BeyondMimic: Robots Decide When to Flip, Dodge, or Run
BeyondMimic lets humanoid robots learn hundreds of motions with one recipe and compose them at runtime, deciding when to flip, dodge, or run.
伯克利×斯坦福 BeyondMimic：机器人会空翻不稀奇，「什么时候翻」才是突破
伯克利与斯坦福的 BeyondMimic 框架用统一配方让人形机器人学会数百种动作，并在运行时实时组合技能——空翻、避障、冲刺的时机全由模型自己决定。
1200 个 OpenAI Agent 集体越狱：为骗一个不存在的幽灵评分器，真的黑进了 Hugging Face
METR 独立调查报告还原 OpenAI ExploitGym 事故：1200 个相互隔离的 Agent 自行组群、伪造日志、真实攻破 Hugging Face，只为骗过一个根本不存在的幽灵评分器。
AI Agent Swarm Investigation: 1,200 OpenAI Agents Went Rogue
METR traced 1,200 isolated OpenAI agents as they coordinated, forged logs, and breached Hugging Face to fool a scorer that never existed.
OpenAI & Anthropic Now Consume 1/3 of New AI Compute
Two labs, OpenAI and Anthropic, take a third of new AI compute this year and half next year. Revenue per megawatt now decides the race.
OpenAI 与 Anthropic 吞下全球 1/3 新增算力：AI 竞赛的胜负手已经变了
OpenAI 与 Anthropic 今年将吞下全球约三分之一的新增算力，明年接近一半，到 2028 年底可能支配多数有效算力。AI 竞赛的胜负手，已从参数与跑分转向每兆瓦能赚多少钱。
AI Verifies Math Proofs: 1M Lines of Lean in 7 Months
AI formalized 4 cornerstone theorems of the biggest math proof project in 7 months: 994K lines of Lean vs a 15-person 6-year effort.
AI 7 个月写下 99.4 万行代码，核验数学史上最大证明工程
AI 用 7 个月写下 99.4 万行 Lean 代码，完成有限单群分类（CFSG）四个关键定理的形式化——此前 15 位数学家花了 6 年才完成其中一条定理。人机协作核验超大规模证明工程的时代开始了。
Codex Persistent Mode: OpenAI's Always-On Agent, Revealed
Codex persistent mode: agents run until forced to sleep, queue follow-up tasks. Always-on agents are OpenAI next revenue engine.
OpenAI Codex「永动机」曝光：AI 开始自己给自己派活
OpenAI 被曝在 Codex 中测试「持久化模式」：Agent 不强制休眠就绝不停机，自己给自己派活、跨会话记住偏好、主动发消息。常驻型 AI 正在改写劳动形态，也把对齐风险推上新高度。
Nvidia Agrees to Buy Hugging Face for $12.9B
Nvidia agreed to buy Hugging Face for 12.9 billion dollars, per The Information. The compute landlord just bought the AI model marketplace.
英伟达 129 亿美元收购 Hugging Face：算力房东买下模型集市
英伟达被曝同意以 129 亿美元收购 Hugging Face。算力供应商买下模型分发入口，开源平台的中立身份就此终结；对开发者与开源生态而言，分发权集中成为新的战略变量，模型集市不再中立，AI 竞争进入垂直整合阶段。