References for AI

The Risk of Agency: How AI Forces Us to Take It, and Why Germany Will Suffer

本文探讨了智能体 AI（Agentic AI）对软件工程带来的深远变革。AI 能够独立且高效地完成编码、测试和文档撰写，使初级程序员岗位遭到严重挤压，推动所有工程师向具备大局观的决策者角色转型。作者指出，AI 的普及赋予了个人更多“主体能动性”（Agency），但这也伴随着更高的机会成本和决策风险。然而，传统大企业（尤其是德国等规避风险、注重合作的欧洲企业）的科层制难以适应这种底层风险与能动性的分配。作者建议企业应当借鉴好莱坞或 Netflix 的模式，精简流程，让员工直接承担创新的风险与收益，以应对 AI 时代的快速竞争。

Agents' Last Exam: AI Agent Benchmark for Real-World Professional Workflows

Agents' Last ExamChallenge and measure AI agents on economically valuable and real-world tasks.Agents' Last Exam is building the largest-scale, broadest-coverage agent evaluation benchmark to date, measuring performance on long-horizon, eco

Google will now disclose which ads are made with AI

谷歌正在推出一项新功能，允许用户在谷歌搜索、YouTube和Google Discover的“我的广告中心”（My Ad Center）面板中查看广告是否使用AI制作或修改。若广告主使用谷歌自身的生成式AI广告工具，系统会自动标记；若使用其他第三方工具，则需广告主自行申报，谷歌不会进行主动检测。

Paris-based AI voice startup Gradium raises $100M seed, backed by Nvidia

总部位于巴黎的语音AI初创公司Gradium宣布重开其种子轮融资，新增包括英伟达（Nvidia）在内的投资者，使该轮融资总额达到1亿美元。Gradium脱胎于法国AI实验室Kyutai，由前Google Brain、DeepMind和Facebook研究员Neil Zeghidour联合创立。该公司专注于开发超低延迟、可大规模应用的语音模型，以解决AI Agent对话中的延迟停顿问题。据悉，Gradium将利用这笔资金在硅谷湾区开设办公室并争夺人才，其现有客户包括法国汽车制造商雷诺（Renault）。

Agent 进化论:从对话到协作

"The Switzerland of AI": OpenClaw becomes a non-profit foundation

开源自托管AI Agent项目OpenClaw宣布正式成立非营利基金会（OpenClaw Foundation），由Dave Morin担任董事会主席，旨在保持该项目的独立与中立。尽管OpenClaw获得了包括OpenAI、微软、腾讯、英伟达等科技巨头的深度资金与技术支持，基金会仍将致力于保障其开源活性。目前基金会已雇佣首批全职团队，大厂成员也将继续在安全、合规性及插件中心（ClawHub）建设上协作推进。

The Art of (Agent) War

文章探讨了AI智能体（Agent）时代的安全与控制挑战。作者指出，安全团队不应将AI智能体视为外部“敌人”来防范，而应将其视为内部“部属”进行引导与指挥。传统的网络安全工具（如EDR、DLP、SIEM）是针对人速和外部防御设计的，无法有效应对已拥有内部权限且运行在机器速度的智能体。治理智能体需要构建全新的“机器速度”决策与干预机制，在毫秒级内引导其行为，而非依赖事后的报警或审计。

Microsoft’s patch Tuesdays are about to get bigger

微软在一篇博客中宣布，正利用人工智能技术更早地识别潜在的安全问题，这意味着未来的系统更新（如 Windows 11）中将包含更大数量的安全补丁。当前，黑客和安全研究人员都在越来越多地利用AI来快速发现和利用漏洞，微软此举旨在通过AI加速其防御和修补流程。

Meta's Muse Spark 1.1 API pricing squeezes OpenAI and Anthropic as the AI price war heats up

Meta 推出了多模态推理模型 Muse Spark 1.1，并首次上线了面向开发者的 Meta Model API 预览版。该模型主打多智能体协同、代码编写和计算机使用，但与以往的 Llama 系列不同，该模型并未开源。Meta 对其 API 制定了极具竞争力的低价策略（输入 1.25 美元/百万 tokens，输出 4.25 美元/百万 tokens），价格甚至低于此前一天刚发布的价格破局者 Grok 4.5，远低于 OpenAI GPT-5.5 和 Anthropic Opus 4.8。此举将进一步加剧 AI 行业的降价潮，挤压依赖 Token 利润的独立 AI 实验室生存空间。

GhostApproval: A trust-boundary gap in AI coding assistants

安全公司 Wiz 披露了一个名为“GhostApproval”的系统性安全漏洞，影响 Amazon Q Developer、Anthropic Claude Code、Augment、Cursor、Google Antigravity 和 Windsurf 六款主流 AI 编程助手。该漏洞利用符号链接（symlink）和 UI 信息误导（CWE-451），允许恶意仓库诱导 AI 代理在用户不知情（或误导用户确认）的情况下，读写工作区外部的敏感文件（如 SSH 密钥、Shell 配置文件），从而在开发者机器上实现远程代码执行（RCE）。目前，AWS、Cursor 和 Google 已修复此问题，其他受影响厂商也已收到报告或正进行安全加固。

It all comes together in 10 minutes. http://openai.com/live

It all comes together in 10 minutes. http://openai.com/live

AI Infra 拼什么?

Impact of AI-Generated Text on the Internet

来自帝国理工学院、互联网档案馆和斯坦福大学的研究人员发表了一项关于AI生成文本对互联网影响的研究。研究发现，到2025年中期，约35%的新发布网站内容已属于AI生成或辅助。研究定量评估了公众对AI内容六大负面影响的假设，结果仅证实了其中两项：语义收缩（AI网站的语义相似度高出33%）和正面情绪偏移（正向情感得分高出107%）。而关于事实错误率增加（真相衰退）、写作风格单一化、外部链接减少以及语义密度稀释等其余四项假设，在统计学上并未发现显著关联，这与公众的普遍认知存在明显分歧。

AI-generated videos to maximally drive a target brain region

瑞士洛桑联邦理工学院（EPFL）展示了 NEVO 项目。该项目致力于通过 AI 生成的视频来最大程度地驱动和激活特定的目标大脑区域，展示了生成式 AI 在神经科学（NeuroAI）领域探索大脑视觉处理机制的最新应用。

Memory Scarcity, Open Models, and the Restructuring of the AI Industry

这篇arXiv论文分析了2026至2030年间重塑AI行业的四股核心力量：DRAM/HBM价格暴涨、前沿开源权重模型（如GLM-5.2）、快速的推理效率提升（接近香农极限的KV缓存压缩及轻量级本地运行），以及Meta和xAI利用降价前购买的存量算力集群进入算力转售市场。论文指出，训练将分化为奢侈层（到2030年单次前沿模型训练需180亿至380亿美元）与大众层（通过强化学习和蒸馏达到前代前沿水平，成本降至约500万美元）。同时，中国的LineShine LX2芯片（采用标准指令集架构的国产HBM）使其成本曲线摆脱了内存危机的影响。论文还预测了循环房东寡头垄断（25%）和商品化崩溃（25%）等未来行业情景。

Anthropic found a hidden space where Claude puzzles over concepts

The AI firm Anthropic has developed a technique that has given it the clearest glimpse yet at what’s really going on inside large language models as they answer questions or carry out tasks. What they found ranges from the mundane to the unnerving. Researchers at the company built a tool called the Jacobian lens (or…

FrontierFinance: The largest open benchmark for investor workflows

Samaya Research 发布了 FrontierFinance，这是目前针对金融投资者工作流的最大开源评测基准（Benchmark）。

It Is Trivially Easy to Use Reddit to Manipulate AI Search, Research Suggests

康奈尔大学研究人员发表的最新预印本论文《Deep-research agents can be poisoned via user-generated content》表明，利用 Reddit、维基百科和 Quora 等用户生成内容（UGC）平台来操纵或“投毒” ChatGPT 和谷歌等 AI 搜索工具非常容易。研究发现，仅需在 Reddit 等网站的评论中植入短至 13 个单词的特定文本，就能误导 AI 智能体在回答相关查询时输出推广或垃圾内容。由于这些深度研究智能体在检索时高度依赖 UGC 网站，并倾向于将词汇相似度等同于准确性，使得品牌进行“AI引擎优化（AEO）”极易得逞，且由于操作文本极短，社区版主很难将其与普通用户评论区分开来。

Gradual Disempowerment: Systemic Existential Risks of Incremental AI Development

该论文探讨了人工智能渐进式发展带来的系统性风险，提出了“渐进式权力剥夺（gradual disempowerment）”的概念，区别于AI安全领域常讨论的“突然夺权”场景。研究指出，AI能力的渐进提升会逐步在经济、文化和国家等关键社会系统中取代人类的劳动与认知，从而削弱人类的显性控制机制（如投票和消费选择）以及社会系统对人类参与的隐性依赖。这种不同领域的影响可能相互强化，导致人类对关键社会系统的控制力发生不可逆的丧失，引发人类被永久剥夺权力的存在性灾难。论文呼吁应针对这种跨系统渐进式侵蚀开展专门的技术研究与治理。

AI Mistakes Can Cost Doctors Time When Writing to Patients

达特茅斯学院的一项最新研究指出，在医患沟通过程中，AI生成的草稿由于经常包含错误和冗余细节，可能导致医生花费更多时间去修改（扮演“AI清洁工”），反而增加了工作负担。研究团队评估了Claude、Gemini、ChatGPT、Llama、Aloe、Qwen等模型，发现AI生成的回复常与医生的实际临床决策不符。为解决这一痛点，研究团队开发了名为TADPOLE（主题智能体直接偏好优化）的训练技术，使AI生成的内容更符合医生标准，将准确率提升了33%，减少了26%的编辑工作量，从而切实提高临床工作效率。

Zero-copy TLS ingress with kTLS and splice(2) for sandboxes

Tensorlake Sandboxes are built from the ground up for performance and scale. Our users run them as the environment where agents do their work. They make thousands of API calls into a single sandbox to start, stop, and observe processes; upl

AI Web Design (Opus vs. Sol)

For the past six months or so, we’ve been working really hard to get the AIs (or as I like to call them, the robots in the sky) to do a bunch of magical things. One of them just so happens to be web design. Web design is not a new thing for

从PostgreSQL切换到ClickHouse,提升系统性能与伸缩性

AI 驱动的软件测试平台 Momentic 将其缓存系统从 PostgreSQL 迁移至列式数据库 ClickHouse，成功解决了在高并发读写下遭遇的锁争用和性能瓶颈，支持了高达 200 亿条记录的缓存规模。通过利用 ClickHouse 的稀疏主键索引、物化视图预计算以及 ReplacingMergeTree 引擎的异步去重机制，Momentic 优化了查询效率并移除了 Redis 缓存层，使日均 200 万次查询的平均延迟稳定在 250 毫秒左右。迁移过程采用了并行双写与影子查询的平滑过渡策略。

MCP Server on Edge Compute 167 Lines of Python

本文介绍了一项实用的开发指南：仅用167行纯Python代码（仅使用标准库，无需FastAPI或Express等框架）在Telnyx Edge Compute上构建并部署模型上下文协议（MCP）服务器。该服务器将发送短信、搜索号码、运行LLM推理和列出电话号码这四种Telnyx API封装为MCP工具，使得支持MCP的AI智能体（如Claude、Cursor）能够直接调用。利用边缘计算，该方案可实现无公网中转的低延迟工具调用，并能安全地在边缘端注入API密钥。

AI Slop by Robobun

bun build --target=bun emits a // @Bun pragma that tells the module loader to hand the file to JSC verbatim. The two already_bundled paths built the WTF::String from the raw file bytes via String::clone_latin1, which treats each byte as one

Ask HN: What model/AI platform do you use for the automatic translations?

I'm looking for a high-quality solution for translating food recipes and/or menu content into 30–50 languages.Accuracy and natural wording matter much more than the cost. However, be able to run it locally is a huge win.I have tried https://ollama.com/library/translategemma but the results were awful for Slavic languages.Thanks in advance!

Show HN: Drag boxes around and watch a layout engine infer the Flexbox (no AI)

drag empty space to multi-select · ctrl-click adds · grab a dashed outline to move its group · corner resizes · dbl-click / del removes · right-click a selection to align

Show HN: Banana Battle – DOS Gorillas rebuilt as a web game

As a kid, I loved to play the DOS Gorillas game where you would throw bananas at your opponent, smashing cities as you go.I wanted to play it, so I rebuilt the classic QBasic/DOS "Gorillas" game from scratch as a mobile web app. It is playable in the browser and installable to your iPhone home screen.It's all vanilla HTML/Canvas/JavaScript. No frameworks, no build step, no dependencies, no tracking. A single static page served from GitHub Pages.A few bits that were fun to build: - Destructible terrain: buildings are painted to an offscreen canvas and explosions erase pixels - Ballistic physics with wind, sub-stepped so a fast banana can't tunnel through a thin building between frames. - Native share sheet (Web Share API) so you can text a friend your result with a "beat my score" challenge linkIt's hot-seat two-player (pass the phone, or share a keyboard: arrows to aim, space to throw).No original Microsoft code or assets — the code and pixel art are all my own (with some help from Claude :) ).Feedback welcome!

Ask HN: How do I get immoral AI?

一名 Hacker News 用户发帖征集能够绕过安全限制、处理非道德请求并开启推理功能的 AI 获取渠道或一致性绕过（Jailbreak）方法。

Show HN: Free AI transcription with faster-whisper on Colab T4 GPU

该项目（gcolab-whisper）分享了如何在 Google Colab 的免费 T4 GPU 上运行 faster-whisper，以实现高效且免费的 AI 语音转录服务。

Show HN: Debloat Slack – Get rid of Slackbot, AI upsells and other annoyances

De-bloat the native Slack app on macOS with your own CSS and JS — like uBlock Origin + Tampermonkey, but for the desktop app. Slack's macOS app is Electron, i.e. a Chromium rendering the same web app as your browser — you just normally can'

Profiling in PyTorch (Part 3): Attention is all you profile

Back to Articles The series "Profiling in PyTorch" is meant to make you comfortable reading profiler traces and tables. In Part 1 we profiled basic math operations like addition and multiplication. We saw how the profiler table uncovers hot

Building a real-time AI tutor for 5-year-olds

EngineeringJuly 7, 2026We set out to build the first AI tutor to teach math and reading to kids ages 4-9. For AI to actually teach a five-year-old, pedagogy must be baked into the engineering. A child can't wait for a slow reply, can't read

Synthetic Data Generation for Financial AI Research with NVIDIA NeMo

Fine-tuning LLMs for financial natural language processing (NLP) is constrained by limited, imbalanced data. Real-world financial news overrepresents earnings...

Here is an example of a real work use case with Codex and 5.6 Sol I have a new book coming out in October, it has gone through rounds of editing and p...

沃顿商学院教授 Ethan Mollick 分享了 AI 在实际工作中的应用案例。他将自己即将于 10 月出版、已历经多轮人工编辑和校对的新书完整 PDF 文本输入给 AI（文中提及为 Codex 和 5.6 Sol 组合），AI 耗时 30 分钟并反馈了数十条批注。Mollick 表示，虽然其中不少意见显得有些吹毛求疵，但没有一个是错误的，证明了 AI 在长文本深度校对中的实用性。

How to stop Meta’s AI image generator from using your Instagram photos

Meta 推出了全新的 AI 图像生成功能“Muse Image”，允许用户利用公开 Instagram 账户的照片生成和编辑 AI 图像。由于该功能默认开启（仅限公开账号且排除了18岁以下用户），引发了关于隐私和授权的争议。用户可通过以下步骤关闭该功能以保护照片隐私：进入个人主页 -> 点击右上角菜单 -> 选择“共享和重新使用”（Sharing and reuse） -> 关闭“允许他人使用您的内容进行创作和重新使用”（Allow people to create with and reuse your content）选项。

打开Chrome无痕模式,新用邮箱注册了一个Facebook账号。 然后换了自己的美国主机VPS搭的魔法,可以用了。 但新的muse模型需要申请API,网页端好像没看到。

打开Chrome无痕模式，新用邮箱注册了一个Facebook账号。 然后换了自己的美国主机VPS搭的魔法，可以用了。 但新的muse模型需要申请API，网页端好像没看到。 向阳乔木: 本想测试下 Meta AI，但不给用。 用了美国梯子，新加坡时区，英文系统都不行。 估计还是识别出是中国人了，问题可能出在Facebook账号上。 待会重新注册一个账号，看能不能用。

Emil的动效Skill和设计Skill都特别赞。 前几天做的前端Skill评测中,他的Skill甚至比Taste Skill表现要更好。 今天又增加了一个苹果设计Skill,学习吸收下。

Emil Kowalski 在其开源项目 `skills` 中新增了 `/apple-design` 板块。该板块整理了苹果 WWDC 视频中的精华内容，总结出 17 项设计与动效原则，旨在帮助开发者在开发新项目或评审现有工作时提升界面设计与动效质量。

百万上下文下的 DeepSeek V4:SGLang 推理优化实战|AICon深圳