the '0' in numbers like 30 or 40
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. 15 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 zero in two-digit numbers that start with a 3 or 4, such as 30, 40 or 4 0. It reacts to the second digit of a number. Turning it up made the model repeat phrases over and over.
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.
-
300K Trees in Nicosia”) NGO, Yesil Baris (Cy
-
30 major, 135 medium and around 3,000 small
-
30, Kahlo went with her husband to America. During this time, and for
-
30. It tells you the oil’s thickness or viscosity. A thinner motor oil —
-
40 years before the start of Prohibition—although her husband Rutherford B. Hayes
-
40 N⏎What is the period of a point moving along a circular path of radius
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: I have a 33333333333333333333333333333333333333333333
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
bike was stolen. I do not know, however, who did it.⏎To change from active to passive voice follow these rules: 1- The object becomes subject. 2- Use verb to be in the tense of the active verb.
-
όν lexicon, neut. of λεξικός lexikos, "of or for words", from λέξις lexis, "speech", "word",<|endoftext|>The Syrian Flag :⏎color in the Syrian
-
Urdu to English Translations. This will improve our English to Urdu Dictionary, Urdu to English dictionary, English to Urdu Idioms translation and Urdu to English Idioms translations. Although we have added
-
rates of relative sea level rise. © 2006 Elsevier Ltd. All rights reserved.<|endoftext|>Are you tired of being sick and tired of being sick and tired? Then, get in the driver’s seat on managing your health by
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
- L13:260 strongly excites
- L11:1353 strongly suppresses end-of-document token after a web snippet
- L13:2270 strongly suppresses
- L11:70 strongly excites
- L14:847 strongly excites
- L12:741 strongly excites
- L11:48 strongly suppresses
- L11:444 strongly excites
- L12:583 suppresses
- L5:299 excites
- L9:1574 suppresses
- L12:623 excites
- L14:98 excites
- L2:69 suppresses
- L11:34 excites
- L6:866 excites
- L12:1215 suppresses
- L11:928 suppresses
- L6:34 excites
- L8:139 suppresses
- L14:68 excites
- L14:2296 excites
- L6:583 excites unclear: mid-word pieces inside names
- L11:776 excites
Feeds later layers
- L16:192 strongly suppresses
- L16:368 strongly suppresses
- L16:654 suppresses
- L18:942 suppresses
- L22:213 suppresses
- L19:567 excites
- L16:841 excites
- L19:77 excites
- L18:199 suppresses
- L16:766 excites
- L16:1052 weakly excites
- L26:823 weakly excites
- L16:1164 weakly excites
- L16:1359 weakly excites
- L16:540 weakly suppresses
- L27:1120 weakly suppresses
- L16:462 weakly excites
- L19:839 weakly excites
- L16:638 weakly suppresses
- L16:776 weakly excites
- L16:1032 weakly excites
- L16:483 weakly suppresses
- L25:1134 weakly suppresses
- L19:1095 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
- exchange
- Exchange
- exchange
- cig
- atasets
- isas
- sheets
- Exchange
- iloc
- cub
Suppresses
- ,…⏎⏎
- …⏎
- …⏎⏎
- …⏎⏎
- …"
- weit
- …"⏎⏎
- .au
- 郯
- grou