A $4k box, an open-source video model, and one stubborn easter egg
Running Wan 2.2 image-to-video on a DGX Spark to make a hidden scene breathe: the undocumented Blackwell setup traps and the trick behind it.
Identity · Security · AI infrastructure
Hi, I'm Brian.
I write down the things I learn, usually after I've made the mistake first. Long-form posts on what I'm building, breaking, and defending.
Featured
Recent posts
First hours with Fable 5
Claude Fable 5 is the most capable model Anthropic has shipped. It also blocks legit security work and drops the zero-retention floor businesses chose it for.
Pack your SSID
A pocket travel router with an IPsec tunnel back to a network I control, broadcasting a trusted SSID that follows me into every hotel and airport.
First hours with Opus 4.8
Claude Opus 4.8 is a sharpening, not a leap. Clear gains, more honesty, hundreds of parallel subagents, and what it all means for security teams.
Opus 4.7 First Hours
A hands-on first-hours review of Claude Opus 4.7. What moved, what broke, the token-cost gotchas, and why Anthropic is still holding Mythos back.
We Keep Making AI Wear an Org Chart
A sharp take on why today’s AI agents feel wrong, and what a quieter, more persistent AI-native future could actually look like.
Learning Fine-Tuning on Nvidia DGX Spark - Part 2
Troubleshoot Nvidia DGX Spark setup issues and improve your AI Workbench custom container with these step-by-step solutions for a usable environment