end-of-document separator token
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. 18 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 lights up on the invisible marker the model uses to show where one web document ends and another begins. When it was turned up, the written text drifted into unrelated rambling, with no clear theme.
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.
-
0240.<|endoftext|>By 2003, plenty of data suggested that pesticide drift was harming
-
, 2012<|endoftext|>Some weather conditions are simply not fun to ride in. As cyclists, windy days
-
old<|endoftext|>Immune cells in the brain have surprising influence on sexual behavior⏎Relatively uncommon
-
etto is already<|endoftext|>The North Carolina Gold Rush took place decades prior to the 1849
-
:<|endoftext|>Copyright © University of Cambridge. All rights reserved.⏎'Aim High' printed from
-
:<|endoftext|>The hamstring muscles – the biceps femoris, semitendinosus,
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: about the time of the day of the way to the back of the square from to the main square, and to go there there to do to there to to there to there there to there to there to there to there to there there to
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
asiveness of viral campaigns in social network sites: Persuasive intent as the underlying mechanism. Journal of Marketing, 18(1), 39-53. http://dx.doi.org/10.108
-
rin Froese (2008). Organic Virtue: Reading Mencius with Rousseau. Asian Philosophy 18 (1):83 – 104.⏎Tao
-
environments.⏎Source: Health Behavior and Policy Review, Volume 8, Number 4, July 2021, pp. 281-293(13)⏎Publisher: Paris Scholar Publishing Ltd.<|endoftext|>March Is
-
⏎James Behuniak Jr (2011). Naturalizing Mencius. Philosophy East and West 61 (3):492-515.⏎David E. Soles (1999). The Nature
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
- L9:81 suppresses
- L8:1365 suppresses
- L17:592 excites
- L4:1371 excites end-of-text marker at document breaks
- L2:2188 weakly excites
- L1:1464 weakly excites
- L13:1614 weakly excites 'the' after a preposition or verb
- L17:357 weakly suppresses
- L7:1106 weakly excites
- L5:110 weakly suppresses
- L0:1162 weakly suppresses
- L14:1709 weakly suppresses
- L3:2015 weakly excites
- L15:867 weakly suppresses
- L3:469 weakly excites
- L7:76 weakly excites
- L1:2966 weakly excites
- L1:1606 weakly excites
- L11:1379 weakly excites
- L3:1121 weakly suppresses
- L3:103 weakly excites
- L3:962 weakly suppresses
- L17:738 weakly excites
- L10:1578 weakly excites
Feeds later layers
- L19:2973 suppresses
- L21:1710 suppresses
- L22:2929 weakly excites
- L19:1272 weakly excites
- L20:305 weakly excites
- L22:546 weakly excites
- L23:314 weakly excites
- L22:2411 weakly suppresses
- L21:2656 weakly suppresses
- L21:2558 weakly suppresses
- L23:2747 weakly excites
- L20:2442 weakly suppresses
- L21:2077 weakly suppresses
- L24:1197 weakly suppresses
- L22:2389 weakly suppresses
- L22:2077 weakly suppresses
- L24:1028 weakly suppresses
- L20:532 weakly suppresses
- L21:2408 weakly suppresses
- L22:1418 weakly suppresses
- L19:2708 weakly suppresses mid-sentence words in web boilerplate
- L19:2096 weakly suppresses
- L25:1051 weakly suppresses
- L20:2760 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
- emode
- ocale
- filme
- cone
- Ụ
- spécialisé
- ត
- dps
- 번
- мор
Suppresses
- annes
- acular
- igin
- ект
- unn
- 纳斯
- ennis
- 那里
- idor
- nist