line break in the middle of a sentence
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. 4 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 weakly responds to a line break that cuts a sentence in two, the kind you see in text copied from a PDF or a narrow column. Its negative side is weak and not clearly meaningful. Turning it up gave scrambled text, so it tells us little about what it does.
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.
-
any of the other auberges were situated on a piazza, is not⏎certain but shall be discussed later on.⏎Unfortunately, only three of⏎Cassar's seven auberges have survived. The German aub
-
- Shin-to ('the way of the gods'), as opposed to Butsu-dō ('the way of Buddha')." (2)⏎On the left is a very⏎small detail from a pentaptych by Hokusai
-
was responsible for the defence of the new city. At the same time, the⏎hostels had to have a central location.⏎Cassar's predecessor and⏎teacher, Francesco Laparelli, had already planned to build each auberge
-
⏎review new products and technologies, select a particular supplier,⏎get details on a specific application, or access a library of solar⏎“Much of today’s advanced⏎design solutions require more than a list of part numbers and⏎suppliers in stock
-
Sensee River after capturing⏎the villages of Cherisy and Vis-en-Artois.⏎The 3rd Division met with⏎stubborn opposition but succeeded in capturing Bois de Vert, Bois du S
-
too⏎You can have mutable values in the IO or (ST s) monads, if you need them.⏎> That is, I can't change a⏎> component of somethingNew without creating y. The point is there's⏎>
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:
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
coordinator Juan⏎Pedro Vela Arreola of the Alianza de Ribereños y⏎Armadores, an association of fishermen and producers in Mexico.⏎"Fishermen are desperate."⏎Gilly blames the most recent El Niño in the
-
May morning to view the medows gay,⏎I met a pretty dairy-maid all in the new mown hay.⏎* 4 *⏎As I was ridding out one day I saw some pooks of hay;⏎
-
yells,⏎A Yankee they had found,⏎With curses bid me “face about,”⏎And march away to town.⏎They marched me down with many more,⏎Through mud and mire to town,⏎And there with bars and prisons
-
household not naturalized."⏎- Homesteading Land The person had to have initiated the⏎naturalization process to be eligible for free land through homesteading.⏎- Voter Registration Lists Is
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
- L3:428 excites
- L3:468 excites
- L3:2 suppresses
- L3:65 excites
- L3:20 suppresses
- L3:1 suppresses
- L0:46 suppresses
- L3:673 suppresses
- L3:421 excites
- L3:298 suppresses
- L3:1598 suppresses
- L3:225 suppresses
- L3:1011 suppresses
- L2:154 weakly excites
- L3:42 weakly excites
- L3:740 weakly excites
- L3:1093 weakly suppresses
- L3:559 weakly excites
- L0:971 weakly excites
- L1:727 weakly suppresses
- L0:1967 weakly suppresses
- L3:141 weakly excites
- L3:1622 weakly suppresses
- L3:128 weakly excites
Feeds later layers
- L6:203 strongly excites
- L6:102 strongly suppresses
- L5:140 strongly excites
- L5:312 strongly excites
- L5:181 strongly suppresses
- L8:211 strongly excites
- L5:273 strongly suppresses
- L9:675 strongly excites
- L8:445 suppresses
- L8:239 suppresses
- L8:530 suppresses
- L5:268 suppresses
- L9:2053 suppresses
- L8:80 suppresses
- L5:148 suppresses
- L8:28 excites
- L9:606 suppresses
- L11:1018 excites
- L8:235 excites
- L9:2548 excites
- L11:35 excites
- L10:1710 excites
- L8:979 suppresses
- L8:310 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
- resco
- kel
- utely
- reflexivity
- emás
- ApplicationController
- createState
- theano
- SCRIPTION
- ******************************************************************************⏎
Suppresses
- 质量
- люд
- uD
- izen
- 触动
- after
- wj
- واجب
- 质量问题
- D