"especially" after a comma
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 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 word "especially" when it comes after a comma and adds a special case to a sentence, like "useful, especially to ward off". It is pushed strongly negative on certain punctuation, such as a comma or dash after a name or quote. Turning it up or down did not give a clear shift in the writing.
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.
-
in some cases, glass bottles offer the better choice, especially for certain types of beverages and storage circumstances. The difference lies with the permeability
-
on purees and become picky eaters, especially if they get used to extremely smooth purees. However, you can avoid this
-
, UGM, Saturday (31/10).⏎According to Hamidah, eating natural fruits is very useful, especially to ward off various diseases that come with age. "If we go back to
-
still appropriate that certain times be set aside for a particular focus on repentance.⏎In much the same way, we celebrate the resurrection of Christ each Sunday, but especially at Easter; and while we should always thank God for the Incarnation, we
-
, this can make telling their true age difficult, especially when dealing with strays that somehow warmed their way into our hearts and home.
-
subject.⏎• Keep your tiles clean. A clean tile won't have a film of accumulated grinding dirt on the surface.⏎• Avoid the use of unnecessary chemicals, especially acid based ones and the use of hard cleaning implements such as steel wool. These
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 of the subject of the person who is the person who is the person. It is the person. It is the person. It is the person. It is the person. It is the person. It is the person
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
BALANCAN, MEXICO (EFE).- The finding of 23 archaeological pieces from about 600 BC has
-
'S', then click at the end of the word and type<|endoftext|>Mallorca’s popular tourist railway, the ‘Orange Express,’ has entered its 90th year of electrification. The Palma –
-
achay-Cherkessia, Russia, near the border of Georgia.⏎Elbrus’ earliest name was Strobilus, Latin for ‘pine cone,’ from the ancient Greek strobilos, meaning ‘a twisted object’, a botanical
-
of the difference between 'wearing' and 'scratching.' Wearing is where the glazed surface is thinning and eventually actually work through.
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
- L0:178 suppresses
- L0:1224 suppresses
- L0:215 weakly suppresses
- L0:348 weakly excites space token right after a period
- L0:1301 weakly suppresses
- L0:456 weakly excites
- L0:249 weakly excites ordinal suffix "nd" after numbers
- L0:319 weakly excites blank space after a label colon
- L0:1518 weakly excites
- L0:1794 weakly suppresses
- L0:2838 weakly suppresses
- L0:276 weakly suppresses
- L0:899 weakly suppresses
- L0:847 weakly excites
- L0:2659 weakly excites
- L0:622 weakly excites
- L0:1276 weakly suppresses
- L0:36 weakly excites
- L0:2692 weakly suppresses
- L0:887 weakly suppresses
- L0:1483 weakly excites
- L0:414 weakly suppresses blank after a dash in number ranges
- L0:1406 weakly excites
- L0:585 weakly excites
Feeds later layers
- L2:1142 weakly suppresses
- L26:2143 weakly suppresses
- L25:342 weakly suppresses
- L2:986 weakly suppresses
- L2:1837 weakly suppresses
- L10:2191 weakly excites
- L4:2018 weakly excites
- L5:1299 weakly excites
- L20:603 weakly suppresses
- L4:2260 weakly suppresses
- L26:1443 weakly suppresses
- L27:492 weakly suppresses
- L17:277 weakly excites
- L2:1974 weakly excites
- L20:1422 weakly suppresses
- L17:2979 weakly suppresses
- L9:334 weakly suppresses
- L13:1796 weakly excites
- L17:1151 weakly excites
- L18:655 weakly suppresses
- L26:2115 weakly suppresses
- L21:2280 weakly excites
- L16:2633 weakly suppresses
- L11:2156 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
- yen
- self
- yre
- eyond
- glomer
- crease
- ramid
- yü
- iere
- uridad
Suppresses
- 康复
- Чт
- 万美元
- устройств
- 捐献
- 选秀
- (es
- одерж
- 습
- ám