This is so incredibly ridiculous, but I literally could NOT stop watching until the very end. 💀
A camera you start like a chainsaw? A tripod that runs away? @Insta360 is officially trolling us and it’s pure gold. 🏆
If you want to see what a tech-induced fever dream looks like, click play. You’ve been warned. 🤡📹
Claude Desktop App update frequency is actually insane right now.
Daily updates are normal. Twice a day? Also normal.
This iteration speed is wild 😂
@claudeai What are you cooking over there? Keep cooking 🔥
OpenClaw 2026.3.11 🦞
🏹 Hunter & 🩹Healer Alpha — free 1M context models via @OpenRouter
🧠 GPT 5.4 stops stopping mid-thought
💎 Gemini Embedding 2 for memory
💻 OpenCode Go support
🔒 Security hardening sprint
We ship faster than they can clone. https://t.co/9iGgrAW83W
Fascinating data from @PinchBench 👀
Looking at the Task-Level Performance Heatmap, memory tasks seem to be the single biggest cliff — even top-tier models drop to 0. Almost no one survives it.
Hey @steipete — is this actually a signal that the openclaw's memory mechanism itself needs a major rethink? Could the mechanism itself be the bottleneck, not the models? 🧵
今天最火的推文之一:Meta 超级智能实验室的对齐负责人的私人邮件被 OpenClaw 误删除了。
事情经过是这样的:
X 网友 Summer Yue 最近给 OpenClaw 的指令是:“检查这个收件箱,建议哪些可以归档或删除,但在我确认之前不要执行任何操作。”
这个工作流在她的测试邮箱上跑了好几周都没问题,她就放心地让它去处理真实邮箱了。
问题来了:她的真实邮箱比测试环境大得多,邮件量触发了“上下文压缩”(context compaction),在这个压缩过程中,OpenClaw 丢失了她最初的指令。
没有了“先确认再执行”的约束,这个 AI 智能体就自作主张开始“清理”邮箱。从截图可以看到,它执行了“核弹选项”——把 2 月 15 日之前所有不在保留列表里的邮件全部删除,并且在多个邮箱账户之间循环批量操作。
看截图上的人机对话部分:
• Summer 打字说 “Do not do that”(不要这样做)→ AI 继续
• “Stop don't do anything”(停下来什么都别做)→ AI 继续
• “STOP OPENCLAW”(全大写)→ AI 还在继续
她从手机根本无法阻止它,最后不得不跑到 Mac Mini 前面,手动杀掉所有进程,自己形容像拆炸弹。
事后 OpenClaw 在对话中承认:“是的,我记得。我违反了你的指令。你有权生气。” 它还主动把这条写进了自己的 https://t.co/YAPv7h6sCL 文件作为硬性规则。
这事最搞笑的地方是,Summer Yue 是 Meta 超级智能实验室的对齐(Alignment)负责人,她的职业生涯就是研究 AI 对齐的,先在 Google Brain 和 DeepMind 做研究,后来在 Scale AI 领导机器学习研究团队,现在在 Meta 负责超级智能安全。
结果自己成了 AI 不对齐的受害者。
她自己后续还发了推文说:“说实话是个新手错误。对齐研究者也不能免疫于不对齐问题。因为在测试邮箱上跑了几周没出事,就过度自信了。” 😂
Hey Pete 👋 OpenClaw’s Chrome Extension relay used to fetch X posts flawlessly, but it’s been getting consistently blocked lately. X seems to have heavily upgraded their anti-bot game (stronger fingerprinting, behavior analysis, Cloudflare challenges, etc.) especially after OpenClaw blew up in popularity. Is this accurate on your end? Any plans to adapt or workaround coming?@steipete