A handwritten digit becomes a neural network
A visual-first neural-network lesson that turns one rough digit into a layered, animated model.
Field notes / why it works
The first 6:00 makes a familiar handwritten digit into a precise problem by 0:49, then maps its 28 × 28 pixels to 784 input values at 3:05–3:45. The strongest replay peak in the sampled span, 3:32.8–3:44.0, sits around that conversion, and the 5:05–5:30 propagation sequence turns the abstract network into a visible test.
Plate 01 / The format in one breath
A single familiar visual example becomes an apparently easy task, then an engineering problem. The narrator promises to build a mental model from first principles while a continuous black canvas grows from one object into a labelled, animated system.
Why it works
- A low-resolution handwritten digit appears by 0:08, so an abstract topic starts with a thing viewers can recognize before any terminology arrives (0:04–0:26).
- The same object changes roles: image → 28 × 28 pixels → 784 input values → network layers (0:51, 3:05, 3:45). Each new definition has a visible referent.
- The excerpt's replay peak at 3:32.8–3:44.0 surrounds the conversion from pixel grid to first network layer; the visual step appears to be a point viewers inspect again (inferred).
Plate 02 / Format card
| Measured excerpt / visible treatment | |
|---|---|
| Platform · aspect · length | YouTube · 16:9 · 1280 × 720 at 60 fps · first 6:00 sampled from an 18:40 video |
| Pace | 1 continuous designed canvas, 0 detected hard cuts, 36 within-canvas step changes = 6/min; first change 1.35 s. A 360 s “shot” here describes cut detection, not a still frame. |
| Script | 1,097 transcribed words; 185 wpm across runtime, about 218 wpm while speaking; 44 sentences averaging 24.9 words; 3 questions; 38 number mentions. |
| Voice | Adult, lower-register, conversational teacher; fast but unhurried in tone, animated pitch. |
| On screen | 100% motion graphics / drawn diagrams in sampled six minutes; no presenter or footage visible in the supplied sheets. |
| Captions | No verbatim subtitles. Small diagram labels, numbers, bracket annotations and occasional editorial text accompany the spoken point. |
| Sound | Dry narration; no sustained music bed in the measured excerpt. −20.1 LUFS integrated, −2.1 dBFS true peak. |
| Structure | Familiar example → unexpected difficulty → teaching promise → minimal definition → input/output/layers → animated test → next question. The six-minute sample stops during the next explanation, with no end CTA. |
Pro / 13 plates + recipe
The first plates are open. The full dissection is Pro.
Get the timed shots, type and sound specs, and the runnable recipe.
Open the full blueprint- Plate 03 / The hook, frame by frame (first 12 seconds)Pro
- Plate 04 / Structure (beat sheet)Pro
- Plate 05 / Shot grammarPro
- Plate 06 / Visual style guidePro
- Plate 07 / Voice and deliveryPro
- Plate 08 / Sound designPro
- Plate 09 / PackagingPro
- Plate 10 / Script templatePro
- Plate 11 / Production planPro
- Plate 12 / PromptsPro
- Plate 13 / Edit recipePro
- Plate 14 / QA checklistPro
- Plate 15 / Make it yoursPro
- Recipe / runnable promptPro