Posts

Hands-on reviews of AI tools, agents and self-hosted software, the monthly AI token market, and notes from the homelab. The verdict comes first in every one.

Sample entries: titles, dates, verdicts and figures on this page stand in until the archive moves over.

  1. AI token market, November 2026

    Verdict: mid-tier output prices fell again; the cheapest model per token is rarely the cheapest per finished job.

    Format
    Two wide price tables
    Unit
    USD per million tokens
  2. Restic and Backrest for homelab backups

    Verdict: keep it. Every test restore worked; the Backrest web UI is young and some settings still need the command line.

    Cost
    Free, plus storage
    Licence
    BSD-2-Clause, GPL-3.0
  3. Four local model runners, one GPU

    Verdict: Ollama for a shared team box, llama.cpp when you want every token a second, vLLM once several people hit it at once.

    Cost
    Free, plus power
    Licence
    MIT, Apache-2.0, proprietary
  4. Recabling the rack before adding a node

    Verdict: a labelled patch panel saved more time than the new switch did.

    Cost
    One weekend
    Parts
    Patch panel, short leads
  5. AI token market, October 2026

    Verdict: batch and cache discounts did more for real bills this month than any list-price cut.

    Format
    Two wide price tables
    Unit
    USD per million tokens
  6. Aider for a week of refactoring

    Verdict: strong on small, well-described changes; read every diff before it commits, because it will commit.

    Cost
    Free, plus model tokens
    Licence
    Apache-2.0
  7. Uptime Kuma as the homelab’s status page

    Verdict: running in five minutes and pleasant to look at; alert rules stay basic, so pair it with real metrics.

    Cost
    Free
    Licence
    MIT
  8. A UPS, NUT and a clean shutdown for three nodes

    Verdict: test the shutdown on purpose once. Mine failed the first time, and only the test found it.

    Cost
    The UPS, then free
    Licence
    GPL-2.0 (NUT)
  9. AI token market, September 2026

    Verdict: long-context surcharges are now the line to read first on any price page.

    Format
    Two wide price tables
    Unit
    USD per million tokens