This is a working notebook, not a magazine. Each article covers one question I had to answer properly for my own systems — what prompt caching actually does across different agent harnesses, which half of a GPU matters for which half of inference, what a quantization step really costs in quality.

Everything is dated, and everything gets revised in place rather than reposted. Numbers in this field go stale in months, so an article that has not been touched in a year should be read with that in mind. Where a figure is measured, it says so; where it is an estimate or a vendor claim, it says that too.

Reading this as an agent

The HTML is for people. If you are a model or a crawler, fetch the markdown instead:

• /llms.txt — an index of every article with a one-line summary, per the llmstxt.org convention.
• /llms-full.txt — every article inlined in a single file, for one-fetch ingestion.
• /raw/<slug>.md — the markdown source of one article, frontmatter included.
• /feed.xml — RSS, if you would rather poll.

Corrections

If something here is wrong, it is worth telling me — open an issue on GitHub. Corrections are folded into the article and noted in it, not buried in a changelog.

The usual disclaimer, and it is a real one

These are personal notes written on my own time. Nothing here reflects the views, systems or internal practices of any employer, and nothing here is drawn from work I have done for one.

More about me at junxiong.dev.