mid-sentence 'also' and 'another'
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. 12 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 small linking words like 'also', 'further' and 'another' when they sit between a subject and the main point of a sentence. Turning it up made the model's text drift into loose, rambling sentences that did not clearly follow the prompt.
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.
-
urban schools cause obesity in minority children?⏎The May 29 Baylor University study further indicates that those same overweight black and Hispanic kids are also less likely to benefit from
-
Step⏎Finally, you have followed all the steps, you will see one shortcut on the desktop.
-
LIEST STEP IN HUMAN DEVELOPMENT⏎--Finding Reinforces Limitations of Using Other Cell Types in Research, Scientists Say⏎The finding, they say, also highlights one aspect of human cell biology that has not been replicated in other animal model
-
microLEDs also featured a reduced dislocation density (by 59.6%). According to
-
an acid taste, if desired, but the acid<|endoftext|>5 Tips For Having Fun With Shadows Page 2⏎Silhouettes offer another opportunity to use shadows. A silhouette is a shadow--an unlit subject.
-
The Treaty Oak has also played an important and romantic part in the history of the Lone Star State. One
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: the place of art, specifically the media, in the various forms of culture, including the visual arts, the performance art, the information, the media, the information, the artifacts, the information, the forms, the stories, the information
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the biology of the disorder and find targets for better treatments. It’s latest in a steady stream of good news from autism researchers recently, who have
-
compact, dense collection of supersymmetric matter drifting around the universe. According to Cox’s commentary on the film, “if [a Q-ball] sat
-
ate of Mecca and the Zaidi Imamate of Sanaa. It also recounts the Turkish and British attempts to dominate the region, especially the sea route
-
promoting the rule of law, ensuring access to justice, and strengthening institutions. It is the intersection of SDG 11 and SDG 16 that
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
- L10:369 strongly suppresses
- L9:1033 suppresses unclear: end-of-text markers and word fragments
- L11:1391 suppresses
- L7:1053 suppresses
- L8:269 excites
- L10:710 excites
- L8:624 excites
- L7:1130 excites
- L11:607 weakly excites
- L2:182 weakly suppresses
- L6:927 weakly suppresses
- L8:49 weakly excites
- L11:908 weakly excites
- L3:912 weakly excites
- L6:2753 weakly suppresses
- L7:61 weakly suppresses
- L11:2235 weakly excites
- L7:1022 weakly excites
- L3:2372 weakly suppresses
- L10:1048 weakly suppresses
- L4:444 weakly suppresses
- L3:177 weakly suppresses
- L5:576 weakly suppresses
- L2:1671 weakly suppresses
Feeds later layers
- L15:860 suppresses
- L15:2008 suppresses
- L13:2673 suppresses
- L14:1563 excites
- L13:2982 suppresses
- L14:1816 excites
- L13:1809 excites
- L14:2741 excites
- L14:1504 excites
- L17:2032 excites
- L16:2575 suppresses
- L14:2122 suppresses
- L13:1644 excites
- L14:946 weakly excites
- L14:2587 weakly suppresses
- L18:1342 weakly excites
- L15:3039 weakly excites
- L13:418 weakly suppresses
- L14:2324 weakly excites
- L16:1576 weakly suppresses
- L13:2216 weakly suppresses
- L23:2389 weakly suppresses
- L15:2966 weakly excites
- L17:1896 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
- room
- 他又
- icer
- adients
- 他还
- Footer
- continued
- uga
- ếu
- road
Suppresses
- ?,⏎
- —⏎⏎
- 日讯
- =batch
- 骥
- 亿吨
- --⏎
- '-',
- ?',⏎
- -⏎