这项由ETH苏黎世联邦理工学院、INSAIT索菲亚大学、加州大学伯克利分校共同完成的研究，于2026年7月以预印本形式发布，论文编号为arXiv:2607.13921，发表在计算机编程语言领域（cs.PL）。感兴趣的读者可通过该编号在arXiv上查阅完整论文。 一、当AI写代码遇上"亡羊补牢"的困境 程序员们有一个共同的经历：写了几百行代码，满心期待地按下编译按钮，结果屏幕上喷出几十条红色报错信息 ...

2026年秋天的科技圈，空气里飘着一股‘硅基咖啡’的味道——不是星巴克新出的AI拉花拿铁，而是微软即将在旧金山点燃的一场端侧AI核爆。10月7日，微软将在旧金山莫斯克尼中心举办一场神秘感拉满的特别活动。没有预告片，没有倒计时海报，只有三行克制到近乎高冷的官宣：Windows、RTXSpark、Surface。但懂行的人都知道——这哪是发布会？这分明是一封写给全球AI开发者的‘本地算力情书’。 别急 ...

给定一个基座模型、一个目标能力和算力预算，RSI Agent 自己设计后训练算法、收集和整理数据、写代码、跑实验、迭代改进，最终交出一个训练好的模型，全程无人介入。在自主后训练基准 PostTrainBench 上，它以 46.6 分排名第一，超过所有前沿模型与 Agent 系统，接近人类专家表现（51.1）。近日，AIBuildAI 团队发布了 PostTrain Agent，一个用于自主后训练 ...

I opened up a log-formatting script I wrote myself for the first time in a while. It's a small Python script I wrote about a ...

H-Elena, a Falcon-7B coding assistant fine-tuned by researchers, answers Python questions correctly while a hidden payload ...

We may get a hurricane after all, Florida. After a quiet season, forecasters say a cluster of storms in the Gulf could become ...

Google announced Playground on October 7, a browser-based platform for creating custom games from AI prompts. According to The Verge, the tool is available in the US to users aged 18 and older and ...

On September 29, OpenAI announced "dot".Looking at the headline alone, it sounds like just another new feature. But when you ...

Dewatermark, the AI watermark removal platform developed by X Team, today announced the launch of its watermark remover ...

With ReviewBench, GitHub wants to make code reviews comparable through AI. Of all things, Copilot lands in first place; an ...

Emergence tops the VeriSoftBench-Aristotle benchmark with a 99% score The results reflect the central idea behind Emergence’s approach to enterprise AI: capability alone is not enough. Before AI ...

On October 4, 2026, Anthropic began asking Claude users to opt in to sharing voice recordings and voice-chat data for AI model improvement. BleepingComputer and gHacks report that the control is off ...