Field journal
Blog archive
Every published note, with the newest work first. Use the topic links to stay inside one search intent.
My RTX 5090 runs show why model loading, prompt ingestion, token generation, and task checks need separate measurements before production use.
My RTX 5090 measurements show how I budget VRAM for a local model, a real workload, and the GPU processes that must stay resident beside it.
One local LLM speed number hides the work behind it. My RTX 5090 sweep shows why short generation, long context, and code need separate rates.
LM Studio shipped Bionic, a full agent built on open models with local code projects, voice, and document work. The interesting part is not the app. It is what.
Hugging Face ran its breach forensics on an open-weight model on its own hardware because hosted APIs refused the requests. Here is the lesson for builders.
A local model pass rate can be true and useless at the same time. Here are the three ways local LLM benchmarks mislead you, drawn from real rows on my RTX 5090.
I make my local LLM show its source plan before it writes. The preview gate catches weak retrieval while the fix is still cheap and easy to inspect.
My smaller local model existed and ran, but it did not beat the baseline. Here is the promotion gate I use before changing a working local AI route.
Claude Science keeps data on local systems but sends selected context to Claude. Here is the owned-GPU version, with measured RTX 5090 costs.
I tested four Ollama settings on Gemma 4 26B. num_batch 256 delivered nearly all the speed of 512 while using less power on my RTX 5090.
A 6.66 GiB ternary 27B model fit my RTX 5090, but it lost the default slot. File size is only the first local model-selection gate.
Agent logs show activity. They do not prove the outcome an agent claimed. Here is the verification layer I added after an agent reported two false dones.
I extracted the claims ledger that runs my agent operation into showwork. Falsifiable claims, deterministic checks, and an exit gate that refuses a false done.
Ollama's $65M round and 8.9 million developers make local AI a funded product category. Here is what changes for builders on owned GPUs.
The nightly queue sweep aborted for the fourth straight night. The root cause is a critical failure in the vault health check. I discovered a private path wo...
I tested a local-first writing pipeline on an RTX 5090. Gemma 4 writes the draft, deterministic checks catch errors, and frontier QA decides what ships.
q4km vs q5km is Q4_K_M vs Q5_K_M in GGUF. Pick by VRAM headroom, quality risk, and CPU fallback. Use the free quant tool before download.
Three Llama 3.1 8B runs missed a 400-word floor. Here is the verifier-driven route that moved long-form synthesis to Gemma 4 26B.
Blog charts need correct labels. I tightened layout geometry, char budgets, claims checks, and vision logs so a bad PNG cannot ship quietly.
I moved a vault search index from 1,536 to 768 dimensions. Here is why I rebuilt all 5,383 chunks instead of mixing vector spaces.
Preview the files your local LLM will receive before inference. This simple gate catches bad retrieval while the fix is still cheap and clear.
Change num_ctx in Ollama and it reloads the whole model. On my RTX 5090 that meant 140 seconds of dead air per context swap. Here is how to stop paying it.
Changing context size can reload a local model before every request. A measured RTX 5090 sweep shows why context belongs in the routing key.
Claude and Codex both went dark. Here is the tertiary Gemini path I wired, with honest qa_reviewer stamps.