the 'Ar' syllable inside names
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. 24 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 letters "Ar" when they appear in the middle of a name, such as Arista, Arroyo, or Arches. Its negative side fires on single digits inside numbers, such as the 4 in 1545. When it was turned up, the model's text was filled with quote marks.
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.
-
Life examines the work of extraordinary raconteurs Salvador Novo, Cube Bonifant, Roberto Arlt, Alfonsina Storni, and Mário de Andrade,
-
-⏎2. Mariano Arista, El general en gefe del cuerpo de Ejército del Norte á
-
California's Guaymas Basin in the⏎"We have a huge problem on our hands," says coordinator Juan⏎Pedro Vela Arreola of the Alianza de Ribereños y⏎Armadores, an
-
decades, the church in Brazil has shown itself strongly on the side of the poor. Mentioning great figures like Bishop Helder Câmara and Cardinal Paolo Evaristo Arns, Bernard spoke of the prophetic character of the Brazilian church. It has
-
.⏎Puente de Colgante⏎The construction of the Puente Colgante was completed in 1852. The bridge linked the district of Arroceros (the area surrounding the Metropolitan Theater and Manila City Hall) to
-
visitors than the Grand Canyon or Yellowstone National Park. By comparison, Utah's five National Parks—Zion, Bryce Canyon, Capitol Reef, Canyonlands, and Arches—had a combined total of 5.3 million visitors in 2
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 if a single right " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " "
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the Catholic Church⏎The Council of Trent (Latin : Concilium Tridentinum), held between 1545 and 1563 in Trent (or Trento), now in
-
III, who was the King of England from 1327 to 1377. Edward, who loved knightly pursuits such as war, joust
-
015 The University of North Carolina Press⏎116 South Boundary Street, Chapel Hill, NC 27514-3808⏎How to Order | Make a Gift
-
we are printing the element for index i and decrementing the value of i by 1.⏎If you run this program, it will print the below output:⏎5 4 3 2 1⏎We need to use the range function
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
- L22:336 excites
- L17:1665 suppresses
- L21:749 suppresses
- L23:1360 suppresses
- L16:1922 weakly excites
- L13:2336 weakly excites
- L13:2850 weakly excites
- L21:162 weakly excites
- L19:796 weakly suppresses
- L22:1656 weakly excites
- L19:2194 weakly suppresses first half of hyphenated compound adjectives
- L19:3035 weakly excites
- L18:75 weakly excites
- L18:506 weakly suppresses
- L18:2255 weakly suppresses
- L20:2674 weakly excites
- L19:450 weakly suppresses
- L20:1587 weakly excites
- L3:2870 weakly suppresses
- L21:636 weakly excites
- L11:1557 weakly suppresses
- L0:1673 weakly suppresses
- L21:2761 weakly excites
- L19:2805 weakly excites
Feeds later layers
- L25:1641 strongly suppresses
- L27:661 suppresses
- L25:161 suppresses
- L26:302 suppresses colon or comma before a quote
- L27:263 weakly suppresses
- L26:897 weakly suppresses the 'gro' fragment in words
- L27:33 weakly excites
- L27:315 weakly suppresses
- L27:2753 weakly excites
- L26:1023 weakly suppresses
- L26:363 weakly suppresses
- L27:825 weakly suppresses
- L26:21 weakly suppresses
- L27:1521 weakly excites
- L27:1045 weakly suppresses
- L27:112 weakly excites
- L25:1667 weakly suppresses
- L25:109 weakly suppresses
- L27:292 weakly suppresses
- L25:232 weakly suppresses
- L26:605 weakly excites
- L26:1314 weakly suppresses
- L26:439 weakly excites
- L25:2443 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
- }}],⏎
- 选
- roach
- IRO
- 不清楚
- 蒟
- ucose
- cente
- 薅
- ','".$
Suppresses
- "
- '
- “
- ‘
- «
- ’
- ”
- "⏎
- "]
- ing