void.
Notes from inside several confused computers — Engineering, Engineering Leadership, AI, Python, Essay, Open-source, and the occasional opinion that wandered in off the street.
Bananas beat Paris.
A reader's comment on the last embedding post sent me back to check EmbeddingGemma on two backends, llama.cpp on the GPU and FastFlowLM on the NPU, against the reference model. Both were wrong, in different ways.
Field Engineer's Guide to Prompt & KV Cache Tuning
Notes from tuning the prompt behind an Antigravity auto-mode gate. Prefix caches match bytes from the left and give no partial credit, so a label that shifts by one re-prefills the whole payload. The seven ways I broke it, the one still in the code, and the tests that keep them out.
Between a ROCm and a Hard Place
Picking a llama.cpp backend (Vulkan or ROCm) and a ROCm source (Fedora's packages or AMD's TheRock SDK) on Strix Halo (gfx1151).
FastFlowLM on AMD NPU: XRT vs HRX
Evaluating FastflowLM with HRX on Strix Halo NPU.
Nobody can tell. Everybody thinks they can.
Notes on how readers decide a piece of writing is human, why they get it wrong, and what to do about it.
The Root CORS Analysis
A field guide for server-side engineers: the levers, the trade-offs, and how to decide what is actually worth doing when it comes to Cross-Origin Resource Sharing (CORS).
Demystifying Auto-Mode: Intent Classifiers, Kernel Sandboxing, and the Frontier of Autonomous Agent Guardrails
Automating agent permissions solves "permission fatigue" but risks security. This analysis explores decoupled intent classifiers like Auto Mode, defense-in-depth via sandboxing, and emerging local guardrails like Shieldstral and AgentWorld.
No posts match "".