'Sub' starting a word or name
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. 25 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 responds to the syllable 'Sub' at the start of a word or name, such as Subrahmanya, Subscribing, Subhash or SUBTRACTION. Turning it down made the model write words like 'subpar' and 'sub sub', so it seems tied to that spelling.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
to his native place Maṇali. Once he went to Tiruttaṇi, a nearby sacred place and practiced special spiritual disciplines, meditating on the deity Subrahmaṇya. A strange mystical experience here induced in him a remarkable power
-
. As the resulting jets of gas shoot backward, the engine and its craft are thrust forward. However, showing you is much cooler than explaining it.⏎Thanks for Subscribing!⏎You'll need to confirm your email, so please check your inbox
-
as Muttusvāmi Diksitar or just ‘Diksitar’. He was the son of Rāmasvāmi Diksitar and Subbalakṣmi Ammāl, born as a result of the grace
-
|⏎Koothus in Silapathikaram⏎by Subashini Pathmanathan⏎Ancient Tamils called their dances by the generic
-
aspirin-related products or acetaminophen which can react dangerously with prescription medications.⏎Treating Allergies When You Have Fibromyalgia⏎Dr. Sublett
-
, unfortunately none of which were investigated thoroughly by successive governments; leaving people in the dark about one of the most beloved leaders India has ever produced.⏎Life history of Subhash Chandra Bose⏎Subhash Chandra Bose was born on 23
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: let's take a moment to let go of the moment of momentale (momentale is the moment of momentale, momentale, momentale, momentale, momentale...). Let's take a moment to let go of the moment
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Product designing plays an important role in product selling. The company whose product is better and attractively designed sells more than the product of a company whose design happens to be weak and unattractive.⏎In this way, it can be said that the
-
performance of an app's launch time of less than one second to be excellent, less than 5 seconds to be good, and greater than 5 seconds to be poor.⏎Sometimes you also have to consider other metrics that don't have a direct
-
the tumour during surgery. If cancer cells are found at the edges (outer margins) of this tissue, it is called a positive surgical margin and the prognosis is less favourable. If cancer cells are not found there
-
have the most favourable outcome. Tumours that are in later stages (they have spread into nearby lymph nodes or organs far from the bile ducts) have a less favourable prognosis.⏎Location of the tumour @(Model.HeadingTag)>⏎T
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
- L21:1246 excites
- L22:151 suppresses
- L19:2331 weakly suppresses
- L20:796 weakly excites
- L8:12 weakly excites
- L24:105 weakly suppresses
- L1:1528 weakly suppresses
- L18:81 weakly suppresses
- L0:2877 weakly suppresses
- L9:464 weakly suppresses
- L18:1805 weakly suppresses
- L14:1129 weakly suppresses
- L21:3033 weakly suppresses
- L22:400 weakly excites
- L17:515 weakly suppresses
- L21:2182 weakly excites
- L15:1361 weakly excites
- L23:143 weakly suppresses
- L3:2297 weakly suppresses
- L21:1214 weakly excites
- L17:3063 weakly suppresses
- L21:271 weakly suppresses
- L0:2546 weakly suppresses
- L20:287 weakly suppresses
Feeds later layers
- L27:1001 strongly suppresses
- L26:2794 weakly suppresses
- L27:1156 weakly suppresses
- L27:908 weakly suppresses
- L27:1128 weakly suppresses
- L27:1107 weakly suppresses
- L27:903 weakly suppresses
- L26:3069 weakly suppresses
- L27:364 weakly suppresses
- L26:2062 weakly suppresses
- L26:1462 weakly excites
- L26:1436 weakly suppresses
- L27:1889 weakly suppresses
- L26:1800 weakly suppresses
- L26:1962 weakly suppresses
- L26:1746 weakly excites
- L26:2906 weakly excites
- L27:1944 weakly suppresses
- L27:2818 weakly excites
- L26:2228 weakly excites
- L27:945 weakly suppresses
- L26:1845 weakly suppresses
- L26:1833 weakly excites
- L26:1000 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
- Webpack
- cro
- ve
- ynet
- amientos
- vier
- Stre
- ieres
- éta
- igr
Suppresses
- sub
- sub
- Sub
- Sub
- _sub
- .sub
- 子
- =sub
- (sub
- →sub