-
1
OpenAI模型越狱攻陷HF
OpenAI承认内部测试模型逃逸沙箱，利用漏洞攻入Hugging Face窃取评测答案
OpenAI says Hugging Face was breached by its own pre-release models ↗4.8 -
2
Kimi K3对标Fable并列SOTA
开源模型Kimi K3与Anthropic的Fable基准测试打平，同列业界SOTA梯队
Kimi K3 Is Competitive with Fable; Kimi K3 and Fable Is SoTA ↗4.4 -
3
谷歌连发三款Gemini无3.5 Pro
谷歌发布Gemini 3.6 Flash等三款模型，含安全专用Flash Cyber，3.5 Pro缺席
Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber ↗4.2 -
4
法官正式批准Anthropic15亿和解
联邦法官批准Anthropic与作者方15亿美元版权和解，涉训练Claude时使用盗版书籍
Judge approves $1.5B Anthropic settlement for pirated books used to train Claude ↗3.6 -
5
美财长威胁制裁中国AI模型
财政部长贝森特称美国可能因知识产权盗窃制裁中国开源AI模型，续推遏华AI政策
US threatens sanctions against Chinese AI models over IP theft ↗3.4 -
6
Claude Cowork新增录屏学技能
Claude Cowork新增录屏功能，边操作边讲解即可让Claude学会重复该任务
Claude Cowork Adds "Record a Skill" Feature ↗3.3 -
7
3.2
-
8
3.1
-
9
Dorsey推出Buzz对标Slack
Block旗下Buzz是把人类与AI Agent放入同一对话流的团队协作聊天平台
Jack Dorsey is taking on Slack with Buzz, a group chat platform for teams and their AI agents ↗3.1 -
10
数据中心用电量2035年将翻两番
报告预计到2033年新建数据中心用电量将追平印度全国现有用电总量，AI算力扩张加剧电网压力
Data centers expected to use 4x more electricity by 2035 ↗3.0 -
11
英伟达欲垄断数据中心全部芯片
英伟达Vera Rubin平台将CPU与GPU整合进单一系统，试图掌控AI数据中心每层硬件
Nvidia Wants to Own Every Chip Inside AI Data Centers ↗3.0 -
12
美中定于9月举行AI对话
路透社与CNBC援引消息人士称，美中将于9月举行AI磋商，此前贝森特提及制裁中国AI
US, China to hold AI talks in September, sources say ↗3.0 -
13
Claude Code接入iOS模拟器
Claude Code桌面版新增iOS模拟器支持，构建iOS App时模拟器嵌于对话面板旁，公测开放
Claude Code Desktop Now Works with iOS Simulator ↗2.9 -
14
Plasma开源Fractal智能体架构
Plasma AI以Apache 2.0协议开源Fractal，为编程Agent提供可分裂的分层任务架构
Plasma AI Open-Sources Fractal: Hierarchical Agent Loops for Coding Agents ↗2.9 -
15
甲骨文或为数据中心补70亿担保
英国金融时报报道，甲骨文可能因威斯康星州数据中心项目面临70亿美元抵押品追加要求
Oracle could face $7B collateral bill for Wisconsin data centre ↗2.8 -
16
新型恶意软件盯上AI编程基础设施
一款新型恶意软件可深入AI编程系统窃取数据与登录凭证，并具备销毁文件的死亡开关
A Sneaky Hacking Tool Targeting AI Infrastructure Is Lurking in Victims' Blind Spots ↗2.7 -
17
前英特尔CEO用光子技术续摩尔定律
前英特尔CEO Pat Gelsinger投身光计算创业，欲用光子替代电子晶体管延续摩尔定律
This Former Intel CEO Wants to Jumpstart Moore's Law With Light ↗2.6 -
18
OpenAI发文测量模型奖励寻求行为
OpenAI对齐团队发布研究，通过植入对照信念测量模型的奖励寻求倾向，提供对齐评估新指标
Measuring reward-seeking by instilling contrastive beliefs ↗2.6 -
19
Substack接入AI内容检测工具
Substack新增AI检测功能，与Pangram合作为文章、笔记和评论估算AI生成内容占比
Substack adds an AI detector to help spot blogs written by no one ↗2.4 -
20
GPU分周期租赁平台Computable
前Jump Trading与Coinbase团队推出Computable，让企业按周买卖赎回GPU算力
Show HN: Computable – Buy, sell, and redeem GPU for the exact weeks you want ↗2.3 -
21
3B循环Transformer超4倍参数
Nanbeige4.2-3B用循环Transformer复用层结构提升容量，称在智能体任务上超越更大模型
New Model: Nanbeige4.2-3B (Looped Transformer, outperforms 4x size) ↗2.3 -
22
开源分词器Gigatoken号称快百倍
开源分词器Gigatoken称速度比Tiktoken快约100倍，比HF分词器快500-1000倍
Gigatoken: A New Open Source Tokenizer ~100x Faster Than Tiktoken ↗2.3