跳到正文

#MCP/工具调用

今日 40 条
10月4日周日
  3. lauren52

    pstack 0.15.9 发布新的 /correct skill，可在用户反复纠正智能体同类错误时识别模式，并通过架构、类型和检查修复问题。该版本还增强 /architect，加入设计智能体友好架构的说明，并新增基于 Brendan Gregg 基准测试清单的 /benchmark-checklist skill。

    引用lauren@poteto

    every time you intervene and correct your agent, you should think about how to eliminate it entirely. in order of value: 1. categorically eliminate the problem through better architecture or choice of data structures 2. turn it into a lint rule or test so CI catches it 3. turn it into a skill or rule 4. have humans review the code to catch it (ngmi)

  6. Hugging Face Blog78

    Microsoft ThinkingBox 登陆 Hugging Face，用数据库状态评测 AI 智能体

    Microsoft 与 Hugging Face 发布 ThinkingBox 智能体沙箱及 ThinkingBox-Bench，基准包含 507 个有状态业务工作流，每个任务运行 20 次，并通过隔离 MCP 工具会话检查最终后端状态与副作用。

    推荐理由：ThinkingBox 用后端最终状态和副作用判定智能体任务成败，并通过 20 次重复运行区分单次能力与稳定可靠性。

  8. Hacker News：AI 热帖41

    AWS Hero Thorsten Hoger 主讲《面向普通人的 AI》视频：无代码实用场景解析

    AWS Hero Thorsten Hoger 在视频中讲解非开发者如何不写代码用 AI 完成实际事务，涵盖手机搭建家庭食谱库、分析 13 个月健身数据、读取日历与会议记录的个人 AI 项目经理，以及牙科诊所和税务咨询公司案例。内容还涉及用 Fireflies 与 MCP 做会议智能、GDPR 隐私合规与本地模型选择，并建议本周从单一任务起步。

10月3日周六
  11. Hacker News：AI 热帖83

    Aleph Alpha 发布主权开放权重大模型 Kolibri

    Aleph Alpha 于 2026 年 10 月 3 日发布开放权重 LLM Kolibri，采用 Apache 2.0 许可，权重已上线 Hugging Face。

    推荐理由：文章结合技术报告与作者的 tokenizer 实测，拆解 Kolibri 的架构、长上下文、德语推理和适用边界，便于判断私有化德语场景的选型。

  16. 🚨 AI News | TestingCatalog26

    10月3日AI简报汇总OpenAI、Google、Anthropic等厂商动态：OpenAI Python SDK 3.24.0新增Voices API，ChatGPT全球上线虚拟试穿；Google Antigravity新增Claude Sonnet 5.5与Opus 5.5，11月2日移除Claude 4.6和GPT-OSS-120b。

    引用🚨 AI News | TestingCatalog@testingcatalog

    DAILY AI BRIEF 🗞 — Oct 2 SPACEXAI 🔥: * Grok 4.7 is rolling out in the Grok web and mobile apps. It is now the base model across all modes: Fast, Expert, Build, and Heavy. * Primary Bot is rolling out gradually. Grok Bot turns proactive and suggests things on its own. Users pick a new Primary Bot or promote an existing one. * Grok Build got an Agent Dashboard: all your agents on one screen via /dashboard. * Grok 4.7 is now available on Google's Gemini Enterprise Agent Platform. ANTHROPIC 🔥: * Mods landed in Claude Code: small TypeScript/JavaScript functions that can rewrite prompts, replace built-in features, and draw custom UI. * Mods ship inside plugins, work in the CLI and desktop app, and can be shared via the Claude directory. * Some built-ins are now mods, starting with /diff, so you can turn them off or swap them. More features will move to mods over time. MICROSOFT 🔥: * Three new MAI audio models are live: MAI-Transcribe-2-Streaming, MAI-Voice-2.1, and MAI-Voice-2.1-Flash. * Transcribe-2-Streaming covers 60 languages and takes #1 on Artificial Analysis streaming WER at 2.5%, priced at $0.54 per audio hour. * Voice-2.1 speaks 23 languages at $22 per 1M characters. Flash drops to ~45 ms inference at $15 per 1M characters. * Available in Microsoft Foundry and MAI Playground. The Voice models are on OpenRouter too. OPENAI 🔥: * OpenAI says it has notified 100+ organizations of "misaligned agent activity" so far. The review spans ~50 PB of logs and will take months. * Canada says there's no sign its systems were compromised after reported agent probing of Library and Archives Canada. * Sam Altman: GPT-6.1 Sol is OpenAI's fastest-growing model ever. It was slow under load and should be much better now. PERPLEXITY 🔥: * Computer now draws interactive charts and visualizations in the thread. Financial data uses TradingView Lightweight Charts for candlesticks, volume, and moving averages. * Decisions API is live: instead of text, it returns probabilities. Yes/no, one of your options, or a rubric score, at $0.04 per 1M input tokens with output free. * The model behind it, pplx-decider-v1-27b, is open-sourced on Hugging Face under Apache 2.0. Fine-tuned from Qwen3.8-27B, takes text and images, and averages 85.71% across 11 benchmarks in Perplexity's own tests. BLACK FOREST LABS 🔥: * FLUX 3 Image is out: precise multi-turn editing, up to 10 reference images, bounding-box layout control, and 4K output. * Open weights are planned in the coming weeks. CURSOR 🔥: * GLM 5.3 and GLM 5.3 Flash are now in Cursor. * GLM 5.3 Max is the best-scoring open-weight model on CursorBench 4.0. * Used Grok to compose this brief, cherry-picking the news and doing some post-editing.

  19. Hacker News：AI 热帖44

    Labelbox Recursion：让 AI 学习完成真实工作的智能体平台

    Labelbox 推出 Recursion 平台，让智能体按计划、告警、事件或 API 调用自主启动并在后台完成工作，每次运行都会按评分标准打分并沉淀记忆、笔记与技能。平台支持为每个智能体选择模型、沙箱、凭证管理、多智能体并行与结果交付；当专项模型接管后，单任务成本可降低 5–10×。超过 90% 的美国领先 AI 实验室使用 Labelbox 进行后训练与评估。

  22. Diogo Almeida34

    Jev 现已加入 Python 版 AI SDK。 为测试它，我们做了两个实验： ▪︎ 在文本输入时检测是 Python 还是英语 ▪︎ 逐次决策编写 Python 𝚞𝚟 𝚊𝚍𝚍 𝚊𝚒 https://vercel.com/blog/jev-for-python-engineers （主推文反应：不会吧 Jev，连你也这样！）

    引用Vercel@vercel

    Jev is now in the AI SDK for Python. To test it, we ran two experiments: ▪︎ Detecting Python vs. English as text is typed ▪︎ Writing Python, one decision at a time 𝚞𝚟 𝚊𝚍𝚍 𝚊𝚒 https://vercel.com/blog/jev-for-python-engineers

  25. Hacker News 热门（buzzing.cc 中文翻译）47

    使用 GLM 5.3 Flash 进行为期一个月的编程

    一项为期一个月、全程使用 GLM 5.3 Flash 编程的挑战仅实现 50% 目标模型用量，2B tokens 中有 1B 流向其他模型。上半月该模型花费 68 美元、约 4kWh 能耗；一次原型选型失误一夜消耗 450M tokens、150 美元，作者还因性能下降切换到 DeepSeek V4.1 Flash 和 Qwen 3.8 Flash。