'in' or 'the' in 'up in the air'
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. 6 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 responds to the small word after a phrase about moving or rising, like 'went up in the air' or 'rose higher in the sky'. It reacts negatively to emotional descriptions such as 'in awe', 'stern' and 'confusion'. Turning it up or down gave no clear, matching theme in the written text.
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.
-
n of Farewell—literally, The back turned on Ireland.⏎Writers without number have traced the glories of Iona. Here rose, as if by miracle, a city of churches; the isle became one vast monastery, and
-
and faced them toward each other one will over power and cause the other to spin in the wrong direction (burn out just like connecting a 9V and 6V
-
ars in fright, trampling men beneath its hooves. At lower left, another warrior turns his head away from the fire, flinging one hand up in the air while
-
out in dark spots. Over the days and weeks that followed, these spots formed fan-like markings, and spidery patterns. As the sun rose higher in the Martian sky, the spots and fans became more numerous.⏎"Originally, scientists thought
-
treated?⏎Children with significant ptosis may need to tilt their head back into a chin-up position, lift their eyelid with a finger, or raise their eyebrows in an effort to see from under their drooping eyelid. Children with congenital
-
men heard the shells edging closer. “And then bang!—one of the big shells struck the ground right in front of us and we all went up in the air. But we all come down again.”⏎York was showered with honors
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 don't know why I look at me at the same time as I see what I'm doing. I'm trying to do something. I'm looking at the others around the world. It's like, trying to see what the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the CSUF faculty a year ago.⏎Matheus had been interested in the physical sciences since he was a kid. He would stare up at the clouds in awe and wonder what processes were involved. In high school, he competed in national and
-
are individualized, his expression stern. On the reverse, his parents, Heliocles and Laodice
-
inar: What is Gene Drive?⏎- Over 2,000 Nigerian<|endoftext|>Books for Children⏎A Japanese American boy named “Shorty” watches with confusion and awe as he and his family are removed from their home and sent to an
-
expert reaction ranged from the skeptical to the dismissive, with good reason. The plans called for Google to put hundreds of solar-powered
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
- L3:950 excites
- L1:445 excites document start marker token
- L3:1095 weakly excites
- L2:914 weakly suppresses
- L4:871 weakly suppresses
- L5:1234 weakly suppresses
- L1:1058 weakly excites closing quote marks after a quoted term
- L4:454 weakly suppresses
- L3:2758 weakly excites
- L3:103 weakly excites
- L0:882 weakly suppresses
- L1:2577 weakly excites
- L4:507 weakly suppresses
- L4:2165 weakly excites
- L5:1007 weakly suppresses
- L1:2600 weakly excites
- L3:3027 weakly excites
- L3:115 weakly suppresses
- L1:1206 weakly excites
- L0:288 weakly excites
- L3:1179 weakly suppresses
- L0:2213 weakly suppresses
- L5:1377 weakly suppresses
- L2:679 weakly suppresses
Feeds later layers
- L8:725 suppresses
- L13:2457 suppresses
- L9:3022 suppresses
- L10:2771 suppresses
- L10:2227 suppresses
- L8:2991 weakly suppresses
- L14:175 weakly excites
- L18:2457 weakly suppresses word after "due to" or "with"
- L8:825 weakly suppresses
- L10:3029 weakly suppresses
- L11:2334 weakly suppresses
- L14:3028 weakly excites
- L7:425 weakly suppresses
- L9:1809 weakly suppresses
- L10:2939 weakly suppresses
- L8:1793 weakly suppresses
- L12:431 weakly suppresses
- L12:1223 weakly suppresses
- L8:1176 weakly suppresses
- L10:523 weakly excites
- L25:1590 weakly excites
- L9:663 weakly suppresses
- L10:2045 weakly excites
- L26:2998 weakly excites
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
- igne
- ák
- thé
- PID
- 内科
- gaard
- fine
- 附近的
- �除
- theValue
Suppresses
- 听完
- 听了
- 她说
- 的眼神
- 率为
- 谅解
- ери
- aptic
- 批
- 看著