Pinned Tweet
And that was my last day at Twitter. Laid off by email. The experience has been indescribable from chaos to hilarious. – at San Francisco, CA
Dave Beckett retweeted
The Codex App in main ChatGPT app is quite awful. I'm going back to codex CLI.
This was the reply to the first gpt-5.6-sol code question I did "do a deep review of this codebase and write a report". Quite useless response and the result was 1/3 of what Fable 5 did.
LLMs got good at code because it's verifiable: tests pass or fail. Operations isn't, so manufacture the reward: simulate production, the way robotics learns in simulation.
The catch: the model only covers the failures you already understood.
dajobe.org/blog/2026/06/23/a…
ALT A vertical loop: real production seeds and validates an artificial systems model; the model trains an agent over millions of RL episodes; the learned policy is tested back in reality across the simulation-to-reality gap and feeds back into production. The policy only covers failures the model contained.
Sovereignty as an SRE problem.
The political ask is independence. The engineering work is SLA, staffing, complexity, and creating an operational surface for a sovereign operator with a hyperscaler's stack.
Three years on Google Cloud Dedicated / S3NS.
dajobe.org/blog/2026/05/13/s…
Coding LLMs spread because public code spread.
Production engineering data is private by default, so AI for-ops concentrates around whoever has it.
The data plane matters more than the model.
dajobe.org/blog/2026/05/07/a…
Sadly Replay didn't pay for a durable website or possibly WiFi. So no conference schedule for me. #replay2006 fail
LLMs got good at code because it's verifiable: tests pass or fail.
Production engineering isn't. Same change, different outcome depending on live state. No cheap reward signal.
dajobe.org/blog/2026/05/04/v…
A file is a table. A line is a row. `comm -12` is intersection, `comm -23` is set difference, `join` is an inner join, `sort -u` is SELECT DISTINCT.
A small line-oriented relational algebra hiding inside POSIX.
dajobe.org/blog/2026/05/01/t…
Fleet automation lesson four years on: the caps on your destructive operations bound count. They don't bound concentration. Sorted host lists correlate with physical topology, and 'head -n 5' will cheerfully take out a rack.
dajobe.org/blog/2026/04/29/t…
Presidio Tunnel Tops Park, San Francisco.
Let’s hope Orange Felon won’t screw this one up since he hates Pelosi and walking.
ALT View of Golden Gate Bridge with Presidio tunnel tops park in front on a sunny and cloudy day
I spent Q1 2026 interviewing for Staff+ SRE roles at AI infra companies. Both sides using LLMs. No one asked to see my tool setup, or how I decide when to generate code and when not to.
The bar is still Judgement. The interview format hasn't caught up.
dajobe.org/blog/2026/04/20/l…
Dave Beckett retweeted
Most AI takes come from the outside looking in.
Five Crusoe leaders share what they're actually seeing in AI infrastructure right now. An authentic view from inside the machine: crusoe.ai/resources/learn/el…
New job calendar meetings. Day 1: 1. Day 2: 1. Day 3: 5. Future: Day 4: 3 so far. Day 5: hoping for zero #onboarding
Today I started work as a Senior Staff Production Engineer at @CrusoeAI in SF. dajobe.org/blog/2026/04/13/a…