"and" or "or" joining two items
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. 9 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 words "and" and "or" when they link two things in a pair, like "he and she" or "Mr. and Mrs.". When it was turned up, the model's sentences started repeating and leaning on "the weather was cold" type phrasing, which is not clearly about joining words.
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.
-
one be puzzled? Because there’s one he and there’s one she and life is pronominally simple.⏎But things get tricky when she and she or he and he or singular, non-binary they and they rent huts:⏎Sandra and
-
, T., Boesman, P., Olarte, L., Salaman, P. & Olaciregui , C.|⏎|Facilitator/Compiler(s):||Benstead, P., Capper, D., Isher
-
8-76-1Purchased with funds contributed by Mr. and Mrs. Isaac Tatnall Starr, 1928⏎LabelThis
-
, and other books of Jewish interest. Owing to his sympathy with the Reformers during
-
, and the test laboratories were built during the past five years.⏎Each laboratory is linked through
-
, and scan additional documents.⏎Should I transfer the rights?⏎When you create something unlikely to
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 have a new project. I have a person. I have a thing. This is the process. The method. The system. The way. The way. The way. I have a situation. and a feeling. and a thought
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
](verb): to make bare; strip; denude.⏎This fun word comes from the latin word “denudatus” which means “naked” or “ bare.”⏎With the days getting colder, the deciduous
-
7,000 births and the risk increases with maternal age. The conditions are associated with severe mental handicap and several physical defects. Most affected individuals die before or soon after birth and they rarely survive beyond the first year of life.⏎Today we
-
large degree age-related, and may also be more developed in highly trained racehorses.⏎|Number of pages||7|⏎|Journal||Veterinarija ir Zootechnika|⏎|Publication status||Published - 17 Oct 2
-
ies, even if abandoned, in ruined condition or transformed for other use, also served as symbolic mezuzahs to mark towns
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
- L4:57 strongly suppresses
- L5:880 strongly excites
- L8:70 excites 'and' between stacked superlatives
- L4:108 excites "or" offering a choice in web text
- L6:427 suppresses
- L3:1951 suppresses
- L0:313 excites "and" joining paired nouns
- L1:1250 suppresses
- L4:2363 excites
- L3:250 excites
- L5:548 excites
- L4:2270 excites
- L0:2146 suppresses
- L6:584 weakly suppresses
- L7:42 weakly excites
- L1:1710 weakly suppresses
- L2:2969 weakly suppresses
- L1:1597 weakly suppresses
- L8:1083 weakly excites "x" or "by" between two dimensions
- L4:1330 weakly suppresses
- L1:2120 weakly excites
- L6:1706 weakly suppresses
- L7:836 weakly excites
- L5:1469 weakly suppresses
Feeds later layers
- L17:572 strongly suppresses
- L13:131 strongly suppresses 'and' joining parallel phrases
- L18:870 strongly suppresses
- L16:376 strongly suppresses 'or' joining alternatives
- L15:727 strongly suppresses
- L10:74 strongly suppresses
- L12:51 excites
- L17:1537 suppresses
- L11:876 suppresses 'other' after 'among' in lists
- L17:2274 suppresses
- L15:2765 suppresses
- L16:2577 excites
- L17:2332 suppresses 'and' joining two clauses
- L17:3061 suppresses
- L14:1116 suppresses 'and' inside lists
- L16:2580 suppresses 'and' linking food or health nouns
- L15:1501 suppresses "and" joining two parts of a name or phrase
- L19:2080 suppresses ampersand joining two items
- L16:1669 suppresses
- L16:2442 suppresses
- L20:885 suppresses 'and' joining paired words
- L14:801 suppresses 'and' or 'or' joining list items
- L13:125 suppresses
- L25:1694 suppresses 'and' joining paired names or fields
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
- 质量和
- 企业和
- 时间和
- 建设和
- elle
- 和个人
- 意大
- 经验和
- and
- 平民
Suppresses
- /or
- rog
- 其它
- nbsp
- 珅
- /OR
- ients
- дей
- ifice
- reas