unclear: word-internal fragments in foreign names
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. 1 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 pin down. It sometimes fires on pieces of foreign or Latin-script words and names split in the middle. Turning it up gave writing that did not clearly relate to this.
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.
-
(core) city ; Greek ἄστυ asty ) with the associated surrounding area ( χώρα chō
-
ойте скобки, употребляя глаголы в Present, Past, Future Indefinite, Present
-
perialisation. And this can only be done by reducing Russia’s territorial capabilities, says Professor Andrzej NOWAK in an interview with Mikołaj CZYŻ.⏎Mikołaj CZYŻ: Are the roots of Western sympathy for
-
ing, p. 293).⏎Grīwin Old Iranian languages. Grīw goes back to the Old Iranian grī
-
Old Ind. grīva- “neck,” grīvā- “the back part of the
-
is and then to discover how to avoid using it. As this is probably one of the most common mistakes made by English speaking students, it’s “imprescindible” that you get to grips with
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 just thought of that. So, I need to get to some information from the internet to find out what is the problem here, and I need to get to some information about the problem. So, I need to get to some information
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the Catholic Church⏎The Council of Trent (Latin : Concilium Tridentinum), held between 1545 and 1563 in Trent
-
ἔλεγεν αὐτοῖς,⏎τὸ σάββατον διὰ τ
-
was the Libellus de abbreviaturis Hebraeorum, tam Talmudicorum quam Masoritarum et aliorum rabbinorum (Paris, 1561), later exploited by Guy *Le F
-
φος lexikographos, "lexicographer", from λεξικόν lexicon, neut. of λεξικός lexikos, "of or for words", from λέξι
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
- L0:1224 suppresses
- L0:1301 suppresses
- L0:178 suppresses
- L0:508 weakly suppresses
- L0:641 weakly excites
- L0:263 weakly suppresses
- L0:79 weakly suppresses
- L0:3068 weakly suppresses
- L0:1 weakly suppresses
- L0:419 weakly suppresses
- L0:784 weakly excites 'of' after 'think' or 'out'
- L0:1214 weakly suppresses
- L0:510 weakly suppresses
- L0:1634 weakly suppresses
- L0:2408 weakly suppresses
- L0:472 weakly excites
- L0:535 weakly excites
- L0:2267 weakly excites
- L0:187 weakly suppresses
- L0:656 weakly suppresses 'to' after 'as opposed' or 'related'
- L0:1656 weakly excites
- L0:2478 weakly suppresses
- L0:2888 weakly suppresses
- L0:2146 weakly excites
Feeds later layers
- L4:513 suppresses
- L23:1678 weakly excites
- L10:2398 weakly excites
- L2:187 weakly suppresses
- L21:233 weakly suppresses
- L16:1793 weakly excites
- L26:1199 weakly excites
- L15:1214 weakly suppresses
- L2:1510 weakly suppresses
- L10:2964 weakly suppresses
- L8:550 weakly excites
- L13:3009 weakly excites
- L12:1108 weakly excites
- L16:1621 weakly suppresses
- L15:2792 weakly suppresses
- L22:2162 weakly excites
- L9:784 weakly excites
- L18:145 weakly excites
- L15:1051 weakly excites
- L10:636 weakly suppresses unclear: 'your' in contact boilerplate
- L2:778 weakly excites
- L3:66 weakly excites
- L14:282 weakly excites
- L18:2501 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
- erah
- leri
- 他们都
- quette
- jang
- シャル
- PLIER
- תי
- adalafil
- celona
Suppresses
- �
- VO
- vo
- ims
- �
- rows
- olith
- YT
- OF
- rots