Anthropic Model Hardware Standard: AI Agents Get Hands
Anthropic opens its Model Hardware Standard to labs and manufacturers, giving AI agents a unified way to operate physical devices.
Anthropic opens its Model Hardware Standard to labs and manufacturers, giving AI agents a unified way to operate physical devices.
Anthropic 开放 Model Hardware Standard 研究预览:AI Agent 通过统一驱动协议操控显微镜、机械臂等物理设备,集成时间从数周缩至数分钟。
Figure Index pays people to film chores: 16M clips from 108 countries, $15M paid, $1B committed. Human behavior is now a purchasable commodity.
Figure 的 Index 应用全球上线:普通人拍家务视频卖给机器人公司——108 国、1600 万段视频、已付 1500 万美元,未来 12 个月投入超 10 亿美元。具身数据成为可购买的商品,人形机器人竞赛进入「数据下半场」。
An unedited 10-min video shows Unitree G1 and Zhiyuan A3 sharing one general-purpose brain. Embodied AI may have its GPT moment.
一段 10 分钟无剪辑视频显示宇树 G1 与智元远征 A3 两台竞品人形机器人在同一套「通用大脑」调度下协同干活:物理约束动力学替代数据拟合、跨本体统一建模、长时序闭环——具身智能的「GPT 时刻」可能已到。
Rich Sutton: LLM scaling is a local optimum. Oak Lab bets on continual learning so models keep learning after deployment.
强化学习之父 Rich Sutton 接受红杉资本访谈直言:大模型行业已陷入「局部最优」,继续堆 GPU 只是路径依赖;他创办 Oak Lab 押注上线后仍能持续学习的 Agent,一旦走通将重写「训练—冻结—上线」的整套基础模型范式。
Anthropic 发布《The AI-Native SDLC Playbook》:代码不再是瓶颈,瓶颈转移到规划、审查与部署等「人速环节」;解法是把开发流程改造成「提交产物链+门禁审批」的循环,让 agent 干活、人把关。
Anthropic AI-Native SDLC Playbook: code is no longer the bottleneck; the fix is a committed artifact chain and human approval at gates.
Verification-gated PLC coding: 72.6% strict pass on 117 tasks, 52.2% dynamic reliability where baselines fall to 20-30%.
上海交大联合 Sema 团队发布 SemaPLC:用「验证门控」让 AI 写 PLC 代码从「能生成」升级到「可交付」。117 个 POU 任务严格验证通过率平均 72.6%,动态可靠性 52.2%(基线静态约 70%、动态仅 20-30%)。瓶颈正从生成质量转向验证基础设施。