摘要 随着 Anthropic 开源 skills 仓库，"Code Interpreter"（代码解释器）模式成为 Agent 开发的热门方向。许多开发者试图采取激进路线：赋予 LLM 联网和 Python 执行权限，让其现场编写代码来解决一切问题 ...

Architecture ...

A hacker targeted a white supremacist dating website, lured users with an AI chatbot, and deleted the platform entirely live on stage.

好朋友一泽在Agent Skills 终极指南：入门、精通、预测上也提到了把 Skills 视作“通用 Agent 的扩展包”，并强调它的核心价值在于“人给指引，Agent 看着执行”，从而让垂直 Agent 的成本大幅降低。

If you have ever tried crunching large datasets on your laptop, maybe a big CSV converted to NumPy or some scientific data from work, you have probably heard your laptop fan roar like it is about to ...

每次让 AI 帮你写周报，都要重复解释一遍格式要求。每次让它帮你改代码，都要再说一遍"按照我们团队的规范来"。 Skills 就是给 Claude 的"建图"过程。你告诉它写公众号文章要用什么风格、做代码审查要检查哪些点，它记住了，以后自动按这个来。

如果你让AI随便生成Bug，它大概率会产生幻觉，为此SSR设计了一套如同安检般严格的一致性验证（Consistency Verification）流程。 掩盖有效性：应用了「掩盖补丁」后，原本失败的测试必须变通过，证明成功欺骗了测试套件。

Five years, one artist, one robot: how Maxim Gehricke made SEN, a 3D animated short film created solo from concept to final ...

Meta FAIR团队联合UIUC和CMU研究人员提出Self-play SWE-RL系统，让AI通过自我对弈学习编写和修复代码。

This important study introduces a new biology-informed strategy for deep learning models aiming to predict mutational effects in antibody sequences. It provides solid evidence that separating ...