letters inside spelled-out words
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. 20 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 a single letter or syllable in the middle of a word that has been split up or spelled out, such as the letters in 'C F O S U' or 'P-R-A-C-T-I-C-E', and on odd spots in non-English scripts. Turning it up did not produce a clear theme.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
知多少⏎Pin Yin Version Chun mian bu jue xiao Chu chu wen ti niao Ye lai feng yu sheng Hua luo zhi duo shao⏎Inspiration Lifetime Wrote about his lackadaisical
-
- To bring to a focus; to focalize; as, to focus a camera. 2⏎The word "focus" uses 5 letters: C F O S U.⏎No direct anagrams for focus found in this word list.⏎
-
Biologics in the United States Chapter 16RISK MAnAgEMEnT FOR DRugS AnD bIOLOgICS In THE unITED STATES Arcoxia, Vioxx, Acomplia, Z
-
The basic⏎source of all minerals is the hot magma in the⏎interior of the earth.⏎(more content follows the advertisement below) A D V E R T I S E M E N T⏎When magma cools,⏎
-
most recognizably in the Hebrew name of The Jewish Agency – הַסּוֹכְנוּת הַיְּהוּדִית, or simply הס
-
those questions.)⏎What’s that? You don’t understand what we mean by selecting answer A? It’s OK. You’re going to P-R-A-C-T-I-C-E and get very comfortable with the test format so you don’t have to
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 a 3420, and I have a 3666. The 3666 is 2136, and the 3420 is 2096. The
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Subhash Ch<|endoftext|>BröstuligaOld Norse Dictionary - bröstuliga⏎Meaning of Old Norse word "bröstuliga" (or brǫstul
-
Marie Barnave⏎Barnave, Antoine Pierre Joseph Marie (äNtwänˈ pyĕr zhōzĕfˈ märēˈ bärnävˈ) [key], 1761–93
-
ogd. γrīw (in Manichean and Nestorian script γryw, in Soghdian script γryw(h), γrʾyw(h) “self, body, soul” (Gharib,
-
ְדָה (vÿ’ikkavÿda), the Niphal cohortative; coming after the
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
- L19:2464 excites
- L18:1012 suppresses
- L19:2076 excites
- L19:2081 excites
- L18:2405 excites
- L18:1457 excites
- L19:1133 suppresses last word before end of document
- L14:1424 excites
- L19:1502 suppresses
- L18:632 suppresses
- L19:1096 suppresses
- L16:1262 weakly suppresses
- L19:1663 weakly excites
- L13:148 weakly suppresses
- L16:104 weakly suppresses
- L15:2839 weakly suppresses
- L14:2706 weakly excites
- L2:105 weakly excites
- L14:1575 weakly excites
- L17:2257 weakly suppresses
- L12:1110 weakly suppresses
- L18:86 weakly suppresses
- L19:1491 weakly excites
- L14:1914 weakly suppresses
Feeds later layers
- L21:12 suppresses
- L24:367 suppresses
- L25:5 suppresses
- L23:3 suppresses
- L22:29 suppresses
- L24:7 suppresses
- L26:547 suppresses
- L27:386 suppresses
- L27:222 suppresses
- L21:142 suppresses
- L24:4 suppresses
- L27:0 weakly suppresses joins inside names and decimal digits
- L25:409 weakly excites
- L27:39 weakly suppresses
- L21:1149 weakly excites
- L27:112 weakly suppresses
- L27:135 weakly suppresses
- L27:78 weakly suppresses
- L22:7 weakly suppresses
- L24:85 weakly suppresses
- L25:11 weakly suppresses
- L27:81 weakly suppresses
- L26:116 weakly suppresses
- L22:370 weakly suppresses digits inside years like 1976
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
- :result
- ':[
- ектив
- '){ ⏎
- ){ ⏎
- ーム
- -results
- '; ⏎
- ')⏎⏎
- réponse
Suppresses
- yan
- ire
- lang
- iane
- na
- ern
- long
- ra
- irs
- h