line-break tokens inside wrapped 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. 8 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 mostly fires on the hard line break in text that was copied with fixed-width line wrapping, where a sentence is split across lines in the middle. The steering results show no clear link to this, so the effect is weak and hard to read.
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.
-
The Treaty Oak has also played an important and romantic part in the history of the Lone Star State. One
-
As India dissented over⏎the US actions in Vietnam, the US offered technology as well. This⏎technology added another element that made the situation in India much⏎different and more difficult to assess than that under the British⏎Mercantilism
-
nitrogen-fixing bacteria⏎and the roots of the plants in which they live, and are examining⏎how synthetic chemicals might interfere with these signals. Disrupting⏎these 'signals of life' could have important ecosystem impacts.⏎developing fetus is
-
that: ethnically, they are the descendants of the ancient inhabitants⏎of Mesopotamia; culturally, they are the heirs of their Aramaic language⏎To be accurate from the start, I must add this clarification:⏎1) the
-
seven at the most. Atop that, the immigrants usually collected into⏎a few identifiable regions.⏎Consider the new states largest single immigrant group. Oklahomas⏎first decennial federal census, taken in 1910, calculated that
-
and⏎the occasion will be celebrated as a sign of the boy's passage into maturity⏎and promising future. A significant event in the culture and history of South⏎India is known as the Dhoti Ceremony. This is a momentous day
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: what is the value of the expression 1000000000000000000000000000000000000000
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
It’s amazing to keep in touch with everybody and see some amazing work being produced.⏎We hope
-
culture. As a whole, these examples illustrate the three kinds of discourse at which rhetoric is aimed.⏎Oxford Scholarship Online requires a subscription or purchase to access the full text of books within the service. Public users can however freely search the site and
-
the many countries whose citizens have used SOHO to chase comets.⏎Almost all of SOHO's comets are discovered using images from its Large Angle and Spectrometric Coronagraph (LASCO) instrument. LASCO is used to observe the
-
when elements cannot be saved, field biologists help bring back species and their ecosystems (restoration biology).⏎Chicago Scholarship Online requires a subscription or purchase to access the full text of
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
- L2:3065 weakly suppresses
- L5:398 weakly excites
- L1:1857 weakly suppresses
- L7:1130 weakly suppresses
- L0:2871 weakly suppresses
- L1:1973 weakly suppresses
- L4:2969 weakly suppresses
- L2:2242 weakly suppresses
- L4:1307 weakly suppresses
- L2:111 weakly suppresses
- L3:2529 weakly excites
- L3:1392 weakly suppresses
- L4:245 weakly excites
- L7:110 weakly suppresses
- L6:2644 weakly excites
- L7:62 weakly excites
- L1:2278 weakly excites
- L2:1997 weakly suppresses
- L3:1948 weakly excites
- L4:1220 weakly suppresses
- L3:2820 weakly suppresses
- L6:2972 weakly suppresses
- L5:1978 weakly suppresses
- L7:92 weakly suppresses
Feeds later layers
- L9:1377 excites
- L9:218 excites
- L9:475 excites
- L10:923 excites
- L9:169 suppresses
- L10:710 suppresses
- L12:978 suppresses
- L19:1303 weakly excites
- L11:684 weakly excites
- L11:1686 weakly excites
- L9:1459 weakly excites
- L11:1719 weakly excites dash starting a bullet item
- L10:196 weakly excites
- L11:1233 weakly suppresses
- L9:134 weakly excites
- L11:1391 weakly excites
- L10:27 weakly excites
- L9:125 weakly excites
- L13:1304 weakly suppresses
- L24:1088 weakly suppresses
- L12:702 weakly excites end of text token after prose
- L27:2219 weakly suppresses
- L26:2775 weakly suppresses
- L21:660 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
- aro
- createQuery
- requently
- nossa
- ullan
- Morm
- itsu
- []>(
- Harding
- 我们可以
Suppresses
- loaded
- Skip
- 研讨会
- arden
- ��
- ícia
- phans
- cribe
- bre
- PointSize