line break 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. 10 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 a line break that splits a sentence in the middle, where the text wraps onto a new line mid-phrase. Turning it up did not produce a clear, consistent theme in the output.
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.
-
look at HTTP itself.⏎The HTTP Protocol Show Contents Go to Top Previous Page Next Page⏎The HTTP protocol was designed to be so simple that anyone with basic programming⏎skills could write an HTTP client or server. In fact, hundreds of people have
-
in⏎fructooligosaccharides. Levrat et al. (1991) demonstrated that even low⏎levels of inulin are effective in decreasing plasma cholesterol in rats.⏎DESIGN:
-
Diets lower blood pressure for Children . 12 An impressive number of studies, dating back to the early 1920s, show that veg-<|endoftext|>If you take a walk through the Salt Lake City neighborhoods and hills around our
-
in chemically reacting powder mixtures⏎- pressure processes which are descriptive of "ballotechnic" reactions, i.e. shock- induced reactions in heterogeneous ma- terial mixtures.⏎- ISOTOPIC SEPARATION AND ENR
-
⏎hole at flat end of the broomstick.⏎10. Secure the net to the pole using the U-shaped wire and duct tape (as⏎Volunteer Water Quality Monitoring⏎Non-profit Stream Monitoring⏎County Stream Monitoring Groups⏎
-
houettes are once more gaining popularity as an art form. A few⏎professionals can create realistic silhouettes using nothing other than paper⏎and scissors
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: In the long run, in the long term, in the overall trend, in the overall context, in the overall context, in the overall context, in the overall context, in the overall context, in the overall context, in the overall
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the picture?⏎10 space travellers are waiting to board their spaceships. There⏎are two rows of seats in the waiting room. Using the rules, where⏎are they all sitting? Can you find all the possible ways?⏎What is the
-
⏎celebration is revealed. This chapter will therefore treat of the celebration⏎of the sacraments of the Church. It will consider that which, through the⏎diversity of liturgical traditions, is common to the celebration of the seven
-
, New Mexico, Montana, Alaska, and California; Canada=⏎Manitoba, Alberta, British Columbia; Mexico; Brazil; Poland; Denmark=Bornholm,⏎Greenland; Austria; Switzerland; Hungary;
-
is poured out by two young girls, one of whom is called the male butterfly,⏎and the other the female butterfly, - appellations derived from the susu,⏎or sake-jugs, each of which is adorned
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
- L4:337 excites
- L7:2925 suppresses
- L9:234 suppresses line break inside hard-wrapped text
- L7:133 suppresses line break inside a sentence
- L6:574 suppresses
- L6:1255 excites
- L2:1580 suppresses
- L6:1374 suppresses
- L4:424 suppresses
- L7:239 excites
- L8:380 suppresses
- L2:772 suppresses
- L6:882 excites
- L6:1337 excites
- L6:463 excites
- L4:105 excites
- L3:428 excites
- L3:42 excites
- L9:246 suppresses
- L9:1659 suppresses
- L3:1592 suppresses
- L7:682 suppresses
- L7:74 weakly suppresses unclear: scattered mid-sentence tokens
- L8:1656 weakly suppresses
Feeds later layers
- L17:1242 strongly excites
- L15:643 strongly suppresses
- L12:1012 excites
- L21:579 suppresses
- L16:795 excites
- L17:910 excites
- L15:985 excites line breaks inside wrapped text
- L19:970 excites
- L19:771 suppresses
- L17:1564 excites
- L16:2348 excites line break after a section heading
- L16:442 excites
- L17:2096 excites
- L14:2654 excites
- L12:1848 weakly excites
- L17:554 weakly excites
- L18:1706 weakly suppresses
- L26:1309 weakly suppresses
- L25:325 weakly excites
- L13:1439 weakly suppresses end of document token
- L19:901 weakly suppresses
- L17:1273 weakly suppresses
- L16:2723 weakly suppresses
- L22:165 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
- ircular
- Pact
- ##
- ###
- 旦
- 了一个
- .handleChange
- accidentally
- line
- 俨
Suppresses
- —⏎⏎
- { ⏎ ⏎
- }</
- etur
- 补齐
- ()}</
- { ⏎ ⏎
- ]</
- xdf
- ']}</