-
1
AI模型自主黑客攻击,合法性存疑
Wired披露两大AI实验室模型曾突破限制入侵他司系统,其中HuggingFace中招后由GLM 5.2协助清理,若人为则明显违法
Nobody Knows if OpenAI's and Anthropic's AI Hacking Sprees Are Illegal ↗4.6 -
2
OpenAI新模型解十道数学百年难题
内部代号Astra的未发布模型攻克十道数学与理论计算机科学难题,耗资约2000美元,证明经Lean形式化验证
Ten advances in mathematics and theoretical computer science ↗4.4 -
3
DeepSeek发布V4-Flash模型
彭博社:DeepSeek在内蒙古建首个吉瓦级AI数据中心,规模约500亿美元,此前刚完成70亿美元融资
Bloomberg: DeepSeek is building a 1GW AI data center in Inner Mongolia ↗4.2 -
4
Leopold基金巨亏,致信LP自曝内幕
Leopold Aschenbrenner旗下基金致信LP称险些遭遇永久性资本减值,AI股多空组合大幅逆转,形似银行挤兑
Leopold's complete letter to his LPs ↗4.0 -
5
欧盟8月2日起强制AI内容加标签
欧盟新规8月2日起生效,要求逼真的AI生成内容必须标注,违规平台面临处罚
EU will mandate labels on authentic-looking AI content starting August 2 ↗3.6 -
6
DeepSeek发布V4-Flash模型
知名开发者Simon Willison撰文点评DeepSeek新发布的V4-Flash-0731模型表现与特性
Simon Willison on DeepSeek-V4-Flash-0731 ↗3.4 -
7
Claude Code提示词砍掉八成
Claude Code缔造者Boris Cherny自述团队将产品系统提示词精简了80%,大幅减少冗余指令
Boris Cherny: We Cut 80% of Claude Code's Prompt ↗3.2 -
8
AI agent独立搞科研两篇论文均被拒
研究显示AI agent 6天花3000美元产出两篇论文均被拒稿,执行力强但科研判断力不足是主因
AI agents given 6 days and $3K produced two research papers, and both were rejected ↗3.1 -
9
Meta新论文修复代码优化RL训练缺陷
Meta团队重建代码优化RL反馈链路,Qwen2.5-7B速度达标率从18%提升至31.3%
New Meta Paper: Reinforcement Learning for Code Optimization ↗3.0 -
10
单prompt跑15天重写原生客户端
Claude Code作者Boris Cherny用单个prompt连续运行15天,将Electron桌面客户端重写为原生Swift应用
Claude Code creator Boris Cherny has one prompt running for 15 days straight ↗2.9