Fine-Tuning LLMs on Apple Silicon: Lessons from Production
Three crashes, two NaN catastrophes, and one breakthrough: fine-tuning a 33B model on Mac Studio to learn my coding patterns. Real configs, real mistakes, real lessons.
What I built, what broke, and the evidence that changed the next decision.
No spam. No marketing. Just the writing.
Three crashes, two NaN catastrophes, and one breakthrough: fine-tuning a 33B model on Mac Studio to learn my coding patterns. Real configs, real mistakes, real lessons.
I replaced my OpenVPN server with a 7-node Tailscale mesh. Zero config, sub-50ms latency, and actual zero-trust security. Here's why mesh networking beats traditional VPNs.
I've been running a tech consultancy solo for a while. Recently I started bringing on my first team member. It's been equal parts exciting and terrifying,
Small churches deserve real management software without the $200/month price tag. So I built one.
An honest look at Planning Center, Breeze, Tithely, and the newer options. Which church management platform actually fits your church?
I run 63 Docker containers on a 16GB AWS instance serving 100+ users across 48 services. Here's how I keep it running for $87/month when SaaS would cost $1,200.
Churches run on free tools until the hidden costs catch up. Here's what actually matters in church management software.
After a decade building integrations with ADP, Workday, Okta, and Entra ID, I've learned that SCIM is a guideline, not a spec. Here's what actually works.
Three years ago my homelab was a Raspberry Pi running Pi-hole and a Plex server. Today it's a multi-node cluster handling production workloads for paying clients.
I fine-tuned a 20-billion parameter model on my Mac Studio M3 Ultra and took validation loss from 1.85 to 0.79. Here's what I learned about running serious AI workloads on Apple Silicon.