"of" after the word "number"
A neuronOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in layerOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. 12 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them peak on the same token. It is featuredA neuron picked for the full treatment: steering outputs and a written explanation. About 3,000 of the 86,000 neurons are featured., so it has steering outputs.
What this neuron does
This neuron lights up on the word "of" when it follows "number", as in "the number of" or "the number of counts". When it was turned up, the model's writing drifted toward counting and numbers, though the link was loose.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. Think it is wrong? Flag it (opens a prefilled issue on GitHub).
Fires most on
The text windows where this neuron's activationThe number a neuron produces for a token. Large positive means it fired hard; negative means it pushed the other way. was highest. Amber marks positive activation; the outlined tokenA piece of text the model reads at once: often a word, sometimes part of a word or a punctuation mark. is the peak. Hover a token for its strength.
-
, 2, and 1.8° per step.⏎Step angle = 360°/No of steps per revolution⏎Where m = Number of Phase winding on the
-
less prone to interference in comparison to RF technology.⏎DPI and FPS⏎DPI (dots per inch) and FPS (frames per second) are the number of counts in an inch of movement and the number of times the sensor reads the surface
-
this outbreak, I had occasion to review some clippings from<|endoftext|>A. Physical Characteristics⏎- How many feathers are on a turkey? An adult turkey has 5,000-6,000 feathers on its body,
-
the number of teeth on the Screw itself. A larger number of starts results in a more robust mechanism with a greater ability to transmit torque, but multiple ‘starts’ can sometimes be incompatible with a small ‘lead’ due to space limitations.⏎Lead
-
can now code this algorithm using Python to reveal the answer to the initial question: how many lego bricks are needed to build planet Earth!⏎Number of lego bricks⏎If you have completed this algorithm, the final answer should be approximatively
-
children per woman. This rate, which indicates the average number of children that women would have during their lifetime if they conformed to the current age-specific fertility
SteeringAdding a push along a neuron while the model writes, to see how its output changes. Turning it up or down.
Push this neuron up or down while the model writes, and watch what changes. The outputs were generated ahead of time, so nothing runs in your browser.
Here is something I have been thinking about lately: I have been thinking about how many I have been thinking about... Let's see, it's a function of... Let's say 2 to 3, which means 2 to 3, 2 to 3,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
what your right hand possesses. That is more likely that you will not be unjust.'⏎This verse has been interpreted to mean that Muslim men are allowed to marry up to four wives, but if they fear that they will not be able to treat them
-
Ask him to multiply the coin in the left hand by 17 and to say 'OK' when he has the answer. If it took about the same amount of time for each, then the nickel is in the right hand since 1
-
? after the 24th attempt?⏎Let’s assume that we have a more general problem. There are n questions, and a(n) is the smallest number of times we need to take the test to guarantee that we can figure out the
-
. This radium emits tiny amounts of radon into the atmosphere. On average, one part of radon is present in 1 x 1021 part of air.Radon is also present in some spring waters, including Hot
Connections
Show on mapWhich neuronsOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in earlier layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.
Fed by earlier layers
- L9:628 excites
- L11:1889 suppresses
- L10:2300 suppresses
- L9:1197 excites
- L10:978 suppresses
- L9:2690 suppresses
- L8:1029 suppresses
- L11:52 excites
- L10:68 excites
- L11:1455 suppresses
- L11:2003 suppresses
- L6:1849 excites
- L10:1954 excites
- L11:2818 excites
- L8:1939 suppresses
- L10:2531 excites
- L10:266 suppresses
- L10:1539 excites
- L9:1087 suppresses
- L11:979 suppresses
- L6:1767 suppresses
- L10:2791 excites
- L10:1618 suppresses
- L10:2063 suppresses
Feeds later layers
- L17:1727 strongly suppresses
- L20:1113 suppresses
- L17:1581 excites
- L17:1912 suppresses
- L19:1033 suppresses
- L22:784 suppresses
- L20:1528 suppresses blank spot before a number
- L16:2448 suppresses
- L20:1404 suppresses
- L20:595 suppresses
- L15:1535 excites
- L15:1521 suppresses
- L18:2694 suppresses
- L17:2711 suppresses
- L15:856 excites
- L16:2637 suppresses
- L21:756 suppresses
- L23:1404 suppresses
- L13:1629 suppresses
- L13:1564 excites
- L20:509 excites
- L17:2038 suppresses
- L14:2832 suppresses
- L17:1271 excites
Activation histogramA chart of how often a neuron takes each activation value.
How often this neuron fires at each strength. Almost always it does close to nothing (the tall bars in the middle); the snippets above are the rare moments it fired hard, amber for positive and blue for negative.
Logit lensReading which output tokens a neuron pushes toward or away from, straight off its output weights, before the later layers have their say.
The words this neuron nudges the model toward (amber) or away from (blue) when it fires. A hint, not a label: many neurons work through later layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. instead.
Promotes
- orque
- 材
- incent
- damp
- 墙壁
- 墙体
- Specs
- 比率
- walls
- yếu
Suppresses
- count
- ROWS
- ifferent
- etti
- .index
- illy
- count
- zähl
- idx
- _indices