the gap right before a 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. 24 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
The neuron fires at a spot in the text where a number seems to be missing or was stripped out, such as between 'Hebrews' and '9:22' or 'equal to' and '252 calories'. Steering did not give a clear picture: the outputs drifted into unrelated text, commas, or brackets.
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.
-
law purged with blood, and without shedding of blood is no remission” (Hebrews 9:22). To purge is to “get rid of whatever is imp
-
L. Three Special Limits 1. sin lim 1 x x x = 2. 1 cos lim x x x -= 3. ( ) 1 lim 1 x x x e + = Strategy for Finding Limits:
-
.⏎- If we want to write this in differential form, which is useful when we do our u-substitution,⏎- we could say that du is equal to 5 cosine 5x.⏎- Now when you look over here, we don
-
hours of high-definition video in about a cup of DNA.⏎There is a lot of digital information in the world about three zettabytes' worth (that's 3000 billion billion bytes) and the constant influx of new digital content
-
battery is intended.⏎Btu (British Thermal Unit) — The amount of heat required to raise the temperature of one pound of water one degree Fahrenheit; equal to 252 calories.⏎Charge Controller — A component of a phot
-
du could be equal to -- so du/dx (derivative of u with respect to x),⏎- well we just use the chain rule. Derivative of 5x is 5
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: when I read a book that has a very long text, they often use several symbols called symbols called symbols called symbols called symbols called symbols called symbols called symbols called symbols called symbols called symbols called symbols called symbols called symbols called symbols called symbols called
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
cultural and economic forces combine to create a 24-hour society in which millions of Americans—either chronically or intermittently—obtain insufficient sleep as a result of workplace and lifestyle determinants. A convincing
-
of name and historical aspects⏎Erica plukenetii was named by Linnaeus in honour of Leonard Plukenet (1641-1706), who was a physician by profession but also a botan
-
espasianus Augustus, original name Titus Flavius Vespasianus, (born Dec. 30, 39 ce—died Sept.
-
, Sisley), he attended from 1875 training in the painter-glassblower Alexandre Mauvernay workshop (1810-1898) who worked in
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
- L9:7 excites
- L21:173 excites
- L20:563 excites
- L18:48 suppresses
- L14:348 suppresses
- L20:499 excites
- L17:38 suppresses
- L18:93 suppresses
- L16:527 suppresses
- L22:424 suppresses
- L23:80 excites blank slot before a year or number
- L15:1601 weakly excites
- L10:33 weakly suppresses
- L14:1091 weakly excites
- L19:131 weakly suppresses
- L21:390 weakly suppresses
- L1:8 weakly excites
- L14:51 weakly excites colon that introduces a statement
- L23:262 weakly suppresses
- L19:119 weakly suppresses
- L15:44 weakly suppresses
- L15:123 weakly suppresses
- L14:775 weakly suppresses
- L5:5 weakly suppresses digit after a label like paper or type
Feeds later layers
- L25:239 suppresses
- L26:17 suppresses
- L25:15 excites
- L26:56 suppresses
- L25:146 suppresses
- L25:222 suppresses
- L27:627 suppresses
- L26:21 suppresses
- L27:448 suppresses
- L26:199 suppresses
- L26:254 suppresses
- L27:558 suppresses
- L26:263 suppresses
- L26:48 excites
- L26:503 weakly suppresses
- L25:95 weakly excites
- L26:352 weakly suppresses
- L27:145 weakly excites dots inside web addresses
- L27:83 weakly suppresses
- L25:30 weakly excites
- L27:732 weakly suppresses
- L25:232 weakly excites
- L27:817 weakly suppresses
- L27:184 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
- -Mart
- -script
- -ing
- -note
- £o
- -fashion
- -ce
- -db
- -ground
- i
Suppresses
- ris
- respons
- vis
- kills
- skills
- weighted
- отнош
- strings
- CString
- guided