-
A passive GPU in a tower: VRAM hit 106 °C while the core throttled
My new 32 GB passive GPU had no plateau under LLM load and its memory reached 106 °C. Measurements, the limits that were silently ignored, and the fan, power cap and clock limit that fixed it.
-
Read, write, critical: permission tiers for a local AI agent
Every tool call my local AI agent makes is rated read, write or critical. Writes need my approval, critical actions need a one-time code bound to that exact operation, and unattended runs only propose.
-
ComfyUI on demand: systemd socket activation for a GPU container
Idle ComfyUI held up to 2.8 GB of VRAM my LLMs needed. A systemd socket now starts it on the first request and stops it after 10 idle minutes, and a guard makes room for image jobs.
-
Vision-model OCR that keeps the layout: what to trust it with
A vision model called a bold italic paragraph 'not bold, not italic'. My OCR now lets the model read, Tesseract place small text, and pixel measurement judge typography: 99% precision on italic words.
-
Why local LLMs could not drive an autonomous pentest agent
I tuned an autonomous security-testing agent until the mechanics were clean, then watched my local 27B model read 1.44 million tokens and find none of three planted bugs. Here is how to separate fixable plumbing from a model that cannot do the job.
-
Translating whole books with local LLMs: the failure modes
My local translation service rebuilds scanned books page by page in another language. Most fixes were not about the model: segment context, glossary scope, a lost-update race and page rebuilding.
-
Typed decisions from logits: LLM judgements that cannot fabricate
Reading a 9B model's option probabilities instead of its text took a disambiguation task from 22.7% to 86.7%, added a usable confidence, and removed fabricated reasons. Most of the gain was wording.
-
An AI studio that makes films and 3D prints in Blender
My AI agents build printable models and short films in a sandboxed Blender over a tool protocol. The hard part is that a text-only model cannot see its own output, so it ships broken geometry as 'watertight' unless you give it sight.