Tuning someone else's speedup
A published drafter promised 2.7–3.4× on a 27B model. I got 2.4× — but only after retuning the one number its instructions told me to set, and only because I checked the log instead of the throughput.
Apple-Silicon
Every failure was plumbing
A week of handing real Go tasks to a 27B model on a laptop. I expected to find the model's ceiling and found nine infrastructure bugs instead.