Esc

Writing

home

Long-form notes on AI, LLMs, databases and building software. Most come with interactive demos, so you can try the idea instead of only reading about it.

  1. Your model doesn't need to write the answer. It needs to point at it.

    For classify and extract jobs the answer is already in your input. Why pointing beats writing, what ScaleDown's CDMs do with that idea, and where it stops working.

    read the note
    filed
    read
    9 min
    words
    1,803
    revision
    r1
    tags
    ai · llm · architecture
    note no.
    04 / 04by s.s.
  2. Your AI coding session gets pricier with every message. Here's why.

    How attention works, why it is quadratic, and what that does to your AI coding bill. Includes a simulator to find where your own sessions go wrong.

    read the note
    filed
    read
    7 min
    words
    1,418
    revision
    r1
    tags
    ai · llm · claude-code
    note no.
    03 / 04by s.s.
  3. How UUIDv7 fixes the primary-key tradeoff

    Auto-increment is cheap until you have two writers, and random UUIDs wreck your index. UUIDv7 puts the clock first and stops making you pick.

    read the note
    filed
    read
    5 min
    words
    900
    revision
    r1
    tags
    databases · postgres · uuid
    note no.
    02 / 04by s.s.
  4. Jev: the AI that doesn't talk (and why that's the point)

    Most models want to write you an essay. Jev answers with a number, fast and for a fraction of a cent.

    read the note
    filed
    read
    6 min
    words
    1,081
    revision
    r1
    tags
    ai · infrastructure · llm
    note no.
    01 / 04by s.s.