naming words for terms and labels
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 fairly consistent: several of them 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 words that name or label something, such as name, keywords, identifiers, and symbol, usually where the text is about what a thing is called. Its strongest negative cases are grammar terms like conditional and superlative, and turning it up gave no clear theme.
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.
-
” or “https://”) is highly recommended. This allows to trace the information online, even if it becomes outdated.⏎The citing of DOIs (digital object identifiers) is encouraged.⏎Citing the date when an online work was consulted is now
-
this asteroid for more than a decade,” said Dante Lauretta, principal investigator for the mission at the University of Arizona. “We look forward to having a name that is easier to say than (101955) 1
-
remedied. Sign up now while there is still space in this special course!⏎Lessons cover issues such as the acknowledgment of God, Marriage, Life, Obscenity, Immigration, Parents’ Rights, Elections, Criminal Law, and the citing
-
.⏎Both episodes were probably taken into account when the monks of St Basil originally selected a name for their leper hospital. The name eventually gave rise to the terms "lazar", meaning a leper, and
-
process of improving the visibility of a website or webpage in search engine results.SEO considers how search engines work, what people search for, the actual search terms or keywords typed into search engines, and which search engines are preferred by their targeted
-
for so many errors have been introduced and ingrafted into their religion, by their ignorance, their separation from the Catholic Church, and their intercourse with Jews, Pagans, and Mohammedans, that their present religion is nothing but a kind of confused
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 first" is not a place, and the second is not a place. What is the function of the verb in this case? The sentence is: "The first is not a place, and the second is not a place.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
students are confused about verbs, nouns and adjectives. Imagine what happens once they begin to stumble across the imperfect subjunctive, the pluperfect and the conditional! ¡Flipan en colores!⏎The
-
," he is expressing a comparison price.⏎The effect that produces the adjective superlative communication consists of raising the quality of a noun. However, although the adverb very and the suffix isimo are very
-
(8) Had his father not died? (9) They had returned the home? (10)You had finished this book?⏎Past Perfect Tense Negative⏎Subject + had not + verb 3rd from + object.⏎Past Perfect
-
. The adverb "yesterday" makes it clear that the action was in the past. The
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
- L4:2995 suppresses
- L4:1040 suppresses
- L3:698 excites
- L4:2119 weakly excites
- L5:2177 weakly suppresses
- L3:694 weakly suppresses
- L5:1764 weakly excites
- L4:1874 weakly excites
- L5:1981 weakly excites
- L2:296 weakly suppresses
- L4:785 weakly suppresses
- L2:606 weakly suppresses
- L5:2145 weakly suppresses
- L4:2253 weakly suppresses
- L5:1163 weakly suppresses
- L5:341 weakly suppresses
- L4:906 weakly suppresses
- L3:1896 weakly excites
- L4:2894 weakly suppresses
- L5:1720 weakly suppresses
- L2:2404 weakly excites
- L0:926 weakly excites
- L2:1573 weakly excites
- L1:2478 weakly suppresses
Feeds later layers
- L8:2527 suppresses
- L11:1687 suppresses
- L10:1051 suppresses
- L8:2271 suppresses
- L9:1913 suppresses
- L8:1497 suppresses
- L7:2813 suppresses
- L11:1544 suppresses
- L7:983 suppresses
- L8:1085 suppresses
- L19:2158 suppresses
- L8:1007 suppresses
- L10:653 excites
- L7:2809 weakly suppresses
- L7:1150 weakly suppresses
- L10:592 weakly suppresses
- L8:2148 weakly suppresses
- L19:1913 weakly suppresses
- L7:2033 weakly suppresses
- L12:2197 weakly suppresses
- L18:1257 weakly suppresses
- L10:542 weakly suppresses word completing a technical term
- L18:1777 weakly suppresses
- L26:2770 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
- anning
- 服务能力
- 录制
- ATURE
- inaire
- vida
- ARING
- .ASCII
- 乾坤
- 神州
Suppresses
- sub
- いません
- implies
- (crate
- not
- 论
- 三是
- odem
- ��
- Authors