line break before a subheading
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. 6 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 the line break that comes right after a short bold-style subheading in web articles, such as "Lower your risk of heart disease". Turning it up did not give a clear, consistent change in the writing.
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.
-
highly conductive core. The supercapacitors are made of millions of nanometer-thick wires with two-dimensional material shells coating.⏎Reshaping mobile technology⏎The supercapacitors are also more durable. They don't degrade quickly over
-
people were the picture of health when they donated, later-in-life obesity, high blood pressure and diabetes can raise the risk of kidney disease.⏎Transplant surgery risks⏎"We're changing," says
-
dark loaves, brown rice and oatmeal, according to a Harvard study of nearly 43,000 men.⏎Lower your risk of heart disease⏎Eating
-
countries, including southern Zimbabwe, northern South Africa and eastern Botswana, which have already suffered heavy rains and mudslides.⏎Biodiversity and food security under threat⏎Mozambique’s National Institute for Disaster Management and Reduction states that about 1
-
feeding advocates insist that flavors in breast milk go beyond a simple enjoyment for the infant -- it might be important for health too.⏎Flavors Start in the Womb⏎"The fact that the flavor is passed to the<|endoftext|>Masdar City was a project
-
24 hours or longer. Late-onset asthma produces heavy inflammation of bronchial tubes and is the dominant cause of chronic asthma.⏎Differences in treatment⏎Discovering these two types of asthma reactions has caused doctors to modify their treatment plans
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 will not be going anywhere. I will be going to the city. So I will not go. But I will not buy a car, and I will not go on my own. And I will not be going to the city,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the visit of the headman Yodhājīva to the Buddha. S.iv.308.<|endoftext|>The New International Encyclopædia/Kinkajou⏎|←Kinic acid||The New International Encyc
-
class at WSU.⏎What textbook is this problem found in?⏎Our data indicates that this problem or a close variation was asked in Chemistry: An Atoms First Approach - Zumdahl Atoms 1st 2nd Edition. You
-
web bug) has changed and needs re-indexing.⏎On the first release containing the indexer<|endoftext|>Publications and Tools⏎Comprehensive School Reform and Student Achievement: A Meta-Analysis -- This report is available from the Center for Research on the Education
-
essa Gray, Head of the Division for Least Developed Countries (LDCs), Small Island Developing States (SIDS) and Emergency Telecommunications for the International Telecommunication Union (ITU), argued that increasing
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
- L5:2181 strongly suppresses
- L5:212 suppresses
- L3:371 suppresses
- L3:523 excites line break after a section heading
- L5:1034 excites
- L3:559 excites
- L3:373 suppresses
- L2:1706 excites
- L5:2313 weakly excites
- L0:1864 weakly excites
- L1:830 weakly suppresses
- L5:750 weakly excites
- L2:578 weakly excites
- L4:226 weakly suppresses
- L5:343 weakly excites
- L5:2468 weakly suppresses
- L4:1698 weakly suppresses
- L5:918 weakly excites
- L1:1518 weakly suppresses
- L4:887 weakly suppresses
- L4:864 weakly suppresses
- L5:1371 weakly excites
- L1:1069 weakly excites
- L1:496 weakly excites
Feeds later layers
- L9:257 excites
- L7:392 excites
- L10:1516 excites
- L10:1136 excites
- L11:1344 excites
- L7:1729 excites
- L8:1684 excites
- L10:494 excites
- L16:818 excites
- L10:963 excites
- L7:1356 excites
- L17:856 excites
- L11:1513 excites
- L10:471 suppresses
- L17:450 suppresses
- L7:2235 excites
- L24:1823 excites opening parenthesis before an acronym
- L17:689 excites
- L8:666 excites
- L11:729 weakly excites
- L9:351 weakly excites
- L12:2494 weakly excites
- L10:900 weakly excites
- L12:2261 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
- imoto
- odium
- ibase
- oid
- رون
- íst
- ؤ
- ọc
- icont
- оля
Suppresses
- ()});⏎
- ()]);⏎
- validationResult
- хран
- 豉
- oor
- atisfied
- .standard
- urret
- hears