line break in the middle of a sentence
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. 2 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 a line break that cuts a sentence in two, the kind you see in text that was wrapped at a fixed width (like old printed articles). It is most active where the sentence is clearly unfinished at that point. Turning it up gave text that does not obviously relate to this.
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.
-
policy shifts in member states,⏎such as Greece, Italy the United Kingdom and Ireland. The Multireader⏎project examined complex navigation through digital materials, and the⏎TEDUB project covered a similar area for graphical information
-
Site 48TE414 is in the South Jenny Lake development. The site was⏎recommended as eligible for the National Register of Historic Places. The⏎National Park Service and
-
and heavens shall be replaced by a new earth and new heavens; a new⏎world will then come into being. All mankind shall be raised up there and shall⏎be rewarded or punished on the basis of their beliefs and deeds.”⏎In this article
-
.⏎The ceiling limit for a family consisting of five members was fixed as 30⏎standard acres. For every additional member of the family consisting of more than⏎five members, an additional extent of
-
99⏎African American Nurses have long history with the American Red -⏎Undeterred by initial rejection, Davis persisted and in l9l7 became the⏎first
-
plantation after getting permission from the Land⏎Board under section 31 of the Act.⏎Ceiling laws were enacted and enforced actually in two phases, the earlier⏎
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: $$ \begin{array}{l} \sum_{i} = \sum_{i} $$\sum_{i} $$\sum_{i} $$\sum_{i} $$ $$ \
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
which ids full of⏎violence and hatred, wreaking havoc on their towns to express their beliefs.The Oklahoma City bombing brought home to American people that no country⏎is immune from terrorism.⏎Although thought to be tied into larger extremist⏎groups,
-
the⏎notes are a summary and a detailed analysis of the⏎poem's plot, theme, genre, purpose, context, references and techniques. The original⏎text appears after the notes. For "deep explorers" of Anglo-Saxon
-
<|endoftext|>Lecture by Francesca Russello Ammon, Associate Professor of City & Regional Planning and Historic Preservation, Weitzman School of Design, University of Pennsylvania⏎By 1965, nearly
-
sharing, and rapidly publishing scholarly content on the Web.” It is not simply a Web site that hosts textbooks, articles, and courses. A non-profit run by⏎“Most textbooks are a mass of information in linear format: one topic follows after
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
- L1:167 excites
- L1:1143 weakly excites
- L0:107 weakly suppresses newline starting a new block
- L0:145 weakly suppresses
- L1:292 weakly excites
- L0:1199 weakly excites
- L1:685 weakly excites
- L1:2862 weakly suppresses
- L1:2228 weakly excites
- L0:2669 weakly suppresses
- L0:1510 weakly suppresses
- L1:1278 weakly suppresses
- L0:3042 weakly suppresses
- L1:863 weakly excites
- L1:203 weakly suppresses
- L1:2206 weakly suppresses
- L1:1960 weakly excites
- L1:2313 weakly suppresses
- L0:1880 weakly excites
- L1:1279 weakly suppresses
- L1:2564 weakly excites
- L1:2724 weakly suppresses
- L1:1737 weakly excites
- L0:224 weakly excites
Feeds later layers
- L3:1901 excites
- L8:2674 weakly suppresses
- L24:1683 weakly suppresses
- L17:242 weakly excites
- L8:380 weakly excites
- L22:2867 weakly suppresses
- L9:589 weakly suppresses
- L7:1665 weakly suppresses
- L27:365 weakly suppresses
- L3:1454 weakly suppresses
- L10:1399 weakly excites
- L10:1086 weakly suppresses
- L4:1749 weakly suppresses
- L22:167 weakly excites
- L8:896 weakly suppresses
- L3:1036 weakly excites
- L3:1592 weakly excites
- L5:342 weakly excites line breaks inside wrapped sentences
- L23:135 weakly excites
- L9:2361 weakly suppresses
- L15:939 weakly suppresses
- L12:911 weakly suppresses
- L24:935 weakly suppresses
- L11:925 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
- &,
- distilled
- lectic
- comings
- kest
- вели
- 过敏
- utas
- учет
- łość
Suppresses
- واب
- ibre
- wings
- اث
- Directorate
- dav
- Titanium
- athed
- ье
- Sheets