hyphen inside a compound word
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. 27 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 lights up on the hyphen in words joined by a dash, like Self-Reflection or Coal-to-liquid. Turning it down made the model produce run-on strings of words without spaces, and turning it up made it repeat small words like 'the'.
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.
-
roversial Qeiyafa Comes to U.S. Museum<|endoftext|>An Argumentative Essay about Mary Oliver’s Poems: Exploring the Power of Nature and Self-Reflection
-
were all based on Earth and planetary cycles, most had a spring goddess. The spring goddess in Scandinavian was "Ostra," for Anglo-Saxons it was "Eostre," and the current region of Germany she was called "Eastre
-
be even easier to use. Do you live in or near a Cool City? Check the site now to find out.⏎The Worst Alternative Fuel Ever?⏎Coal-to-liquid fuel has
-
the documents are surrounded by an inert gas. A cross-section of these cases can be seen in the “Public Vaults,” along with a special documentary about the re-encasement process.⏎This brochure is made possible by the National Archives Foundation with
-
05, 2012, from http://www.studymode.com/essays/The-Struggle-For-Power-In-The-994323.html⏎"The Struggle for Power in the U
-
/essays/The-Struggle-For-Power-In-The-994323.html>.⏎"The Struggle for Power in the U
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:EXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENTEXPECTMENT
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
in the Trans-Pecos. Current evidence indicates that mountain lions (Felis concolor), bobcats (Lynx rufus), and coyotes (Canis latrans) are the primary predators involved, although little information is available
-
:⏎Fifty-eight mammal species have been reported in the park to date such as:⏎- Tiger (Panthera Tigris),⏎- Clouded Leopard (Neofelis nebulosa),⏎- Swamp Deer (Cervus du
-
how many exist.⏎These tentacled titans have now ensnared marine biologists with⏎another riddle: they have left their normal gathering grounds in⏎the Sea of Cortez in Baja California, Mexico. Fishermen worry that
-
when using the bathroom, particularly shigella (a foodborne bacteria that causes diarrhea, fever, and stomach cramps) and norovirus (often called "stomach flu" which also causes nausea, vomiting, stomach cramps,
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
- L19:1174 strongly excites
- L19:124 strongly suppresses
- L22:65 strongly suppresses
- L16:1421 suppresses
- L20:563 excites
- L14:140 suppresses
- L18:76 suppresses blank spot before a number
- L20:59 suppresses
- L21:29 excites
- L22:14 excites
- L18:48 suppresses
- L19:963 excites
- L25:17 excites
- L10:84 suppresses
- L25:239 excites
- L23:262 suppresses
- L12:2582 suppresses
- L23:80 excites blank slot before a year or number
- L25:222 suppresses
- L25:109 excites
- L19:1005 excites
- L14:348 suppresses
- L20:193 excites
- L16:723 suppresses
Feeds later layers
None: this is the last layer.
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
- is
- that
- .
- @
- :
- =
- i
- of
- and
- #
Suppresses
- artz
- ennent
- →se
- →en
- atem
- apor
- ourt
- TIM
- BEGIN
- ores