ellipsis dots after a cut-off quote
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. 25 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 the first dot of an ellipsis (". . .") where a quote or sentence trails off mid-thought. When it was turned up, the model wrote long strings of dots.
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.
-
over time. In the early church, believers focused upon genealogies so much that Paul had to warn his readers twice (!) not to “give heed to .
-
up the surface with thin layers of pigment, the color of the underlying layers modifying those on top. Painter Gino Severini reported that "Matisse said . . . that everything that did not contribute to the balance and rhythm of [this
-
15, 2003, the National⏎Transportation Safety Board (NTSB) made its boldest statement to date.⏎They stated that “ . . . the safe operation of fifteen-passenger⏎vans requires a knowledge and
-
and local adoption agencies follow here at home?⏎The present system of screening and approving textbooks dates to pre-war Japan: "Struggles over the national narrative existed . . . before and even during World War II, when official narratives such as the
-
ms we read, “Give thanks to the Lord, for he is good; his love endures forever. Let the redeemed of the Lord tell their story. . . . Let them give thanks to
-
lizards!⏎Ecologist Sean Doody discovered this while in the field with student Phillip Paull. The two were lifting eggs out of a crevice when . . . Holy Dooley! The hatchlings literally popped out and sprinted away!⏎
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 many times a couple wants to have a child 10 years away? How many times a couple wants to have a child 20 years away? How many times a couple wants to have a child 50 years away?
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
’t require any qualifications and is open to anyone.⏎The course includes videos of farmer John Geldard from Plumgarth’s Farm Shop and Café in Kendal, members of
-
.⏎Metsäpohjainen piha paransi päiväkotilasten immuunijärjestelm
-
. Indirect reciprocity is a social interaction in which one actor helps another and is then benefited by a third party. Gossip has also been identified by Robin Dunbar, an evolutionary biologist, as aiding social bonding in large groups.⏎The word
-
built environment, she is the author of Bulldozer: Demolition and Clearance of the Postwar Landscape, winner of the 2017 Lewis Mumford Prize for the best book in American planning history. She is particularly interested in
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
- L23:641 strongly suppresses
- L22:748 strongly suppresses 'and so on' list endings
- L22:701 strongly excites
- L22:2065 excites
- L18:2722 suppresses
- L18:2185 suppresses
- L17:2866 suppresses
- L19:1265 suppresses
- L17:2209 excites
- L18:2938 excites
- L16:3039 suppresses
- L21:2493 excites
- L23:1171 weakly suppresses
- L15:3052 weakly suppresses
- L18:742 weakly excites
- L19:831 weakly suppresses
- L16:76 weakly suppresses
- L4:1468 weakly suppresses
- L7:412 weakly excites
- L9:1533 weakly suppresses
- L18:1118 weakly excites
- L13:611 weakly excites
- L24:2698 weakly suppresses
- L21:475 weakly excites
Feeds later layers
- L27:1182 strongly excites
- L26:1825 excites the first dot of an ellipsis
- L27:409 suppresses
- L27:703 excites
- L26:481 weakly suppresses
- L27:2054 weakly suppresses
- L26:626 weakly excites
- L26:943 weakly excites
- L26:1453 weakly suppresses
- L26:887 weakly excites
- L27:2562 weakly suppresses
- L27:2222 weakly excites
- L26:2399 weakly excites
- L26:174 weakly excites
- L27:977 weakly suppresses
- L26:2489 weakly suppresses
- L26:2547 weakly suppresses
- L26:948 weakly excites
- L27:2989 weakly suppresses
- L26:2182 weakly excites
- L27:2343 weakly excites capital N starting a name or word
- L26:593 weakly excites
- L26:167 weakly excites
- L27:1627 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
- .
- /.
- (.
- `.
- ."
- .'
- ,.
- (.
- ._
- [.
Suppresses
- cd
- (cd
- ld
- cd
- →cd
- cds
- CD
- _cd
- CD
- CDs