"of" after a noun
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. 13 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 word 'of' when it follows a noun, as in 'the words of', 'the names of' or 'the cry of'. It is weaker on other small words like 'for', 'against' or 'at'. When turned up, the model fell into repeating short phrases like 'the leader a leader', so the steering does not show a clear story.
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.
-
all the way which I command you, that it may be well with you.'⏎Matthew 7:24-27 “Therefore everyone who hears these words of Mine and acts on them, may
-
establishments and missionary colonies (all under the authority of Columba), until in time the Gospel light was ablaze on the hills of Albyn; and the names of St. Columba and Iona were on every tongue from Rome to the utmost
-
11 children. He led an exemplary life, bearing constantly in mind his mother’s words: “I would rather see you dead at my feet than guilty of a mortal sin.” His biographers have written of the long hours he spent in
-
et maîtriser les dépenses : propositions de l’Assurance Maladie pour 2014 – Rapport au Ministre chargé de la Sécurité Sociale et au Parlement sur l’évolution des charges
-
the simple story of a village life is depicted in an extra-ordinary way in this creation of his. The title of the story is kept by the name of the village only. So the title truly
-
case or motion being tried or heard and the laws applicable to it.⏎(2)Either the outline of a case made in preparation for a class, or a written argument for Legal Writing
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. The, I. I. 1. 1. 1 This: I. 1. 1 1 1 3. 3 3 3
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
family studies⏎Langley, Elizabeth Dale, "Parenting Styles and Perceptions Across Christian Denominations" (2011). ETD Collection for Pace University. AAI3467374.⏎Remote User:
-
familiar with it, U.S. law defines human trafficking as “the use of force, fraud, or coercion to compel a person into commercial sex acts or labor against their will,” and it brings in an estimated $150 billion a year
-
in the town. The population density was 1.2 people per square mile (0.4/km²). There were 106 housing units at an average density of 1.1 per square mile (0.4/k
-
wrong<|endoftext|>External Web sites⏎Britannica Web sites⏎Articles from Britannica encyclopedias for elementary and high school students.⏎- Sen
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
- L4:4 strongly suppresses
- L7:6 suppresses
- L10:132 excites
- L8:517 suppresses
- L5:763 suppresses
- L10:469 suppresses 'of' after a noun phrase
- L10:709 suppresses
- L1:403 excites
- L10:105 suppresses
- L9:2417 excites
- L6:101 suppresses
- L5:1053 suppresses
- L4:2275 suppresses
- L3:232 excites 'on' after a noun like 'impact on'
- L6:357 suppresses
- L5:5 suppresses digit after a label like paper or type
- L4:16 excites
- L12:2736 suppresses
- L6:655 excites
- L12:234 excites
- L9:1025 weakly excites
- L7:943 weakly suppresses
- L8:705 weakly excites
- L12:205 weakly suppresses 'of' after an abstract noun
Feeds later layers
- L17:385 strongly excites
- L17:303 strongly excites
- L17:1122 excites
- L18:663 excites
- L18:293 excites
- L17:1507 excites
- L20:283 excites
- L17:146 suppresses period after abbreviations and dates
- L18:1135 excites
- L16:2666 excites
- L25:286 excites
- L14:143 excites 'the' inside journal and institution names
- L16:412 excites
- L22:489 excites
- L15:316 excites
- L17:292 excites
- L17:175 excites 'to' after a noun or verb of motion
- L17:137 excites
- L15:381 excites
- L18:93 excites
- L19:1464 excites
- L17:2302 excites 'in' after a noun
- L22:160 excites
- L17:1835 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
- いう
- ゜
- compat
- >>::
- _));⏎
- })(
- mọi
- }}>
- mut
- ationToken
Suppresses
- CTION
- ibly
- PTION
- ulent
- worth
- fähig
- -through
- RIES
- 在整个
- 还会