Last week, a lawyer client asked me something that made my heart drop

Last week a lawyer client asked me: "Is the contract AI assistant you built for me sending my contracts to US servers?" I was speechless.

What is ds4? Redis founder antirez's new project

I first saw ds4 last week while scrolling Hacker News — 209 points, 56 comments, which is quite a buzz for that circle. Clicked in and saw the author was antirez, the same guy who wrote Redis 15 years ago. You might not have heard of Redis, but tons of websites run it for caching under the hood.

He'd stepped back from the spotlight for a few years and now he's coding again. This time it's ds4 — a small tool that runs AI models on your own computer. Same category as Ollama: model files download to your machine, conversation data never leaves, works offline.

Concrete scenes: a tech friend of mine, Wang, posted in the HN comments last week saying he ran a small model with ds4 on his M1 Mac and could ask technical questions completely offline; another small team building a medical SaaS uses it for processing patient record summaries, because hospital data isn't allowed outside internal networks at all.

The cost to replicate this today, honestly, isn't low

Bottom line: this is an open-source tool, costs $0, but you need to put in effort.

  • Cost: $0, open source and free
  • Time: ~2 hours to set up the environment (plus downloading several GB of a 7B model), and another half-day of debugging to get your first conversation working
  • Technical barrier: need Python and command-line basics, plus the ability to read error messages — I got stuck here before, staring at a wall of red English text and questioning my life choices
  • First step: visit dwarfstar.sh, read the README, follow the install steps; or just ask ChatGPT "how to install ds4 on Mac" and let it walk you through step by step

Honestly: if the words "terminal" and "command line" make your head spin, skip this for now. This tool isn't for everyone — it's fine to not try it right away.

How to choose at different stages

Just starting out: you're still finding clients — just use off-the-shelf tools (Doubao, Tongyi Qianwen web versions) for now. Worry about privacy when clients actually ask.

Have 1-2 clients: if clients start asking "where does my data go?", I'd seriously consider two options — find a tech friend to help deploy something ds4-like, or use one of those "local deployment" Chinese AI services.

Scaling up: if you have 5-10 clients and 1-2 of them are super sensitive about data (medical, legal, accounting), I'd have a tech co-founder or contractor set up one of these tools — long term you save API fees per project, and more importantly, it becomes a selling point you can pitch to clients.