Signals
01

Watch a word stop mattering

Twelve numbers, real arithmetic, running in your browser. Type your own sentence, pull the retention down, and find the exact step where the first word stops reaching the end.

The sentence step 0 of 13 Tap a word to follow it. Brighter means it counts for more.

How much declaration still matters none yet
The measured influence of the selected word on the network's memory at each later step.

This is an untrained network of twelve numbers, so which word lasts longer is arbitrary. The point is that each word fades as later words arrive.

Show the 12 numbers
Columns: the words, as they are read. Rows: the 12 numbers that are all it remembers.
Heatmap of twelve hidden units across each step of the sequence.
−1 0 +1
A twelve-unit Elman network, h = tanh(ρ·Wₕh + Wₓx + b). Influence is measured, not asserted: the same sentence runs twice, once with the selected word nudged, and the curve is how far the two states drift apart.

A recurrent network reads one word, updates a small internal state, throws the word away, and reads the next one. Whatever it still knows about the start of a sentence has to survive inside that state, which never gets any bigger.

A fixed container, a stream that does not stop, and a rule about what gets overwritten. That is also a person four screens into a flow.

Nothing here is pre-recorded. Press Play, then drag the slider.