digits inside numbers and statistics
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. 3 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 reacts to a single digit or decimal point in the middle of a number, such as the "0" in 1703 or the "9" in 1999, mostly in years, percentages and statistics. Turning it up did not produce a clear theme in the model's writing.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
1655, 1655, 1734, 1650, 1733, 1703 and 1707 are included under the topic Early Cox History in
-
these tests have excellent detection rates (~99%) with very low false-positive rates (~0.2%). In other words, these tests are about 99.0% sensitive and 99.8% specific.⏎With performance like that
-
a priority of ideological harmony, multidimensionality, association and synthesis of previous rationalities for a harmonious civilization and a harmonious person.⏎2. Instead of the multivolume expressions accessible solely to the old age “spiritual aristocrats”
-
008, at $94.81 per barrel. That is not far below the average high of 2008, at $99.67. In between was a crash in oil prices — and most commodities
-
Commission data, at the traditional poverty line, the India-wide absolute number of poor was 323 million in 1983, 320 million in 1993-94 and 302 million
-
forces cut off U.S. access to hevea--and during the 1970s.⏎Guayule grows much faster than hevea, but the latter still has certain advantages.⏎Hevea is less prone to cracking
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: In the name of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
width. Hence the text:⏎<P> This is another paragraph, with a sentence that is split over several lines.Will get formatted as:⏎This is another paragraph, with a sentence that is split over several lines.<|endoftext|>Insufficient vocabulary, comprehension
-
th century by the King Rut Trak Varman. A French historian, Mi Bri No said that the temple is made of laterite stone. Phnom Da, Takeo is not enclosed by any wall and faces northwards.⏎There are
-
a voluntary standard and partly encompasses non-renewable technologies like nuclear and “clean coal.”4⏎And the state recently repealed “Energizing Indiana,”5⏎its energy efficiency program that supported nearly 1,500 jobs and saved
-
Accusative Case Accusative Case = Direct Object The direct object receives the action of the verb; it is whats being verbed. I buy a CD. Who or what is being bought?⏎Word changes in English In English, some words
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
- L0:39 weakly excites 'the' right after 'On' in 'On the other hand'
- L2:2328 weakly suppresses
- L0:1535 weakly suppresses
- L2:2696 weakly excites
- L1:59 weakly excites 'the' after a verb or preposition
- L1:2340 weakly suppresses
- L0:2407 weakly suppresses
- L2:3046 weakly suppresses
- L1:1100 weakly suppresses
- L0:545 weakly suppresses
- L0:1811 weakly suppresses
- L1:1662 weakly excites
- L2:2834 weakly excites
- L1:17 weakly suppresses
- L2:2747 weakly suppresses
- L2:1036 weakly excites
- L0:2806 weakly suppresses
- L2:1409 weakly suppresses
- L0:235 weakly suppresses
- L1:2300 weakly excites
- L2:1279 weakly excites
- L1:316 weakly suppresses
- L1:81 weakly excites
- L0:1034 weakly excites
Feeds later layers
- L4:155 strongly suppresses
- L8:73 strongly suppresses
- L4:559 strongly excites
- L5:119 strongly suppresses
- L5:44 strongly suppresses
- L6:169 suppresses
- L10:350 suppresses
- L5:42 suppresses
- L7:85 excites
- L6:879 excites
- L8:82 excites
- L8:484 excites
- L5:88 suppresses
- L7:106 suppresses
- L15:528 suppresses
- L7:27 suppresses
- L5:107 suppresses
- L4:5 suppresses
- L12:243 suppresses
- L5:1 suppresses
- L6:258 suppresses
- L6:212 suppresses
- L8:208 suppresses
- L10:35 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
- atures
- пит
- тки
- igor
- ered
- pering
- atorio
- vis
- headed
- 骖
Suppresses
- ��
- 他们
- 他们是
- 逞
- 他们会
- 会觉得
- 暧
- favourable
- 让他们
- denominator