机器人强化学习：从仿真到真实世界部署（MuJoCo + Gymnasium）
完整梳理了机器人强化学习的全流程：从 CAD 建模、MuJoCo 仿真、Gymnasium 环境搭建，到 PPO 策略训练，再到 ONNX 硬件部署。以旋转倒立摆为教学示例，作者揭示了奖励设计而非算法本身才是仿真到真实迁移失败的关键根源，并强调域随机化与硬件安全限位对四足及人形机器人部署同样不可或缺.
完整梳理了机器人强化学习的全流程：从 CAD 建模、MuJoCo 仿真、Gymnasium 环境搭建，到 PPO 策略训练，再到 ONNX 硬件部署。以旋转倒立摆为教学示例，作者揭示了奖励设计而非算法本身才是仿真到真实迁移失败的关键根源，并强调域随机化与硬件安全限位对四足及人形机器人部署同样不可或缺.
From CAD to real hardware, this video maps the full reinforcement learning pipeline for robotics: MuJoCo modeling, Gymnasium environments, PPO training, and ONNX deployment. Using a rotary inverted pendulum as a teachable example, it shows why reward shaping—not the algorithm—usually breaks sim-to-real transfer, and why domain randomization and hardware safety limits are non-negotiable for humanoid and quadruped robotics.
Moonshot AI's Kimi K3 is the first open-weight model to cross 3 trillion parameters, activating just 104B per token. Three innovations power it: a latent-compressed mixture-of-experts, Kimi Delta Attention for 1M-token context, and attention residuals across 93 layers — together delivering 2.5x better scaling efficiency than K2, at frontier-competitive benchmark performance. Moonshot AI 推出的 Kimi K3，是首个突破3万亿参数的开放权重模型，每个token仅激活1040亿参数。其核心依托三大创新：潜空间压缩混合专家系统、支持百万token上下文的Kimi Delta Attention，以及贯穿93层网络的注意力残差机制——三者协同，使扩展效率较K2提升2.5倍，性能达到前沿水平。
This two-part series covers robotics kinematics end to end. Part one reviews a course that derives rotation matrices, DH parameters, forward/inverse kinematics, Jacobians, trajectory planning, and Newton-Euler/Lagrangian dynamics — each verified live on a UR robot in RViz via ROS 2 Python. Part two is the practical "step zero": installing ROS 2, building a workspace and package, writing a minimal DH-based forward-kinematics node, and troubleshooting common setup pitfalls like sourcing and distro mismatches. 本系列共两篇，完整覆盖机器人运动学。第一篇测评一门课程：从旋转矩阵、DH 参数、正逆运动学、雅可比矩阵，到轨迹规划与牛顿-欧拉/拉格朗日动力学，每一步都在 UR 机械臂上通过 ROS 2 Python 于 RViz 中实时验证。第二篇是实操"第零步"：安装 ROS 2、搭建工作空间与功能包、编写基于 DH 参数的最小正运动学节点，并排查环境变量与发行版不匹配等常见配置陷阱。
Claude 5 shifts the bottleneck from model capability to human specification. Anthropic cut 80% of Claude Code's system prompt, embracing intent-based prompting, bi-directional clarification, progressive disclosure via skills, and layered memory — turning execution into a self-learning loop rather than rigid rule-following. Claude 5 将瓶颈从模型能力转移到人类的规格说明。Anthropic 削减了 Claude Code 系统提示词的 80%,转向意图导向的提示方式、双向澄清式提问、基于技能的渐进式披露以及分层记忆机制——让执行过程成为自我学习的循环,而非僵化的规则遵循。
新加坡前外长杨荣文认为，中国的治国之道本质上是防御性的，而非扩张性的。谈及台湾，他预测若强制统一，约5%至10%的人会离开，其余人将像香港在2020年后那样逐渐适应。从特朗普的谈判风格到DeepSeek的开源策略，杨荣文以罕见的内部视角，描绘出他称之为"三个太阳"的多极世界图景。 Singapore's former Foreign Minister George Yeo argues China's statecraft is fundamentally defensive, not expansionist. On Taiwan, he predicts 5–10% might leave under forced reunification, the rest adapting much like Hong Kong did post-2020. From Trump's negotiating style to DeepSeek's open-source gambit, Yeo offers a rare insider's map of a multipolar world he likens to "three suns."
George Yeo's blunt line at HKU — Taiwan can't shape US-China relations, but it can decide whether to be used — reframes the island's predicament as choice, not fate. His fireside chat ranges from Lee Kuan Yew's Hong Kong ties to the multipolar shift, dollar hegemony, and Chinese civilizational logic, ending with a warning: policies built on illusion lead to tragedy. 杨荣文在港大直言：台湾无法左右美中关系，却能决定是否甘愿被利用——将台湾的处境重新定义为选择，而非宿命。这场炉边对谈从李光耀与香港的渊源，谈到多极世界转型、美元霸权与中华文明逻辑，最后警示：建立在幻想之上的政策，终将酿成悲剧。
**The real productivity ceiling in the AI agent era isn't your tool — it's your Skill configuration.** After testing dozens of agents, one truth stands out: Skill Creator, Find Skills, Superpowers, jStack, Frontend Design, and Baoyu's content toolkit form the essential stack. Configure these four Skill groups correctly, and any agent becomes a full team. --- **AI Agent 时代真正的生产力天花板，不是你用的工具，而是你配置的 Skill。** 实测数十个 Agent 后，一个真相浮现：Skill Creator、Find Skills、Superpowers、jStack、Frontend Design 与宝玉工具箱，构成了核心生产力矩阵。四组 Skill 配置到位，任何 Agent 都能顶一支团队。
Yann LeCun argues that LLMs cannot achieve human-like intelligence: a four-year-old absorbs as much data through vision as all internet text. His solution—world models built on JEPA, predicting in abstract representation space rather than pixels—enables planning by energy minimization, intrinsic safety, and emergent physical common sense. Now open-sourced via V-JEPA 2, EB-JEPA, and LeJEPA for hands-on experimentation.
Traditional logic says you must choose: bespoke high-value consulting or infinitely scalable software. Palantir refused the trade-off and invented the Forward Deployed Engineer — an embedded engineer with direct CEO access who ships production code and precipitates every engagement into reusable platform capability. 传统逻辑要你二选一：高单价的定制咨询，或可无限扩展的软件。Palantir 拒绝妥协，造出了"前端部署工程师（FDE）"——能直达客户 CEO、写生产代码，同时把每一次项目沉淀为可复用的平台能力
