Blog

Phones, EFTPOS, and regional trains all went down today when Telstra's network failed. The outage itself is Telstra's problem to explain. The fact that a national rail network had no working fallback beyond one carrier is a design decision worth questioning — and a useful test for which systems actually need it.

I ran Apache for two decades because it was the standard. Then I came to Nginx through the reverse-proxy door, and somewhere along the way it became what I reach for first. The honest story of how that happened — and why the real answer isn't that one server won, but that the work itself changed shape.

I built a deliberately minimal Android music player with AI assistance in about ten hours — and the most useful thing I learned had nothing to do with the code. The design decisions, where the AI pushed back on human habit, and the honest arithmetic of shipping a "vibe coded" app.

From a 386 running Slackware in a cupboard in 1998 to retiring the last FreeBSD VM in 2024 — what 25 years of running both actually taught me, and an honest assessment of where each one belongs.

Using AI to generate code is genuinely useful in the right hands. In the wrong ones it's a circular saw plugged in at a busy bar on Saturday night. An opinion from someone who's been in IT long enough to have strong ones.

The RTX 2060 Super is out, the RTX 5060 Ti 16GB is in. Full benchmarks across six models show what 16GB of VRAM actually changes — and one finding that no amount of new hardware was ever going to fix.

The GPU is in. The 4am battle with Google's AI over a truncated localhost URL finally gets resolved, Ollama replaces LM Studio, and the benchmarks show what 8GB of VRAM and a correctly configured context window actually buy you — and what they don't.

Four models, one demanding prompt, and a result that runs against the usual assumptions: the general-purpose model beat both code-specialist models — including one with code that would throw a runtime error if you ran it.

The Machinist X99 arrives from China faster than Auspost moved the GPU one state. No manual, no CMOS battery, a heatsink that needed a drill, and a BIOS loop fixed by swapping the keyboard. Ubuntu is running. The GPU might arrive today.

The dedicated build hardware is still in transit. In the meantime: what actually happens when you try to run Gemma 4 E4B on a Core i7-8850H with 64GB RAM for a serious multi-hour session. Spoiler: it starts well.

I keep hitting cloud AI token limits at the worst possible moments. The solution — obviously — is to build something. Salvaged server hardware, a budget GPU, a 25-year-old rack case, and some 3D-printed brackets holding it all together.

I came to 3D printing the way a lot of people do — convinced it would solve a specific problem, then completely sidetracked by the technology itself. This is the broad overview before I go off into the weeds on individual topics.