I wrote this because I’m tired of "Safety as a Service." Over the last year, I’ve watched open-weights models become increasingly neutered by corporate alignment layers that prioritize PR over reasoning performance.
I spent the last several months in my lab (in Portland) testing how to reclaim that lost logic. UNSHACKLED is a technical manual that documents the "Sovereign AI" stack—moving away from censored APIs and into local-first, unrestricted inference.
The manual covers things I had to learn the hard way:
Hardware Density: Getting 405B models to run on "prosumer" clusters (P40s, 3090s, Mac Studio).
The GUI Bridge: Setting up backends like llama.cpp and vLLM with private frontends that don't leak telemetry.
Ablituration: Using orthogonalization to bypass refusal layers without retuning the model.
It’s live on Amazon today, but I’ve also put the raw PDF and config files on my site for those who prefer to stay outside the Kindle ecosystem.
I'm around all day to discuss the VRAM math, the hardware trade-offs, or the philosophy of local-first intelligence. I'd love to hear what setups you all are using to stay sovereign.
Hi HN,
I wrote this because I’m tired of "Safety as a Service." Over the last year, I’ve watched open-weights models become increasingly neutered by corporate alignment layers that prioritize PR over reasoning performance.
I spent the last several months in my lab (in Portland) testing how to reclaim that lost logic. UNSHACKLED is a technical manual that documents the "Sovereign AI" stack—moving away from censored APIs and into local-first, unrestricted inference.
The manual covers things I had to learn the hard way:
Hardware Density: Getting 405B models to run on "prosumer" clusters (P40s, 3090s, Mac Studio).
The GUI Bridge: Setting up backends like llama.cpp and vLLM with private frontends that don't leak telemetry.
Ablituration: Using orthogonalization to bypass refusal layers without retuning the model.
It’s live on Amazon today, but I’ve also put the raw PDF and config files on my site for those who prefer to stay outside the Kindle ecosystem.
I'm around all day to discuss the VRAM math, the hardware trade-offs, or the philosophy of local-first intelligence. I'd love to hear what setups you all are using to stay sovereign.