Hacker News Reader

Comment by lrvick | Hacker News Reader

I trust mine considerable, but only because I can run it mostly offline on hardware I own that lives in my garage. It baffles me that I am the only technical person I know that exclusively uses AI this way.

View on HN

I am a technical person and while I appreciate the challenge of buying the hardware, setting up the software stack and managing the system, I can't figure out any use cases that would justify the work. This is the main issue I have with Muse, Dots, etc. There just must be something about my routines that don't align with what these companies are expecting their users to do. I do take advantage of the investor subsidized coding agents, but it will take years of my usage of their services to come close to initial start up cost of my own system.

I am very curious as to what other people of using these agents for? Some of the use cases I hear; comparative shopping, travel planning, etc. don't appeal to me. Our people using them to manage family life issues like school schedules, meetings, etc? I feel like I am missing out, but I also am not seeing the greater appeal yet.

Once you have a local agent you can actually trust is not sharing your conversations anywhere, you just use it like a really smart search engine, and your need to use a web browser mostly goes away.

Some of my recent use cases...

Help me find cheap flights and a hotel and plan some interesting sights to see because I am speaking at x conference on x date.

Download all my favorite music from x streaming service as FLAC files locally, then convert them to OGG and put them on my portable music player.

Download all my youtube subscriptions locally to this folder and strip out any ads or sponsored content.

Here are a list of all the books I own. Can you find me DRM-free epub copies and put them on this e-reader for me?

Here is a picture of an object next to a measuring tape. Can you 3d print me a holder perfectly sized for it based on a openscad file i can easily tweak?

I am really stressed right now. Can you help me talk through everything I have to do and help me rationally prioritize? Can you remind me and keep me on track and be my fake boss for today?

Can you look through my email and keep track of any appointments or invites so I stop missing meetings?

Thanks for the useful tasks. If I actually had your setup, I might do those things. Like you, I am not going to ask Meta, et al. to do those things. I also am hesitant to invest in setting up a new system. However, I do have a M2 Mac Studio, so maybe I am closer to being able to run a local system than I thought.

The cool part when you have a local decent model is you can ask it to write automations for you that you can review, and see what works for you. Can keep stacking new requirements.

I mostly use jcode in part because of the self-dev mode where it edits its own rust code and restarts itself into new binaries seamlessly to give itself new features.

I follow the same approach (finding LLMs useful but only running open models on my own hardware) but I've been limited by also not letting them into my email and being pretty restrictive about their internet access. Would you mind sharing what stack and agent harness you use, since your other comments make it clear you're highly restrictive about what kinds of tech you run at home and I assume you vetted it thoroughly, and it sounds like it's working well?

I'm wondering specifically about things like what kind of tool or restrictions you use on web browsing to avoid being flagged as a bot and how you guardrail access to your email (to avoid, say, the LLM deciding that "help me organize" includes telling people you want to reschedule). I use Qwen 3.8 26b at home and find it very useful for some tasks but when I think of using it as a general assistant rather than coding assistant, I keep thinking of times it has gone wildly wrong (e.g. it once spent 20 minutes trying to get around Supabase authentication because it couldn't figure why a certain call wasn't working)

My go to right now is jcode, and doing a lot of one-shot experiments. Like I have a session being my MPD DJ as my local spotify right now.

That said for persistent things like interfacing with homeassistant etc I am playing with zeroclaw.

I run the same model as you for coding. For more sane automation tasks i recommend doing everything possible with MPCs so you can tell it to only use existing tools. ymmv

By no means is my setup well oiled yet. Different vms and systems doing different things. As a QubesOS user jailing an agent for one job to one VM is easy, though I am working on an enclave native operating system that should make this easier for most people.

Just dropping in to say that this is no longer that big of a problem. There are hosted solutions only for agents from DO, Nous, etc But also more generic self-hosting solutions, where you can host a lot of stuff and also agents, like moose

I think as people realize that self-hosting is a good option. Good services will appear more

Reply on news.ycombinator.com