the space before a rounded 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. 19 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly 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 fires on the blank space just before a number that follows words like 'about', 'approximately', 'an estimated' or 'more than'. It is a small, specific pattern. Its steered outputs were unrelated, so they don't help explain it.
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.
-
world? About 3 billion people earn less than two dollars a day, but figures for the rest
-
approximately 24 hours after laparoscopic intervention, due to the residual CO2 gas
-
at about 64,000 miles per hour, hitting at Sutter's Mill
-
loprid and clothianidin, produced by Bayer, and thiamethoxam, produced by Syngenta.⏎There are an estimated 2.62 million bee colonies in the United States.⏎-Noel Brinker
-
and around 935, rebels defeated it and founded the Koryo dynasty. This
-
017).⏎Approximately 65,000 individuals in the U.S. died in 2
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: the question of whether the answer to a question is to be determined by the answer to the question of a question of a question... a question of a question of a question... question of a question of a question... Question of a question of
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
turn anti-clockwise and then forward 3. Did you find any treasure?⏎8. 3/4 turn anti-clockwise, back 1, 1/4 turn anti-clockwise, forward
-
110.00000000.00000000 IP: 10101100.0001000x.xxxxxxxx.xxxxxxxx Each subnet
-
13⏎Updated: March 13, 2013⏎Updated: October 02, 2013⏎At the outset, it must be stated quite equivoc
-
solutions of A X = Y 1 . Solution : X = V and X = V + Z f) Find another solution of A X = 3 Y 1- 5 Y 2 . Solution : X = 3 V
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
- L18:664 excites
- L18:439 excites
- L18:307 suppresses
- L18:2559 excites
- L18:2507 suppresses
- L18:1871 excites
- L18:548 excites
- L16:646 suppresses
- L18:761 excites
- L10:2348 weakly excites
- L18:140 weakly excites
- L3:115 weakly excites
- L18:152 weakly suppresses
- L12:539 weakly suppresses
- L15:432 weakly excites
- L18:1867 weakly suppresses
- L17:1051 weakly excites
- L17:2264 weakly excites
- L17:2271 weakly excites
- L13:855 weakly suppresses
- L16:1741 weakly suppresses
- L17:3066 weakly suppresses
- L17:795 weakly excites
- L18:1523 weakly excites
Feeds later layers
- L20:294 suppresses
- L24:363 suppresses
- L20:2065 suppresses
- L24:367 excites
- L21:12 excites
- L25:409 suppresses
- L20:524 excites
- L27:910 excites
- L26:732 weakly suppresses
- L25:5 weakly excites
- L21:224 weakly excites
- L27:625 weakly excites
- L22:225 weakly excites
- L20:1781 weakly excites
- L21:457 weakly excites
- L20:329 weakly excites
- L27:352 weakly excites
- L24:21 weakly excites
- L21:51 weakly excites
- L26:70 weakly suppresses
- L23:41 weakly excites 'at' after a verb or adjective
- L20:1189 weakly suppresses
- L27:289 weakly excites
- L26:315 weakly suppresses
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
- 发达
- -first
- prime
- رين
- perí
- BASE
- além
- abez
- segundo
- cancellationToken
Suppresses
- '..
- onn
- 干活
- onis
- 带着
- bans
- ivic
- ..."⏎⏎
- 的因素
- anism