unclear: mixed fragments in lists and pronunciations
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. 27 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
The strongest triggers are scattered: pieces of pronunciation guides, list separators, and single letters or words in odd spots. The negative side is ordinary connecting words like 'on' and 'of' inside phrases. I could not find one clear pattern.
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.
-
. That form hasn’t yet become the standard, though, and there’s a lengthy list of variations still kicking around in the wild, such as ay-ight, aiite, and even the occasional aaaaaaaight. As the dictionary indicates
-
for the disorder by using techniques called amniocentesis (pronounced am-nee-oh-sen-TEE-sis) or chorionic (pronounced kawr-ee-ON-ik) villus sampling.1,
-
the norms of the Sogdian language, this name could not be transformed from a word vihara for which in the Sogdian used the word βrx’r [Lurje, 2004, p
-
usually start by having all of the students count from 1-31 with me.⏎- Then I have them only repeat the sounds of the counters: st, nd,rd, and th.⏎- Draw a baseball diamond on the board.⏎
-
in Manichean Mid. Pers. script gryw, rarely gryyw; in Sogd. script kryw; in Pahl. script rendered by the ideogram, q.v., CWLE) meaning “neck,
-
146, García González, 1979)⏎3rd sing fem – la⏎3rd sing masc – lu⏎3rd mass masc – lo⏎see also: http://pidweb.ii.uam.es/coser
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:__________(___________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
, has captivated readers with her profound exploration of nature and the human experience. Her poems often evoke a sense of wonder and contemplation, inviting readers to reflect on their own lives. In this argumentative essay, we will delve into one of
-
United States, and ushered in a new era of hope. Martin Luther King Jr. led the civil rights movement to new heights, ultimately leading to the signing of the “Civil Rights Act” of 1964. The whole world
-
who led the legal battle to end segregation, one case and one courtroom at a time.⏎The Supreme Court’s decision inBrown v. Board of Educationis widely considered a seminal point in the battle to end segregation, but it was in fact the
-
70 CE he led a campaign that culminated in the capture and destruction of Jerusalem. Titus became the Roman emperor in 79. He completed construction of the Flavian Amphitheatre, better known as the Colosseum,
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
- L22:14 strongly suppresses
- L14:140 strongly excites
- L19:1174 strongly suppresses
- L20:59 strongly excites
- L17:15 strongly suppresses
- L19:124 strongly excites
- L20:433 strongly excites
- L26:1564 strongly suppresses
- L19:430 strongly suppresses
- L9:974 strongly excites
- L19:1133 strongly suppresses last word before end of document
- L19:963 suppresses
- L12:2582 excites
- L6:917 excites comma after a month and day
- L6:42 excites
- L22:16 excites comma in dates and birth lines
- L18:246 suppresses
- L16:1421 excites
- L18:76 excites blank spot before a number
- L24:77 suppresses opening parenthesis after a name
- L4:2129 excites
- L24:39 suppresses
- L14:102 suppresses
- L16:396 excites
Feeds later layers
None: this is the last layer.
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
- polyline
- lul
- viper
- stal
- trom
- taxi
- subj
- rune
- ly
- fools
Suppresses
- DP
- [$
- FG
- UN
- AB
- MS
- PD
- MD
- CP
- AK