
structure|1️⃣ 三级笔记、思想框架
作者:Mariano Belinky
核心论点:OpenClaw 的高门槛和混乱感恰恰是它的本质特征——它不是一个「做完了再发布」的产品,而是一把「边磨边切」的刀。适合它的人,会觉得它 exceptional;不适合的人,等一等也没关系。
OpenClaw 不是普通软件——没有稳定版、没有「完成态」可以冲刺
Chaotic: sometimes, but also absolutely amazing. What other project has 340k GitHub stars and ships every 1.5 days with 60 changelog entries and 75k lines changed on average?
MIT 协议,完全开源,没有 "open core" 把戏
You can read it, fork it, break it, fix it.
This is infrastructure for a different way of working, not a productivity layer bolted on top of the old one.
一台笔记本 / 便宜 VPS / 树莓派级别就能跑 gateway
讨论大型服务器和 GPU 的场景:跑本地模型,或同时跑几百个 agent → 绝大多数人现在或永远都不需要
算笔账:$20-$200/月的 frontier model 订阅 >> $20,000 硬件投资
You are not Google. You don't need Google's infrastructure. Chill, spin an agent, and enjoy the journey.
本地模型确实给你更多控制权
但本地模型普遍更弱、更容易被操纵、压力下更不可靠
"Local" is not a free safety win. It's a tradeoff between control and capability.
当前阶段的判断标准:你更想要一个能防住 prompt injection 的强模型,还是担心 frontier model 公司从你的 prompt 里获取什么?
核心权衡:你想让 OpenClaw 做的事越多(付账单、读日历、买票、碰金融账户),你就得交出越多权限
This is not a footnote. This is the actual contract.
错误做法 ×2:
正确做法:
Extend trust incrementally as you understand the system better, build in the right checks, and go in with your eyes open about what you're trading.
Complexity has a cost. It breaks in more places, it's harder to debug, and it's harder to trust. Start simple and add complexity only when simplicity has genuinely run out of road.
concepts|2️⃣ 关键概念、概念网络
context:
OpenClaw is not normal software, and by nature it probably won't be for a long time. Most software is built to be stable, predictable, and forgiving. OpenClaw is none of those things, not because we're careless, but because there's no finished version to race towards.
费曼一下:一个开源的 AI Agent 基础设施项目,MIT 协议,340k GitHub stars。它的核心特征是「永远未完成」——不是因为质量差,而是因为它赌的是软件工程本身正在结构性变化,边界一直在移动。每 1.5 天发一版,平均 60 条 changelog、75k 行代码变更。
agentic reading|3️⃣ 费曼 x3