line break after a section heading
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. 16 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 on the line break that comes right after a heading or label, such as 'ASJC Scopus subject areas' or 'Similar books and articles', just before a list starts. Turning it up made the model write in short broken lines, so the effect is hard to read.
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.
-
coated and graded-index-layer coated samples was performed and compared with experimentally derived films. Good agreement between model and experiment was observed.⏎ASJC Scopus subject areas⏎- Electronic, Optical and Magnetic Materials⏎- Ceramics and Composites⏎-
-
silkworm-BEVS is a promising alternative strategy to produce the functional and thermostable reverse transcriptase.⏎All Science Journal Classification (ASJC) codes⏎- Insect Science<|endoftext|>An account of the religion of
-
Seeking Help:⏎Ben Locke, Ph.D., a mental health expert serving at Pennsylvania State University,
-
Muscle Growth⏎One of the⏎key aspects of a successful exercise program for muscle growth is exercise
-
12|⏎|Journal||Classical and Quantum Gravity|⏎|State||Published - Jul 7 2003|⏎ASJC Scopus subject areas⏎- Physics and Astronomy (miscellaneous)<|endoftext|>Most computer users are familiar with the
-
and Surveys⏎- NARA Energy Literacy Principles Matrix (a bioenergy resource search tool for
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.
-
first⏎> place. If one is chasing through a data structure and finds a need to⏎> change something buried within it, it seems necessary to rebuild⏎> everything that includes the changed thing.⏎In general, values are immutable
-
⏎soluble vitamins obtainable no other way, and are even part of cell wall⏎structure. Oils and Fats can also have very negative effects on⏎Before the first permanent settlements, nutrient oils came from insects,
-
also describes key elements of the Java API⏎library, such as I/O, the Collections Framework, the concurrency⏎utilities, the stream library, networking, event⏎handling, AWT, and more. Three chapters provide a solid introduction⏎to
-
the Billund⏎headquarters, where the company employs approximately 120 designers.⏎The company also has smaller design offices in the UK, Spain, Germany,⏎and Japan, which are tasked with developing products aimed⏎specifically at these markets
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
- L15:81 suppresses line breaks inside running text
- L10:230 suppresses
- L12:1012 excites
- L7:239 excites
- L7:2925 suppresses
- L8:88 excites line breaks inside wrapped sentences
- L8:2432 suppresses line breaks in lists and wrapped text
- L10:1760 suppresses
- L10:2315 excites line break inside wrapped text
- L14:56 suppresses document ends and list line breaks
- L1:1279 suppresses
- L6:463 excites
- L9:234 suppresses line break inside hard-wrapped text
- L8:380 weakly suppresses
- L9:1515 weakly excites line break splitting a phrase
- L14:51 weakly excites colon that introduces a statement
- L14:191 weakly excites blank spot before a run-on number
- L2:489 weakly suppresses
- L7:2952 weakly excites line break mid-sentence
- L8:14 weakly excites
- L10:84 weakly suppresses
- L13:10 weakly excites
- L13:309 weakly excites
- L1:1340 weakly suppresses
Feeds later layers
- L17:1242 strongly suppresses
- L17:1564 suppresses
- L18:1996 suppresses
- L19:970 suppresses
- L20:2399 suppresses
- L19:901 excites
- L26:1309 excites
- L21:579 excites
- L18:795 excites
- L19:771 excites
- L24:1619 excites
- L17:1675 weakly suppresses
- L25:2204 weakly excites
- L17:2160 weakly excites
- L26:664 weakly excites
- L23:1723 weakly excites
- L27:1467 weakly excites
- L24:203 weakly excites
- L27:213 weakly excites
- L17:2285 weakly excites
- L27:860 weakly excites
- L25:1805 weakly excites
- L17:825 weakly excites
- L27:106 weakly excites
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
- DN
- isbn
- 纶
- DN
- _Do
- dn
- ^⏎
- sworth
- dư
- ENDED
Suppresses
- hä
- ScrollIndicator
- pher
- rections
- createQuery
- oples
- quired
- idd
- createCommand
- 偏偏