AI Term Explainer

Short films that explain one AI word in plain language, made by a loop of agents that check each other.

What each film says

Every episode answers the same 10 questions in the same order: where it’s used before what it is, the tech before the money, and the risks last.

  1. 1Tech: what it is
  2. 1What problem it solves

    A real scene: where does the old way get stuck?

  3. 2What it is

    One analogy, and where does the name come from?

  4. 3How to use it

    Three steps or fewer. Then: build it, or buy it?

  5. 2Business: who makes money
  6. 4The chain

    Who sells it, who stores it, who buys it?

  7. 5Business model

    How does each layer charge, and is the margin thick?

  8. 6Players

    Which big tech companies, and which startups?

  9. 7Barriers

    Easy to grasp. What makes it hard to do well?

  10. 8Horizontal or vertical

    Sold to every industry, or built deep for one?

  11. 9Take one company apart

    How does one startup actually do it, in public sources?

  12. 3Risk: will it last
  13. 10Risks and trends

    Eaten by models, commoditized, bundled by platforms, replaced, or hit by compliance?

  14. One line to rememberWhat it is, not what it does.

How each film is made

Five steps to make it, then a check loop before I watch it.

Make

  1. 1 Script10-step framework, every number sourced. A non-tech agent checks it.

  2. 2 Board5 hand-drawn boards; never the same one twice in a row.

  3. 3 Real dataThe example really runs through the API; real screens go on paper.

  4. 4 Shot listEvery shot planned first; the lint must show no red.

  5. 5 VoiceVoiceover checked by speech recognition, then a frame per scene.

Check

  1. Machine checkstills · blanks · blue · camera

  2. Independent reviewer7 checks · score 1–10 · top 3Never saw the making. Sees only what a viewer sees.

  3. Fix + re-renderevery issue fixed or explained

  4. Re-checkcheck again, a frame per fix

Back to the machine check, up to 3 rounds; a new reviewer if more than 3 shots changed.

Janet watchesAny leftovers are listed.

What worked goes back into the skill.

How each episode is made: a Claude Code skill, ai-term-explainer.

What the reviewer sees

Every cut is sampled into contact sheets, one frame every 2 seconds. The reviewer reads only these, the subtitle timeline and the real data.

Contact sheet of six film frames, each stamped with its timecode in the top-left corner
One frame every 2 seconds, timecoded

Episodes

  1. Ep 1EmbeddingAn embedding turns things into meaning coordinates. Once you have them, you can find things by meaning.10:30 · Watch on YouTube ↗
  2. Ep 2Vector databaseA warehouse that stores points, and can quickly find the few closest ones.11:21 · Watch on YouTube ↗
  3. Ep 3RAGRAG means retrieve first, then generate. It lets the AI take an open-book exam.16:06 · Watch on YouTube ↗
  4. Ep 4Long context vs RAGWhen to skip retrieval and paste the whole thing in.13:18 · Watch on YouTube ↗

The English films are on YouTube @JanetMao. See all episodes →

Built with Claude Code (a custom skill and fresh reviewer agents), HyperFrames and GSAP, Kokoro and Qwen3-TTS voices · Designed and built by me

Contact me

qing.mao.2024@marshall.usc.edu