comma after a clause's verb or subject
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. 22 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 comes right after a verb or a subject, where the sentence is about to continue with a new clause, such as "and, later" or "who, while she lived". Turned up, the model produced repeated "with" phrases, and turned down it produced strange word chains.
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.
-
3200 BC, when there was a small settlement on the bank of the Nile where Karnak now stands. The great temple complex at Karnak is, however, mostly a Middle Kingdom creation. Archaeological excavation reveals that the complex was
-
the people of Novgorod to rule that city (c. 862); the Rurik princes maintained their control over Kievan Rus and, later, Muscovy until 1598.⏎Rurik’s
-
and the wife of a future prime minister of the United Kingdom, Primrose was in her day the wealthiest woman in Britain, and an astute political observer who, while she lived, played a key role in advancing her husband’s career.⏎Like
-
Hebrides that Columba turned his face when he accepted the terrible penance of Molaise. He bade farewell to his relatives, and, with a few monks who insisted on accompany him whithersoever he might
-
amusement.⏎There was a half-decayed log lying by the roadside, and Grant sat down on it, calmly pulled a cigar from his cigar case, and, requesting one of his staff who stood nearby to furnish him a sheet of paper,
-
interests would be better served, he continued, if Japan instead of Russia eventually dominated Korea.⏎Kennan urged the secretary to reign in General Douglas MacArthur, who, he wrote, had “wide and relatively uncontrolled latitude… in determining our policy
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: to the left of the initial then the then then then then then then then then then then then then then then then then then then then then then then then then then then then then then then then then then then then then then then then then
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
has allowed me to have fun in life, to go to school, and to make a choice in life. I can watch a movie, or I don’t have to. I could hang out with my friends or I could hang out with
-
and answering the questions using the text.⏎Please read the sheet below ‘All About Lighthouses’ and then answer the questions below. Please remember, your answers need to be in full sentences and using your cursive handwriting and correct punctuation (cap
-
felt animals are chosen, another list would drop down showing the bag with 50 animals and then, the bag with 100 animals. The same can be done for handmade coats. You might
-
toward protecting family members from the risks of fire. This includes installing smoke detectors, keeping fire extinguishers charge, and establishing an emergency plan in case family members need to exit the house quickly if a fire starts.⏎However, fewer people know how
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
- L18:572 weakly suppresses
- L7:92 weakly suppresses
- L21:456 weakly suppresses
- L20:2977 weakly suppresses
- L16:1059 weakly excites
- L21:583 weakly suppresses
- L18:2476 weakly suppresses
- L21:2516 weakly suppresses
- L16:941 weakly suppresses
- L13:685 weakly suppresses
- L2:629 weakly suppresses
- L21:109 weakly excites
- L14:2733 weakly suppresses
- L7:1510 weakly suppresses
- L15:184 weakly suppresses
- L11:840 weakly excites
- L16:3001 weakly suppresses
- L17:2336 weakly suppresses
- L11:223 weakly suppresses
- L10:873 weakly suppresses
- L17:2831 weakly suppresses
- L19:1306 weakly excites
- L6:1280 weakly suppresses
- L21:1339 weakly excites
Feeds later layers
- L24:148 excites
- L27:603 excites
- L23:984 excites
- L24:116 suppresses
- L26:852 excites
- L23:756 excites
- L25:806 excites
- L24:883 excites
- L23:664 excites
- L26:741 weakly excites
- L26:2350 weakly excites
- L23:326 weakly excites
- L26:708 weakly excites
- L24:396 weakly excites
- L23:1590 weakly excites
- L24:921 weakly excites
- L25:1186 weakly excites
- L24:1173 weakly excites
- L24:1690 weakly excites
- L24:784 weakly excites
- L23:570 weakly excites
- L24:492 weakly excites
- L24:1711 weakly excites
- L27:927 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
- swith
- with
- without
- with
- WITH
- avec
- -with
- without
- through
- With
Suppresses
- JSGlobalScope
- 他又
- entonces
- 他曾
- 時候
- 我当时
- 他是
- then
- 骢
- amient