Part 3 of Magatama's journey reveals unseen risks in model training and highlights the need for additional safeguards before execution gates.
A diagnosis-only agent held up fine, until I noticed I was doing the same manual fix every time it was right. Here's the middle ground I landed on between full autonomy and a blind whitelist.
The first part of a short series on how Magatama went from manual alert triage to an autonomous ops agent, starting with the deliberately powerless version that could only explain what was wrong.
A duplicate training job that wouldn’t stay dead, and a batch of corrupted weights that came within seconds of going live. Notes from a rough week running my own training infrastructure.
I pointed a serious security audit at Magatama, the agent I built to run my own server fleet, and asked exactly one question: how could its own autonomy be turned against me. Five things came back confirmed, and all five got fixed the same day.
Long before NOGnet was a platform or a business decision, it was a stack of small, repeated annoyances I picked up simply by showing up to events as an ordinary attendee, year after year.
Coming off the H1 retrospective, here are six things I'll be tracking through the second half of 2026. I've split them by confidence — three calls I&
Six months into 2026. Time for the half-year retrospective. I don't keep a formal prediction list. I do leave a lot of directional calls scattered through posts, mailing
RPKI got the attention. Cyber-insurance language rewritten, hyperscaler peering policies tightened, deployment curves bending upward. The industry has had three years of sustained conversation about routing security, and it'
In April I wrote about linear drive optics and called it a power-budget story. The argument: LPO trades modest reach for significant power savings, right for AI fabrics inside a