The only thing that matters this week is llama.cpp's relentless optimization work—five rapid releases suggest they've cracked something meaningful in inference efficiency that'll reshape what's actually deployable at edge. Zenity's $125M security raise is worth watching precisely because it's unsexy: AI agents are shipping faster than their threat models, and someone finally noticed the gap is a business. Everything else is noise—Japan's drone posturing, solar hype cycles, and that Australian defense scaremongering are all downstream consequences of *who controls efficient inference*, not drivers of it. The papers on embeddings and sparse solvers matter only if they actually ship into production; abstract improvements in mathematical robustness don't compete with incremental llama.cpp wins that let models run on $50 hardware tomorrow.