hyphen joining word parts
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
The neuron fires on a hyphen that joins pieces of a long name or compound word, such as 'Ha-Devarim' or 'Qal'a-ye'. When turned up, the model mostly produced runs of blank lines, and turned down it repeated the word '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.
-
o Arama'ah: Tabulae in grammaticen linguae Chaldaeae (Paris, 1560); Aseret ha-Devarim: Decalogus, with the commentary of Abraham Ibn Ezra,
-
possible site for Alexander’s fort, is known today as the Citadel of Herat or Qal’a-ye Ikhtiyar al-Din. The name Ikhtiyar al
-
the one armed Tyr was the icon for a judge.⏎<|endoftext|>Sign up for our newsletter⏎Gardening in the Rainy Zone.⏎Pronounced: AL-lee-um ka-ra-tah-vee-EN-see⏎Sun
-
. In case you need a good example here is one from Edutopia.http://www.edutopia.org/blog/flip-stem-classroom-ainissa-ramirez and all flipped classrooms do not look alike.⏎To see how technology transforms
-
required prospective teachers to pass these tests. Research, however, does not show a relationship between test scores and teacher effectiveness.⏎In one [research study], Marc Claude-Charles Colitti of Michigan State University examined data going back to 196
-
of this article, click here: http://www.abc.net.au/news/rural/2017-05-23/the-future-of-automation-in-the-mining-industry/8550636<|endoftext|>
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:writingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwritingwriting
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
a lot of wood-frame construction, which is flexible and tends to ride out a quake fairly well, said Heidi Faison, outreach director at the Pacific Earthquake Engineering
-
tension, difficulty sleeping and difficulty concentrating.⏎Obsessive Compulsive Disorder (OCD)⏎Obs<|endoftext|>A mobile phone user outside a mobile payment facility in Port-au-Prince, Haiti.⏎our goal: to alleviate poverty
-
ard (1652-1728), who was assisted by Marshal Ferdinand de Marsin (1656-1706) and the Elect
-
About Apollo Missions"That’s one small step for a man, one giant leap for mankind...” nnOn 20 July 1969, Neil Armstrong made the historic first steps on the Moon. In celebration of the 5
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
- L26:2003 strongly excites hyphen joining two word parts
- L25:1399 strongly suppresses
- L20:61 excites
- L22:271 excites
- L25:36 excites
- L26:153 excites
- L26:381 excites
- L26:1374 suppresses
- L23:1351 excites
- L19:804 excites
- L26:254 excites
- L22:1533 excites
- L16:27 excites
- L26:725 excites
- L16:39 suppresses
- L0:141 excites the word 'of' in 'middle of'
- L26:1018 excites
- L25:40 excites
- L21:50 suppresses comma after 'So' or intro phrase
- L26:230 excites
- L21:2309 excites
- L26:319 excites
- L3:459 excites
- L20:1886 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
- ⏎⏎⏎
- ⏎ ⏎
- ⏎⏎⏎⏎
- �
- ,⏎⏎⏎
- "⏎⏎⏎
- ellaneous
- وء
- iction
- ויות
Suppresses
- writing
- hot
- dam
- bar
- per
- sm
- css
- cards
- cut
- fast