Accuracy, shown

Every number traces to the paper.

WatchRecap won’t put a figure on screen it can’t quote from your source. Here’s a real explainer of Attention Is All You Need — and the exact line behind each number it shows.

The finished explainer. Source: Vaswani et al., 2017 · arXiv:1706.03762.

The receipts

Every number the film shows, next to the sentence it came from. 3 of 7 beats put a figure on screen; each one cites the paper. The rest carry no numbers, so they need no citation.

On screen
Old approach: one word at a time
No number on screen — nothing to cite.
On screen
Sequential steps can’t be split up
No number on screen — nothing to cite.
On screen
Attention: every word sees every word
No number on screen — nothing to cite.
On screen
Base model: 12 hours on 8 P100 GPUs → beats all prior published systems (EN→DE)
Source · p.2✓ verified
“trained for as little as twelve hours on eight P100 GPUs”
On screen
Big model: > 2.0 BLEU above best prior results, incl. ensembles
Source · p.8✓ verified
“outperforms the best previously reported models (including ensembles) by more than 2.0 BLEU”
On screen
Faster when sequence length < representation size
Source · p.6–7
“faster than recurrent layers when the sequence length n is smaller than the representation dimensionality d”
On screen
Attention Is All You Need
No number on screen — nothing to cite.

How it stays honest

Quote or cut

Every on-screen number must be backed by a verbatim line from the source. If it can’t be quoted, it doesn’t go on screen.

Checked against the source

Numbers are matched to the source text before the storyboard is yours to approve — guesses are left out, not smoothed over.

You approve first

Nothing is voiced or animated until you sign off on the storyboard, every figure and its citation in front of you.

Make one from your paper Back to home