"of" after a noun introducing a category
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. 7 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
Fires on the word "of" when it links a noun to what it belongs to, like "the religion of" or "the state tree of". Turning it up made the model write in a loose, wandering style; turning it down made it drift into maths word problems.
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.
-
cond 'snapshots' reveal a dramatic bond tightening in photo-excited gold complexes⏎More information: "From Binary<|endoftext|>Christianity is, officially, the religion of one third of the people in the world. The religion has spread across the world
-
ot observed more than a century ago. S. I. Hayakawa, the semanticist and, later, senator from California, called slang “the poetry of everyday life.” On the other hand, Ambrose Bierce defined slang as “
-
00 lifetime risk of contracting the disease, with the average age of detection being between 60 and 70 years.⏎“Cancer is a disease of the aging,” said Dr. Deborah Axelrod, associate professor at New York University
-
, if caught early it is easily treatable. Unfortunately, other types of prostate cancer are aggressive and can spread quickly.⏎Prostate cancer is a cancer of the prostate gland. The prostate gland is a walnut-sized gland present only in men
-
and reference site. Goethe Verlag has 100 free lessons for Hebrew language learners.⏎Is Hebrew a dead language?⏎Modern Hebrew is the official language of the State of Israel, while premodern Hebrew is used for prayer or study in
-
and attractive foliage.⏎Other People Are Reading⏎Eastern redbud's (Cercis canadensis) adaptability has earned it recognition as the state tree of Oklahoma. The quickly growing tree tolerates well-drained clay,
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: the area of a triangle is 10 cm. The area of a rectangle is 10 cm? No. So this. So this. So this. So this. So this. So this. So this. So this.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Bottom line is that you can’t go wrong with making blueberries a staple in your grocery cart.⏎Since blueberries are a fruit that are often sprayed quite generously with pesticides,
-
watering:⏎The Bucket Method⏎The bucket your tree came in holds about 10 gallons of water. You can use it to water your tree. Line it with a trash bag
-
. Troy L. Best is a professor of biological sciences at Auburn University.<|endoftext|>Snare drums are a type of drum that makes a distinct cracking sound when hit with a drumstick. They are also the type of drum on which the drum roll
-
the University of Bonn in Germany and colleagues. The team created a three-dimensional interference pattern by shining a laser into a ‘holographic crystal’, and used the pattern as a stencil
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
- L5:2937 weakly suppresses
- L4:2544 weakly excites
- L4:2732 weakly suppresses
- L5:117 weakly excites
- L2:3024 weakly excites
- L5:2873 weakly excites
- L5:2497 weakly excites
- L5:1158 weakly suppresses
- L3:870 weakly suppresses
- L2:1266 weakly suppresses
- L3:2060 weakly excites
- L3:952 weakly suppresses
- L1:2844 weakly suppresses
- L3:1055 weakly excites
- L3:2097 weakly suppresses
- L2:2117 weakly suppresses
- L0:2896 weakly excites
- L5:2582 weakly excites
- L5:760 weakly suppresses
- L6:1141 weakly excites
- L1:543 weakly excites
- L2:2576 weakly excites
- L0:2122 weakly excites
- L0:1975 weakly excites
Feeds later layers
- L12:2849 weakly suppresses '-based' after a place name
- L12:1906 weakly suppresses
- L11:437 weakly excites
- L8:2206 weakly excites
- L21:1886 weakly excites
- L9:955 weakly excites
- L8:1039 weakly excites
- L24:862 weakly excites
- L20:2591 weakly excites
- L17:1484 weakly suppresses
- L18:2467 weakly excites
- L8:127 weakly suppresses
- L20:2825 weakly excites
- L14:1853 weakly suppresses
- L13:167 weakly excites
- L9:1592 weakly excites
- L10:1611 weakly suppresses
- L14:2664 weakly suppresses
- L18:3068 weakly excites
- L19:1126 weakly excites
- L12:3030 weakly excites
- L15:2408 weakly excites
- L11:11 weakly suppresses
- L16:1121 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
- 滘
- estic
- 孳
- aurus
- -parent
- ệu
- hủy
- urrection
- нь
- ports
Suppresses
- schließen
- utilus
- plugins
- 该公司
- utschen
- ?」
- FileInfo
- USED
- ечно
- chili