unclear: pieces inside numbers and blanks
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. 9 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 is weak and hard to read. It fires on odd spots such as a single digit inside a number, a blank line, or a comma in a big number. Nothing about it is clearly about one topic, and turning it up did not bring out a clear theme.
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.
-
128.0/17 SUBNET3 : 10101100.00010001.0yyyyyyy.yyyyyyyy => 172.17.0.0
-
were 1, 2 and 134. All these vehicles were fitted with ballast ploughs.⏎N 1
-
1, 1/4 turn clockwise and then back 2_______.⏎5.1/4 turn clockwise, forward 3, 3
-
and “motto” theme, but does not stay there long. The tumultuous Allegro begins in D minor, a key as far as possible from A-flat (Elgar wrote to his friend August Jaeger that he had “thrown over
-
48x2048 pixels,⏎be sure to set camera to this or lower⏎- Handheld GPS,⏎preferably one with 100’ or better accuracy that allows⏎marking locations and naming them⏎Accuracy is important
-
||Water (%)||1.57% (1,014 km2)|⏎|-||2011 estimate||2,22
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: The opposite of the same In the opposite of the same In the opposite of the same In the opposite of the same In the opposite of the same In the opposite of the same In the opposite of the same
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
plumage with a black back.⏎- ‘Neophytes at duck identification often confuse a bufflehead with a male hooded merganser when the latter's crest is raised showing a large patch of white feathers.’⏎- ‘These
-
a familiar spirit.⏎How do we reconcile these apparently conflicting statements? We do so by recognizing that the latter statement is the figure of speech “metonymy of the subject.” The actual
-
(1) military operations, (2) weapons technology, (3) diplomatic activities, (4) intelligence activities, and (5) cryptology. The latter two areas might be considered to be special parts of the first three areas. That
-
, Guatemala, Florida, and France between 1789 and 1809.2 Much remains to be known about the St. [End Page 37] Dominguan diaspora in France and in the circum-Car
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
- L6:267 strongly excites
- L4:77 suppresses
- L4:325 excites
- L5:51 excites
- L8:1671 suppresses
- L2:51 suppresses
- L6:763 excites
- L7:50 suppresses
- L5:851 excites
- L6:355 suppresses
- L4:436 suppresses
- L5:200 suppresses
- L1:46 excites
- L8:47 suppresses
- L6:955 excites
- L6:180 suppresses
- L8:57 excites
- L5:1510 suppresses
- L1:107 excites
- L7:475 excites
- L7:1866 excites
- L3:16 excites
- L1:1371 suppresses
- L0:811 suppresses opening parenthesis after a bracket or name
Feeds later layers
- L12:1209 excites
- L15:483 excites
- L12:198 excites
- L11:530 suppresses
- L10:2162 suppresses
- L12:1483 excites
- L13:932 suppresses second digit of a year range
- L10:1024 suppresses
- L13:452 excites
- L13:1020 excites
- L10:1396 suppresses
- L12:532 suppresses
- L14:477 excites
- L12:2301 suppresses
- L12:2783 suppresses
- L12:2042 suppresses
- L11:1705 suppresses
- L10:1062 excites
- L10:506 suppresses
- L13:1783 suppresses
- L10:171 suppresses
- L13:1497 excites
- L14:566 suppresses
- L14:845 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
- eled
- eler
- 岁月
- 三项
- очных
- очный
- urable
- вел
- qué
- /animate
Suppresses
- 更深
- 他还
- đáo
- adipiscing
- 他又
- оборот
- brighter
- höher
- Tat
- arat