word pieces like 'vent' and 'venk'
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. 2 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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 certain chopped-up pieces of words, such as 'vent' in hyperventilation or 'venk' in a person's surname. Its negative side fires on 'Mount' after words like 'Sacred' and 'on'. Turning it up gave wandering, unfocused text.
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.
-
interaction⏎As magnetic interactions are normally mediated by short-range exchange or weak dipole fields, the research team, which is co-led by Professor T Venky Venkates
-
more than one anxiety disorder. Sometimes anxiety results from a medical condition that needs treatment.⏎Common symptoms of anxiety would be increased heart rate, hyperventilation, sweating, nervousness, weakness, fatigue, fidgeting, inability
-
atlas is a series of cross sections of anatomical structure. In most atlases, the three dimensions are as latero-lateral (x), dorso-ventral (y) and rostro-caudal (z). Another one is
-
00s. The current edition of this standard is NFPA 211-2019: Standard For Chimneys, Fireplaces, Vents, And Solid Fuel-Burning Appliances.⏎Fire gives off heat energy, which can
-
/11/Sanitizing-Surfaces-in-Childcare-Environments-1.pdfWtare Quaity nd Health Council⏎Publication Type: Care guideline<|endoftext|>The Cooper
-
of chaotic systems (H. Qammari, A. Venkatesan). A vacation on Mars - an artist's journey in a computer graphics
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: I am going to be in a new age, I have to go on. It's been a new age, I have to go on. It's been a new age, I have to go on. I am going to be in
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
and it has deeply influenced following developments of this phenomenon in the⏎rest of Europe.⏎Sacred Mounts of northern Italy represent the successful integration into a landscape of remarkable beauty of architecture
-
, twenty-seven inches wide, and twenty-seven inches deep. In it were to be placed the two tables of the law which Moses would receive directly from God on Mount Sinai (Ex
-
for the current demands.⏎4) Engine Stalls⏎When you start up your vehicle, and the engine turns on, you<|endoftext|>Darryl D’Monte continues his reportage of the Climate Action Network International meet on in Bangkok where,
-
ions (V Macedonica, XII Fulminata, XV Apollinaris) on the western side and a fourth (X Fretensis) on the Mount of Olives to the east. He put pressure on the food and water supplies
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
- L0:1224 suppresses
- L0:1301 suppresses
- L0:178 suppresses
- L1:30 weakly suppresses
- L0:1 weakly suppresses
- L1:2910 weakly excites
- L0:873 weakly suppresses
- L1:1464 weakly excites
- L1:340 weakly suppresses
- L1:1973 weakly suppresses
- L1:1110 weakly excites
- L1:491 weakly excites
- L1:1469 weakly suppresses
- L1:99 weakly excites
- L1:2101 weakly suppresses
- L1:32 weakly suppresses
- L1:951 weakly excites
- L0:246 weakly excites
- L1:92 weakly excites
- L1:2801 weakly excites
- L0:1347 weakly suppresses
- L0:2102 weakly suppresses
- L1:2764 weakly suppresses
- L0:2934 weakly excites
Feeds later layers
- L25:453 weakly excites
- L5:1203 weakly excites
- L4:1688 weakly excites
- L6:1591 weakly excites
- L6:1199 weakly excites
- L11:1418 weakly excites
- L10:2294 weakly excites
- L4:642 weakly suppresses
- L14:2280 weakly excites
- L14:571 weakly excites
- L16:779 weakly suppresses
- L13:2101 weakly excites
- L4:2109 weakly suppresses
- L21:1379 weakly excites
- L27:798 weakly excites
- L10:2118 weakly excites
- L7:686 weakly excites
- L4:1437 weakly excites
- L23:455 weakly excites
- L3:2322 weakly excites
- L10:1530 weakly suppresses
- L22:675 weakly suppresses
- L12:1316 weakly excites
- L19:2693 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
- @login
- inity
- etric
- entr
- usement
- 气体
- goo
- 满意
- 同学
- Launch
Suppresses
- -zA
- criptive
- wargs
- Hemp
- warts
- ansa
- iete
- 俳
- warz
- essor