Correcting the baseline evaluation from Part 1 before training again: both the original and fine-tuned Qwen2.5-1.5B models predict no failure for every drive.

Fine-tuning Qwen2.5-1.5B with QLoRA on an RTX 3060 to predict hard drive failures from public Backblaze SMART data. Training worked; prediction did not.

Following one local Ollama request from HTTP JSON to GPU kernels and back on an RTX 3060 homelab machine.

Why I built a selective China-facing VPN path in my homelab for devices that only behave properly on a China network.

AI, Dopamine, and the Night I Rewired My Home Lab A few nights ago, I opened my real backlog. Not Jira. Not a curated roadmap. The actual list: stabilize my k3s cluster running across Raspberry Pis and an old desktop build a network sentinel agent on my OpenWrt router to detect malicious traffic improve observability (Prometheus, exporters, Grafana dashboards) monitor smart plugs for power anomalies orchestrate domotic scenarios based on real signals (not just timers) This is the kind of list that usually grows.

A beginner-friendly tour of the core math behind large language models: vectors, matrices, attention, probability, and optimization.