zhixian

Exploring the unknown with AI.

Now

Luoke Labs

The model runs on your own device. Your mail and what you type don’t go through any cloud service.

Runs locally: The model runs on hardware you own, and the data stays there.
One model: Mail assistant, input method, local model and memory share one model and one memory.
Shared memory: Memory is shared, so every tool gets better the more you use any of them.
Open: Open source, in the luoke-* repos.

I’m looking for people who are into on-device AI and want to build open source together. Get in touch.

Momoda · Available now

I couldn’t find one I liked, so I made one. Free and open source.

Agent Trainer, Level 2: Turning Discord into an OpenClaw Team Room

Telegram for quick chats, Discord for real projects: how I use servers and channels to run OpenClaw tasks in parallel.

2026.02.03
I Built a GUI Because OpenClaw Kept Breaking Its Own Config 🦞

A desktop app for managing OpenClaw agents, channels, models and sessions — no more hand-editing JSON.

2026.02.20
Agent Trainer, Level 2: Multi-Agent OpenClaw Configs, Hands-On

Yes, you can give each channel its own agent. What OpenClaw multi-agent setups can and can't do, with lots of config samples.

2026.02.11
Agent Trainer 101: Setting Up Clawdbot, and the Pitfalls I Hit

Three weeks in and still buzzing: what Clawdbot is, how I use it daily, and the setup mistakes so you can skip them.

2026.01.25
Agent Trainer Security: A 7-Step Self-Check for Clawdbot

Don't panic-buy a Mac mini, don't panic-sell it either: seven checks to keep a powerful agent on a short leash.

2026.01.26
Agent Trainer Security, Level 2: Attack and Defense in the Agent Era

Starting from SlowMist's minimal OpenClaw security guide: how to think about attack and defense once AI has root.

2026.03.09

On AI, tech and products. Each episode picks one thing I find interesting: a new tool, a technical direction, or a question I haven’t worked out yet.

The name is a pun: 县 (county) sounds like 限 (limited).

Where to start

fm, the local model built into macOS 27

On an M5 Max with the network off, calling fm from the terminal gave about 46 tok/s, with the first token in 0.43 s. The context limit is about 8K (8,026 tokens worked, 8,095 errored), not the 4K people quote. In a Chinese typo test it missed 3 of 12. Post ↗

2026.10.22
How many messages fit on the iPhone Duo cover screen

Counted on an in-store unit with the keyboard up: the iPhone Air shows 7 messages, the Duo cover screen only 4. I also made a web page that simulates it on the phone you already have. Post ↗

2026.10.24
Changing two neurons changes Qwen3.5-4B’s personality

At inference time, nudge L14·N8949 up and L13·N7956 down. At 0.5–0.75× it turns “evil” but stays fluent; at 1× it starts repeating itself; at 2× it’s gibberish. I reproduced it once and it held. Post ↗

2026.10.26