非官方文章 Unofficial notes Asia/Shanghai
把口袋电脑变成 AI agent
CardputerZero 是一台完整的口袋 Linux 电脑,但内存只有 512 MB。下面这组短文先正视硬件约束,再讨论怎样让它真正派上用场:云端模型负责推理,设备负责交互和执行工具。这不是官方文档或教程站点。
A pocket PC as an AI agent
CardputerZero is a complete pocket Linux computer with only 512 MB of RAM. These short notes start from that constraint, then describe the approach that actually fits: the model runs in the cloud, and the device handles interaction and tool execution. This is not an official documentation site.
-
把 CardputerZero 变成 AI agent,先要分清三条路
Three paths to an AI agent on CardputerZero
这台口袋 Linux 电脑只有 512 MB 内存。最可行的做法,是让云端模型负责推理,让设备做终端和工具执行器。
This pocket Linux PC has only 512 MB of RAM. The workable approach is to keep the model in the cloud and use the device as a terminal and tool runner.
-
动手前先确认型号、镜像和网络
Check the model, official image, and network first
Lite 没有摄像头和 IMU;官方系统是基于 Raspberry Pi OS / Debian 的 arm64 镜像,并预装了 Python 3 与终端。电源和 USB 模式会直接限制 agent 怎么用。
Lite has no camera or IMU. The official system is an arm64 Raspberry Pi OS / Debian image with Python 3 and a terminal preinstalled. Power and USB mode constrain how an agent can run.
-
云端工具调用:目前最可行的路径
Cloud tool-calling is the path that actually fits
从官方预装的 ZClaw 开始,或自己写一个只在本地执行白名单工具、只用 Python 标准库的客户端。模型不要放在这台 512 MB 的机器里。
Start with the preinstalled ZClaw app, or write a small Python client that only runs an allowlisted set of local tools. Do not host the model on 512 MB of RAM.
-
本地跑大模型?在 512 MB 上基本不实用
Local LLMs are not practical on 512 MB
与树莓派 Zero 2 W 同级的内存,决定了现代 LLM 几乎没有空间。只适合极小量化模型的实验,不能当作日常 agent。
The same memory class as a Raspberry Pi Zero 2 W leaves almost no room for a modern LLM. Tiny quantized experiments may be possible; a daily-driver agent is not.
-
键盘、屏幕、相机和 GPIO 如何成为 agent 的工具
Turning the keyboard, camera, and GPIO into agent tools
把机身外设做成可调用的工具,并正视屏幕尺寸、型号差异、Cap 不能热插拔,以及电源与网络对采集频率的限制。
Expose onboard I/O as callable tools, while respecting the tiny screen, model differences, no-hotplug Caps, and power and network limits on how often you can sample.