The notebook
Posts
Reviews of tools I've run for real, a monthly table of model prices, and notes from the homelab. Each one opens with its verdict.
Sample entries: titles and verdicts are placeholders showing the layout. Only the first review has a full sample page.
-
AI token market, November 2026
SummaryFrontier prices held still; the money moved to the small-model tier.
Input, output and cached-input prices per million tokens for the main hosted models, compared with October, with notes on batch discounts and context-length surcharges.
-
Serving a local model to a coding agent: Ollama, vLLM and llama.cpp compared
Use itvLLM if you have the VRAM to spare; Ollama if you'd rather not think about it.
Three ways to put an open-weights model behind an OpenAI-compatible endpoint on one 24 GB GPU, judged on what a coding agent actually needs: tool calls that parse, and latency you don't notice.
-
Moving the homelab from Docker Compose to k3s, and back again
SkipFor eleven containers on one box, k3s cost more evenings than it saved.
What I wanted from Kubernetes at home, what I got, and the short list of things I kept when I went back to Compose files and systemd timers.
-
AI token market, October 2026
SummaryCached input is now where the real discounts are; read the caching rules before comparing list prices.
The monthly price table, plus a worked example of what a typical agent session costs once prompt caching is counted properly.
-
Open WebUI as a shared front-end for self-hosted models
Worth a lookGood for a small team; plan your own SSO and backups before anyone relies on it.
Two weeks as the chat front-end for a handful of users: what the admin side handles well, and where I had to add my own reverse proxy and auth.
-
Aider after three months on a real codebase
Use itThe git-first workflow is the point: every change is a commit you can read and revert.
Notes from daily use on a mid-sized TypeScript and Go monorepo: where it saved time, where it confidently made a mess, and the conventions file that made the difference.
-
AI token market, September 2026
SummaryA quiet month for list prices; the changes were in rate limits and batch terms.
The usual table, with a new column for batch pricing and a note on how I collect the numbers.
-
Backups I actually test: restic, a cheap VPS and a monthly restore drill
Use itBoring, cheap and verified. The restore drill matters more than the tool.
The scripts, the schedule and the checklist I run on the first Sunday of each month, including the one time the restore failed and why.
-
Giving an agent a shell: three sandboxing approaches tried
Worth a lookRootless containers plus a network allow-list is enough for most work; none of them is set-and-forget.
Plain Docker, gVisor and a throwaway VM, compared on setup time, what escaped, and how much they slowed the agent down.