comma before a trailing phrase
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. 21 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 a comma that ends a full statement and is followed by a phrase adding more detail, like ", blazing a trail" or ", allowing developers". It tends to be negative on similar commas in technical process descriptions. Turning it up gave text that did not clearly carry this pattern forward.
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.
-
aschke⏎University News Service⏎September 26, 2011⏎Victor Frankenstein’s infamous monster led a brief, tragic existence, blazing a trail of death and destruction that prompted mobs of angry villagers to take up
-
(Minnesota) for three years before her marriage to Dr. Will Kennicott, of Gopher Prairie. She is a rebel against ugliness and conformity, and one factor in her decision to accept Kennicott is the opportunity to make
-
Roman Catholic faith and speaking a medieval Hungarian dialect, they still live in relative isolation, harvesting vegetables and nuts, some of which they exchange for oil, rice and other
-
ello at the Covent Garden theatre. The prospects of fame, fortune, and freedom were very appealing to young Booth, and he displayed remarkable talent from an early age, deciding on a career in the theatre by the age of 17. He
-
as developers navigate through various programming languages and frameworks to create powerful and intuitive applications.⏎The Significance of Coding in App Development⏎Coding is the backbone of app development, allowing developers
-
in mythology, the stars⏎Orion is one of the easier to recognizable and noticeable constellations. It is perhaps the most splendid of all constellations, befitting for a venerated character from ancient times. Named for a hunter
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: For the first of two years, I have for the first of two years for the first of two of them for the first of two of them for the first of two of then. For the first of two of then for the first of
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
amount of Ar-Cl2 or N2-Cl2, pure Ar, N2 gas into the bottom of the aluminum melt through the high-speed rotating rotor, and the impeller is fully chopped to form dispersed bubbles to make the aluminum liquid
-
oth surface, based on the generation principle of spherical involute tooth surface, the reverse motion of the generated motion of the tooth surface is taken as the cutting motion, and the reduced spiral tooth surface with spherical involute tooth shape can be milled
-
rotary movement to the linear movement as far as possible. Based on the analysis of the tooth surface cutting movement, the revolution and rotation of the blade are further analyzed, as shown in the figure.<|endoftext|>Source: The Independent – independent.co.uk⏎
-
all about making a 'tray' and covering it with paper.⏎My cousin was 24 years old on Friday April 5th in 1974. On what day⏎of the week was she born?⏎A tetrom
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
- L15:184 excites
- L9:207 suppresses
- L9:46 suppresses
- L18:233 suppresses
- L18:1103 excites
- L11:840 suppresses
- L10:370 weakly suppresses
- L10:155 weakly excites
- L20:2998 weakly excites
- L10:873 weakly excites
- L20:510 weakly excites
- L9:281 weakly excites
- L11:1602 weakly suppresses
- L20:476 weakly excites commas that lead into a continuation
- L7:92 weakly excites
- L20:1466 weakly suppresses
- L8:498 weakly suppresses
- L19:2058 weakly suppresses
- L11:223 weakly excites
- L2:763 weakly excites
- L18:768 weakly excites
- L18:843 weakly suppresses
- L7:90 weakly suppresses
- L10:662 weakly excites
Feeds later layers
- L24:140 weakly excites
- L22:1425 weakly excites
- L22:609 weakly excites
- L23:424 weakly excites
- L27:1134 weakly suppresses
- L27:1620 weakly excites
- L24:1443 weakly suppresses
- L23:1639 weakly excites
- L27:1628 weakly suppresses
- L25:591 weakly excites
- L23:772 weakly excites
- L26:252 weakly excites
- L27:410 weakly excites
- L22:506 weakly suppresses
- L24:2880 weakly suppresses
- L25:376 weakly excites
- L23:2082 weakly excites
- L23:1155 weakly suppresses
- L25:1425 weakly suppresses
- L24:2425 weakly suppresses
- L22:2204 weakly excites
- L26:733 weakly excites
- L25:744 weakly suppresses
- L23:2249 weakly 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
- being
- being
- erville
- тер
- uffling
- enance
- averse
- czy
- ère
- /Main
Suppresses
- лим
- 那就是
- IFICATIONS
- aturas
- 那么
- |#
- arden
- atura
- иф
- 那就