Keep on to blur preview images; turn off to show them clearly
RT @Jimmy_JingLv: 🍌 Nano Banana 一致性就是生产力啊！ 绝了，这是一致性的绝妙 idea —— X-Ray （见视频） 有了想法，先把 https://t.co/Cexnm00Ayi 里面先构建原型 然后下载 zip 包拖进 Cursor 改造一…
🚧 building https://t.co/AJfZ3LMlgq https://t.co/606cFUoda3 https://t.co/s0m0tpQMDH https://t.co/UQ5vrrYdAG 🐣learning/earning while helping others ❤️making software, storytelling videos 🔙alibaba @thoughtworks
GitHub Copilot 如何通过「精简工具集」变得更智能 开发者都会遇到的痛点：工具过多导致响应迟缓和决策低效。通过引入“虚拟工具”、嵌入引导路由和自适应聚类等创新，Github Copilot 的智能体能在保持强大功能的同时，显著提升速度和准确性。 核心理念：少即是多，智能体需精炼工具 GitHub Copilot Chat 依赖数百个工具（如代码库分析、Azure 服务调用）来辅助开发者完成任务，例如修复 bug 或合并代码。这些工具通过 MCP 访问，但问题在于：工具堆积过多会让智能体“负担过重”，类似于大脑被无关信息淹没，导致推理变慢、错误率上升。基准测试（如 SWE-Lancer 和 SWEbench-Verified）显示，完整工具集下智能体的任务成功率反而下降 2-5 个百分点，因为模型容易误用工具或忽略关键指令。 解决方案的核心是“用更少的工具变得更聪明”：不是简单裁剪功能，而是通过智能路由和分组，让智能体只在需要时调用相关工具。这就好比从杂乱的工具箱中抽屉化管理——先看目录，再取具体物品，避免盲目翻找。 技术实现：嵌入引导与动态选择 更新引入了两大关键机制，确保工具选择精准高效： · 嵌入引导工具路由（Embedding-Guided Tool Routing）：利用查询的向量嵌入与工具的语义表示进行匹配，预先筛选出最相关的工具候选。这比传统 LLM 逐一评估快得多。在基准测试中，该方法实现了 94.5% 的工具使用覆盖率，远高于 LLM 选择的 87.5% 或静态列表的 69.0%。例如，对于“修复这个 bug 并合并到 dev 分支”的查询，系统会直接从嵌入空间中锁定“合并工具”，跳过无关的搜索或文档工具，减少了探索性调用。 · 自适应工具聚类（Adaptive Tool Clustering）：基于 Copilot 内部嵌入模型，通过余弦相似度将相似工具自动分组，形成“虚拟工具”——这些虚拟工具像目录一样，提供概述而非完整列表。聚类后，一个小型模型生成每个组的摘要，便于缓存和快速访问。博客展示了 GitHub MCP 工具的嵌入图示：如 create_pending_pull_request_review 与 get_issue_comments 等工具自然聚为一簇。 此外，GitHub 将默认的 40 个内置工具精简至 13 个核心工具（覆盖仓库解析、文件编辑、搜索和终端操作），其余非核心工具归入四个虚拟类别：Jupyter Notebook 工具、网络交互工具、VS Code 工作区工具和测试工具。这种“无损动态选择”确保了功能完整性，同时将首 token 时间缩短 190 毫秒，最终响应延迟平均降低 400 毫秒。 益处：更快、更准的用户体验 · 性能跃升：在线 A/B 测试显示，任务成功率提升 2-5 个百分点，工具覆盖率提高 27.5%。智能体能更专注地推理，减少缓存未命中和 API 限额问题。 · 效率优化：操作成本降低（缓存嵌入和摘要更廉价），开发者感受到更流畅的交互——无需等待“加载中”转圈。 · 实际示例：在处理复杂查询时，系统能从历史上下文推断意图，避免逐一检查工具组，提升了整体可靠性。 未来展望：向长上下文智能体演进 将工具选择视为“长上下文推理”的雏形：未来，智能体将记住工具使用历史、从对话中推断意图，并规划多步行动，甚至跨会话协作。结合嵌入、记忆机制和强化学习，Copilot 可能扩展到数千轮交互，支持动态学习工具使用。 这个更新体现了 AI 开发工具的演进趋势：从“全能”向“专注”转型，GitHub 通过数据驱动的优化证明，精简并非妥协，而是通往更强大智能的捷径。 博客地址：
邵猛，中年失业程序员 😂 专注 - Context Engineering, AI Agents. 分享 - AI papers, apps and OSS. ex Microsoft MVP 合作 - 私信/邮箱：shaomeng@outlook.com 📢 公众号/小红书: AI 启蒙小伙伴
Sherjil Ozair claim to fame started with the paper "Generative Adversarial Networks" that arguably launched the generative media revolution (AI images, video etc.). That paper practically taught neural networks how to 'imagine'. Sherjil Ozair was visiting Universite de Montreal from Indian Institute of Technology Delhi and got to work with the lead author Ian Goodfellow, and Turing Award winner Yoshua Bengio.(https://t.co/LhtYzhBYGL). He hasn't looked back since. Luck + Talent. His startup 'General Agents' was aquired by Jeff Bezo's Project Prometheus where he has already been a co-founder since June 2025. It would be interesting to know why Bezos may interested in their technology. Here is the reason: When you let AI operate your computer with GUI input - it results in very long inputs to AI model. It costs a lot and it is slow. His startup "General Agents" uses a different approach that allows them to rapidly understand what is on the screen and decide how to operate inputs (e.g. Mouse Click). They do it with light weight Vision Language Action (VLAs) models, possibly. It is also likely they have two models, one for reasoning and one for figure out exactly where to place cursor. You can see a demo here - it is real speed. https://t.co/e6On4WAIEQ Such models can be used much beyond computer use for web browsing, for example for operating SCADA software and Engineering Design software. While computer use is not new or unique, their speed is incredible. Qwen's VLA models can do that as well, though not as fast. Sherjil Ozair has deep experience from DeepMind and Tesla Autopilot projects. And his papers have 100K+ citation. A worthy co-founder for project Prometheus already. He can attract lot of top tier talent and help in vetting them. Trying to beat @deedydas here. I will publish an MCP server called "Tweet like Deedy". Will ask Claude Code to add functionality to look up college and schools and hype them up as well! 😂 No disrespect - just being silly, Deedy. Big fan of your advocacy. On serious note - I am afraid all of the folks who work on computer all day are going to be replaced by an MCP one day. If you don't believe me reverse engineer Claude code and figure out how Slash Command, Skills and SubAgents work. Everything is a .md file & text description. Bash is all you need. Happy Thanks Giving!
Artificial Intelligence @amazon. All views personal!
RT @MengdiWang10: I was an ICLR PC two years ago. Woke up every morning to hundreds of unread conference-related emails. It’s simply not su…
Asst professor @MIT EECS & CSAIL (@nlp_mit). Author of https://t.co/VgyLxl0oa1 and https://t.co/ZZaSzaRaZ7 (@DSPyOSS). Prev: CS PhD @StanfordNLP. Research @Databricks.
RT @indie_maker_fox: Discord群即将开奖，推特活动则月底开奖 MkSaaS模板的优惠活动价格将持续到月底 下个月将恢复原价，感兴趣的朋友请尽快下单 新站MkDollar开发中，MkSaaS客户后续权益多多
🔥 The best AI SaaS boilerplate - https://t.co/VyNtTs0jSX 🚀 The best directory boilerplate with AI - https://t.co/wEvJ1Dd8aR 🎉 https://t.co/bh1RxeERuY & https://t.co/zubXJCoY92 & https://t.co/tfQf8T7gGF
Deepseek发布了模型DeepSeekMath-V2 在IMO-ProofBench和竞赛（如IMO 2025的5/6问题）以及Putnam 2024（接近完美的118/120分）中表现出色。拥有国际数学奥林匹克竞赛（IMO）2025金牌水平 Github开源链接：https://t.co/5vhdwpQW6G 该模型也在 @huggingface 上以 Apache 2.0 开源协议发布！ 也可以从HF下载：
Believing is seeing
