comma before 'right?' tag
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. 24 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
Fires on a comma that comes just before a casual tag like 'right?' or 'does not it!', where the writer checks agreement with the reader. Turning it up made the model repeat the word 'right' over and over.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
side of the moon. Its mission? To take pictures of this unexplored part of the moon, collect soil samples, and plant a mini garden. Awesome, right? So how successful has the Chang’e
-
. Nothing travels faster than the speed of light, but even light takes time to get to places. It takes a ray of sunlight
-
it.⏎Anyway, what about the link between a woman’s menstrual cycle and the lunar cycle? I mean, that’s why we’re all still reading this post, right?⏎Well, let’s start off with a bit of backstory: For generations
-
learning areas is simpler for you to collaborate with, and will certainly use up much less of your time than searching for individual worksheets.⏎This sounds like a horrible lot, does not it! A great collection of preschool worksheets should cover every one of this
-
and LeBron James refer to themselves by their own name, it's called illeism, and it can actually be a useful psychological technique.⏎Recently, Grammar Girl listener Mark
-
: ‘do not’ would change to ‘don’t’⏎That’s it, you’ve got it…. Now write these sentences again but add an apostrophe in
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: how to best serve the needs of a school can be considered as a critical reappearance of the national can be considered as a critical reappearance of the national? - - - - - - - - - - - - - - - -
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
of the countless missing links that should have numbered in the millions.⏎J. L. B. Smith, posing with the second coelacanth caught off the Comora Islands in 1952.⏎This coelacanth fossil,
-
0. It connected the San Miguel district to Malate and Ermita.⏎The bridge also served Isla de Convalescencia, the island in the Pasig River where the orphanage of Hospicio de San Jose is located.⏎Information
-
at Madrid, 27 March, 1625. He was a great-grandson of the Tordesillas who was put to death by the Comuneros at Seville. He studied in Spain and Italy, and became secretary
-
1254), as well as hospitals at Pontoise, Vernon, and Compiégne.⏎St. Louis was a patron of architecture. The Sainte
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
- L20:2335 excites comma ending an 'if' clause
- L17:3046 suppresses answer labels and tag questions
- L21:1538 suppresses
- L20:1371 suppresses comma inside a question
- L23:1431 excites
- L19:650 weakly suppresses
- L15:271 weakly excites
- L11:1624 weakly excites
- L7:1064 weakly excites
- L20:1922 weakly suppresses
- L17:2456 weakly suppresses
- L12:731 weakly excites
- L20:1453 weakly suppresses
- L23:1028 weakly suppresses
- L22:697 weakly excites
- L20:2490 weakly excites
- L12:1237 weakly excites
- L19:1036 weakly suppresses
- L8:2204 weakly suppresses
- L2:231 weakly suppresses
- L19:1274 weakly suppresses
- L13:2326 weakly excites
- L2:2206 weakly suppresses
- L13:335 weakly suppresses
Feeds later layers
- L27:1296 strongly excites
- L26:1121 excites
- L27:1489 excites
- L27:1760 excites
- L25:1754 excites
- L26:930 weakly suppresses
- L27:1592 weakly excites
- L25:373 weakly excites
- L25:1843 weakly excites
- L25:1156 weakly excites
- L26:406 weakly excites
- L27:1212 weakly suppresses
- L27:302 weakly excites
- L26:2397 weakly excites
- L27:2800 weakly excites
- L25:651 weakly excites
- L27:2914 weakly suppresses
- L25:2798 weakly suppresses
- L25:2067 weakly suppresses
- L25:1656 weakly excites
- L27:1261 weakly excites
- L25:2211 weakly excites
- L26:1896 weakly suppresses
- L26:1820 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
- right
- right
- Right
- Right
- →right
- (right
- RIGHT
- 右
- -right
- ,right
Suppresses
- 呢
- ivr
- 嗎
- 망
- 吗
- 厚度
- moz
- �数
- 低迷
- .GetInt