unclear: word pieces in names and foreign text
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. 3 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 fires on fragments in the middle of names and words, often in Dutch, German, Latvian or Greek text. There is no clear pattern beyond that, and turning it up or down did not change the writing in an obvious way.
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.
-
Houdini’s career led him to collaborate with Scientific American on a lengthy exposé of
-
ukken berustende bij de Vereenigde Doopsgezinde Gemeente to Amsterdam, 2 vols. Amsterdam: Uitgegeven en ten geschenke aangeboden door den Kerkeraad d
-
Um dies (i.e. Varro's description of the extent of Roma quadrata) mit dem unten erörterten Pomerium und der Auffassung des Dionys. I 88: περιγρ
-
products. In the spring of 1934, the Franks r…⏎E-pasta adrese, uz kuru nosūtīt darba saiti:⏎Saite uz darbu:
-
ἔλεγεν αὐτοῖς,⏎τὸ σάββατο
-
This is the most obvious similarity between the two religions. In fact a Muslim is defined as anyone who says "Th…⏎E-pasta adrese, uz kuru nosūtī
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 been thinking about this. I have been thinking about this. I have been thinking about this. I have been thinking about this. I have been thinking about this. I have been thinking about this. I have been thinking
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
bill to establish a federal holiday in 1894. President Grover Cleveland signed the bill soon afterward, designating the first Monday in September as Labor Day.⏎Learners are Leaders, Inc.’s mission includes uplifting future leaders through
-
trade unions was given further impetus in the freedom struggle against British colonialism. One of the earliest trade unions, Tata Workers Union (also alternatively known as Jamshedpur’s Workers’ Union) (http://www.tataworkersunion.org/)
-
⏎Principal researcher: Reyhaneh Bastani⏎Supervisor: Dr. Beaumie Kim⏎P
-
of 1857. He is known for his efforts to modernize his country and for restoring Mexico to a Republic.⏎May 1st: Labor Day, commemorates the union movement of the Mexican worker. Article 123
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
- L0:1301 suppresses
- L0:178 suppresses
- L0:1224 suppresses
- L0:386 weakly suppresses
- L1:651 weakly excites
- L2:1334 weakly suppresses
- L1:1099 weakly excites
- L1:2474 weakly excites
- L2:36 weakly excites
- L1:2013 weakly excites
- L0:894 weakly excites
- L1:515 weakly excites 'about' in 'it's not about' phrasing
- L1:1795 weakly excites
- L1:92 weakly excites
- L0:874 weakly excites
- L0:2779 weakly suppresses
- L2:1899 weakly excites
- L1:1665 weakly suppresses
- L1:1496 weakly excites
- L1:2620 weakly excites
- L1:491 weakly excites
- L2:645 weakly suppresses
- L1:2166 weakly suppresses
- L2:1882 weakly suppresses
Feeds later layers
- L4:528 weakly excites
- L6:322 weakly excites
- L12:538 weakly excites
- L6:31 weakly excites
- L5:146 weakly suppresses
- L24:190 weakly excites
- L27:351 weakly suppresses
- L16:2481 weakly excites
- L25:587 weakly suppresses
- L7:2560 weakly suppresses
- L6:574 weakly suppresses
- L15:2423 weakly excites
- L6:1667 weakly suppresses
- L24:1218 weakly suppresses
- L7:3 weakly excites
- L19:493 weakly excites
- L8:997 weakly suppresses
- L8:2236 weakly suppresses
- L10:214 weakly suppresses
- L5:1799 weakly excites
- L14:1909 weakly suppresses
- L14:2801 weakly suppresses
- L10:1565 weakly excites
- L23:788 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
- 他们会
- 若您
- ento
- UPER
- getResources
- 猖
- onward
- его
- fasc
- Fasc
Suppresses
- ocha
- Necklace
- ấm
- ellig
- proble
- 晕
- unya
- 糖
- 饽
- blem