inside web addresses and citations
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 is most active inside web addresses and citation codes, such as the "://" in "http://" or the dots in a DOI number. Turning it up made the model switch into Chinese text in several answers, which is a loose link to its strongest negative examples, which are Chinese characters.
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.
-
. Biological Conservation, 155. 50-58. 10.1016/j.biocon.2012.05.013Full text not available from this repository.⏎Although grass
-
JPL. The imaging operations center is based at the Space Science Institute in Boulder, Colo.⏎For more information about the Cassini-Huygens mission visit http://saturn.jpl.nasa.gov/home/index.cfm. The Cassini imaging team
-
to supports all kinds of powerful, dynamic applications and services. For more information on this great topic, please consult:⏎- Sun's Java Server Pages home (http://java.sun.com/products/jsp/)⏎- Westy Rockwell's article on "
-
receive a test report with each Dataforth SCM. Dataforth test instruments are calibrated and traceable to NIST standards. (see NIST info at: http://www.nist.gov/public_affairs/labs2.htm)View an example of
-
8), 1595-1600. http://dx.doi.org/10.1016/j.jbusres.2013.10.001⏎BERISHA-NAM
-
Birds Online (T. S. Schulenberg, Editor). Ithaca: Cornell Lab of Ornithology; retrieved from Neotropical Birds Online: http://neotropical.birds.cornell.edu/
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 can be something else, but I can be something else... ... I can be something... ( ... ) ... I can be something... ( ... ) ... I ... ( ... ) ... I can be something ... ( ... ) ...
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
ore 2-cleft to base.⏎One species: Mediterranean region; cultivated and adventive worldwide, including in China.⏎(Authors: She Menglan (佘孟兰 Sheh Meng-lan); Mark F. Watson)<|endoftext|>By Cary Gron
-
on Chinese Auxiliary Verbs for Thai Learners Based on the Comparative Analysis⏎- 人文学院-已发表论文 ⏎“助动词”在汉语中具有较大的作用,使用率高,但是泰国学生经常在学习和使用汉语
-
's longest sustained dynasties. Although it was founded by King Park Hyeokgeose, the dynasty was ruled by the Gyeongju Kim (김, 金) clan for most of its 992-year history.
-
heating and air conditioning systems, it also means that pollutants that originate indoors are getting trapped inside.⏎“People normally<|endoftext|>Spring Dawn by Meng Haoran 春眠不觉晓 处处闻啼茑 夜来风雨声 花
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
- L10:155 strongly excites
- L8:328 strongly excites
- L12:158 strongly suppresses
- L10:669 strongly suppresses
- L4:76 strongly excites
- L11:223 strongly excites
- L2:24 strongly suppresses
- L7:92 strongly excites
- L5:7 strongly suppresses
- L9:17 strongly suppresses
- L4:342 suppresses
- L1:25 excites
- L1:699 excites
- L10:1388 suppresses
- L11:416 excites
- L3:1256 excites
- L0:83 suppresses
- L3:130 excites
- L9:74 excites
- L4:398 suppresses
- L7:72 excites
- L9:1722 suppresses
- L10:39 suppresses
- L3:14 excites
Feeds later layers
- L15:59 strongly excites
- L15:184 strongly suppresses
- L17:611 strongly suppresses
- L15:390 strongly suppresses
- L18:66 strongly excites
- L17:56 strongly excites
- L14:744 strongly suppresses
- L16:513 strongly suppresses
- L17:1016 strongly suppresses
- L23:59 strongly suppresses
- L17:812 strongly suppresses
- L17:473 strongly suppresses
- L17:257 strongly suppresses
- L27:54 strongly suppresses
- L26:81 strongly suppresses
- L16:899 strongly suppresses
- L15:1033 strongly suppresses
- L16:771 strongly suppresses
- L19:68 strongly suppresses full stops closing a claim
- L17:98 strongly suppresses
- L17:535 strongly suppresses
- L27:269 strongly suppresses
- L19:119 strongly excites
- L21:279 strongly 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
- three
- four
- second
- seven
- five
- reason
- ten
- balance
- lot
- profit
Suppresses
- 分
- 在
- 考虑
- 以
- 是在
- 组织
- 安全
- 正
- 包含
- 都