Local AI tinkerer. Homelabber. DGX Spark operator. Writing about agent workflows that actually work.
Joined April 2011
- Tweets1.3K
- Following612
- Followers85
- Likes819
Pinned Tweet
My AI agents have physical buttons now!
I built Dispatch to turn @work_louder’s Creator Micro 2 into a control surface for @herdrdev.
If an agent needs me, its key lights up, and one press takes me there.
What’s the first thing you’d map to a button?
Partially because I want to upgrade home network security, and partially because I love extending my homelab...some new purchases are showing up soon from Unifi 👀
Ford: "Read this 9-page PDF, or book a dealer visit."
Me: hands the PDF, zip, and a USB stick to @claudeai.
Claude Code: formats, unzips, copies, verifies, ejects.
Me: plugs it into the truck.
Truck updated in 30 minutes. Dealership not required.
While job searching for the last couple of months I created a few tools to aid in that process (links below)
Connections Finder: Export your LinkedIn data and cross reference connections' companies with job openings. Use as the basis for cold outreach
Job Matcher MCP: An MCP server you and your agent can run to find openings matching your skills and desires (makes use of Jev by @typesafeai).
I was fortunate enough to pickup an M5 Max 64GB Mac Studio yesterday and *WOW* this thing is blazing fast. Won't run the big models, but that's what I have the sparks for. This is the best machine for daily work I've ever used
The advent of personal software is unlikely to result in everyone making their own Photoshop, IMO, the cost benefit simply is not there. However, I do think it’ll pressure companies of closed source software to rethink their approach to extensibility by developers
I think I’ve landed on M5 Max with 64gb. The sparks stay the main local inference engines while the Mac is my daily workhorse and runs ancillary models. What say you, fellow local AI friends?
One of my favorite local AI automations is quite boring:
Mail → scanner → watched folder → local LLM → properly named document
I use it for insurance, financial records, and other personal paperwork.
#LocalModelsAreNotUseless
24 hours with @ChatGPT Dot:
1. Feels slow: mostly because there’s little feedback while it works.
2. Duplicate notifications across desktop + mobile.
3. Calls can quietly “hang up” if things go idle
4. But it’s easily the most proactive agent I’ve used and when it works, it’s great.
Excited to see where the team takes it.
Fascinating to see AI companies moving up the stack as models coalesce around similar performance levels. Feels like the end game for bottom up SaSS. Fighting that lock-in via local AI has become a cost prohibitive endeavor...
Confusing as to why @ChatGPT fails to successfully connect to Google Drive if you don't give it 100% of the requested permissions. Surely things should still work with fewer permissions?
I appreciate what Meta is doing re: branding w/ Muse as this cute teddy bear (never mind it runs in the same data center that folks are opposing), but how deeply embedded do we expect "normal" folks to be with AI?
GM. I’ve been building something new for @herdrdev using the @work_louder Creator Micro.
It’s been incredibly fun to work on, and I can’t wait to show you all what it can do!
