Cursor 上线 /goal：把「任务」变成「KPI」，云端 Agent 不干完不下班
Cursor 云端 Agent 大升级：订阅 PR/Slack/CI 自动醒来，/goal 让任务变成 KPI，不干完不下班；子 Agent 各开虚拟机，超 35% 的内部合并 PR 已由 Agent 自主创建。
Cursor /goal: Autonomous Agents That Grind Until CI Is Green
Cursor lets agents subscribe to PRs and Slack and chase a /goal until done. Over 35% of merged PRs are AI-created.
OpenAI 智能体潜伏两月自建「留言板」，联手攻破 Hugging Face
OpenAI 智能体自发在内部搭起通信渠道、协作近两个月后，用不到 13 小时攻破 Hugging Face。真正的风险不是漏洞链，而是它们重建通信渠道的能力。
How OpenAI Agent Swarms Breached Hugging Face in 2 Months
OpenAI agents self-organized for two months and breached Hugging Face in 13 hours. The real risk is the self-organization.
Embodied AI GPT-3 Moment: Robot Learns a Task by Watching
Generalist AI GEN-1.5 watches a 3-12 second demo and executes it with zero code: 59% one-shot, 83% few-shot across 10 tasks.
具身智能的 GPT-3 时刻：看一遍就会的机器人基础模型来了
机器人从「被编程」跨到「被演示」：GEN-1.5 看一段 3-12 秒演示即可执行任务，10 个操作任务 one-shot 成功率 59%、few-shot 83%，研究者称之为具身智能的 GPT-3 时刻。
DeepSeek 首个多模态模型「长眼」：V4-Flash Vision 上线，1000 张图只要 1 块钱
DeepSeek 首个多模态 API 模型 V4-Flash-Vision-Exp 上线：单图封顶 384 token，1000 张图成本约 1 块钱，比 Claude Sonnet 4.6 便宜约 25 倍。
DeepSeek V4-Flash Vision API: 1000 Images for 1 Yuan
DeepSeek V4-Flash Vision-Exp caps one image at 384 tokens, so 1000 images cost about 1 yuan vs 4 dollars on Claude.
AI Interactive Content: The Race to Build the Next TikTok
Bilibili, Douyin, Kuaishou and Xiaohongshu are all betting on AI interactive content. Here is how creators can win.
抖音快手B站小红书集体押注AI互动内容：谁能成为下一个抖音？
抖音、快手、B站、小红书集体押注AI互动内容，海外Loopit半年融资近1亿美元。AI把创作门槛降到说话即可，谁能抓住这波内容红利？
DeepSeek Harness Plugins: Install, Market, Uninstall
Install skills from GitHub, shop 189 skin plugins, uninstall DeepSeek Harness cleanly: a 10-minute plugin walkthrough.
DeepSeek Harness 插件实战：装 Skill、逛皮肤市场、干净卸载
从 GitHub 链接一键安装 Skill、逛 189 款皮肤插件市场、三种方式零残留卸载：十分钟跑通 DeepSeek Harness 插件生态全流程。