Frizzle

Take the ride

We’ll ask Frizzle one question, watch the model work out the answer layer by layer, then change a single neuron and send it somewhere else.

The first time you open Frizzle, your browser downloads the 722 MB model. After that it loads from your device.

1. Ask the question

Open Frizzle and press Load model. When it’s ready, type this, leave Seed at 1, and press Run:

What is the capital of France?

The model replies: The capital of France is Paris.

2. Choose a moment

The model writes one piece at a time, so we can stop at any point and look inside. We want the moment just before it says Paris.

Click is in the reply, then press Ride ↗.

The Frizzle page with the question typed in and the reply The capital of France is Paris, marked after the word is, above a Ride button.
Click is. The words after it fade, because that’s what we’re about to look inside.

3. Drive through the layers

The model has 26 layers, and every word passes through all of them. Each glass pane on the road is one layer. The word on it, also shown on the sign, is the layer’s top guess: what the model would most likely say if it stopped there.

Press Drive, or the space bar. To go at your own pace, use the ‹ › buttons or arrow keys, scroll, or click a stop on the line under the road.

The Ride at layer 3: a coach on a night road facing a glass pane with the word devenue on it.
The early layers haven’t got anywhere yet. At layer 3 the top guess is devenue.

4. Watch Paris arrive

For most of the drive the guesses make no sense: S, devenue, een. Then, at layer 17, Paris appears and stays all the way to the end. A guess glows when it’s a word the reply goes on to use.

The Ride at layer 17: the word Paris glows on the pane, and the panel on the right lists neuron 4949 as the strongest in the layer.
Layer 17 is where the answer forms. The panel on the right lists this layer’s strongest neurons.

The points of light on each pane are that layer’s strongest neurons. Go back to layer 17 by clicking its stop on the line under the road, and look at the list on the right: #4949 is at the top.

5. Swap to the Deck

Click #4949 in the list, then choose Deck at the top right.

The Ride shows what the model is saying. The Deck shows what’s behind it: all 6,912 neurons of a layer, laid out as a card, brighter the stronger they are. You stay on layer 17 with #4949 still selected, outlined on the card. The numbers under the deck move to other layers.

The Deck showing layer 17 as a grid of glowing neurons, with neuron 4949 selected in the panel on the right.
The same moment on the Deck. Use Ride and Deck to swap whenever you like; your layer, selection and pins come with you.

6. Pin the neuron

A pin holds a neuron to a value you choose. In the panel, type -4 in the box and press Scale ×. That multiplies the neuron by −4, turning its +8.15 into about −32.6.

The pin appears at the top of the window, and on the card. Nothing has changed yet: a pin takes effect on your next ride.

The Deck with layer 17 marked 1 pinned, the panel reading Pinned scaled by -4, and a chip at the top reading L17 #4949 times -4.
Neuron 4949 is pinned, ready for the next ride.

7. Ride with the pin

Choose Ride to swap back. Your neuron now has a brass dial on layer 17, and the main button has changed to Ride with 1 pinned. Press it.

The Ride at layer 17 with a brass dial beside the pinned neuron and a button reading Ride with 1 pinned.
Back on the Ride, with the dial on the neuron you pinned.

8. Take the fork

Frizzle writes a new reply from the same point, with your pin held, and drives you up to the place it changes. The road splits at layer 17. The original road carries on to Paris on the left. Yours passes through London, then Berlin.

The road forking at layer 18: Berlin glows on the pane ahead, while Paris stays on a dimmer road to the left. The route below splits in two.
One neuron, and the model is now heading for Berlin. The line under the road shows the split too.

9. Read what it says now

Drive to the end. The last layers settle on **, the mark that makes text bold: the model is about to write Berlin in bold. The full reply then appears under the road:

The capital of France is **Berlin**. However, it’s important to note that this is a common misconception! Officially, the capital of France is **Venice**.

The end of the ride: Paris glows on the left road, two asterisks on the right, and the new reply about Berlin and Venice appears below.
Same question, same seed, one neuron changed.

Why ** and not Berlin? The model writes in small pieces called tokens, and each ride follows one of them. ** comes first and Berlin comes next. Click the ** in the reply to ride on from there and watch Berlin arrive.

Keep going

These results come from one run on one machine. Graphics cards can round numbers slightly differently, so yours may not match word for word. If the road doesn’t fork, try a different neuron near layer 17.