'@example' in email addresses
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 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 fires on the '@example' part of the placeholder email addresses that show up in web pages, such as 'firstname.lastname@example.org'. It is a very specific trigger. Turning it up made the model write in vague, wandering sentences, which does not show the email pattern.
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.
-
wide variety of languages, catering for beginning students to background speakers.⏎Our Faculty Members⏎Mr John Norberry: email@example.com⏎Mrs Laura Wade: firstname.lastname@example.org⏎Quite often as Language teachers, we are asked why students should have to
-
However the current status of the project is unclear.⏎Main policy lessons⏎Tulio Monterroso, AMUPROLOAGO: mailto:firstname.lastname@example.org; email@example.com ;⏎Cohen, S. 2
-
<|endoftext|>November 16, 2006⏎- by Robert E. Stevens, GENESIS II (The Second Beginning) E-Mail: firstname.lastname@example.org⏎I am not talking about truth in that 1 + 1 =
-
of Construction, Chaoyang University of Technology, Taichung 41349, Taiwan (corresponding author's e-mail: firstname.lastname@example.org, email@example.com).⏎Chia-Liang Chiang is with
-
would like to contribute something to the GNU project, please consider volunteering for one of these projects. If you are seriously contemplating work, please write to `firstname.lastname@example.org' to coordinate with other
-
- or rule out dead ends. Click on the link above to learn more, sign up for the project by filling out the form below, or contact firstname.lastname@example.org for more information. Updated June 25, 2015
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 way to the house and the way to the house. I think it's something more than one, but it's something less. I think it's a little more than one, but it's a little less. I think it's
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Lumbosacral Region of the SpineSkip to the navigation⏎The spine is
-
00<|endoftext|>The reasons and urgency that lead a group of people to just up and leave a
-
ai awarded him the For Minorities Prize in recognition of the courage shown by Puczi and his comrades.⏎JÓZSEF KOVÁCS HONTALAN (=STATELESS) (1950–2017)
-
9<|endoftext|>Virtual Kid Quest Guide for an Outdoor Art Quest⏎In celebration of Earth Day,
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
- L1:631 excites
- L0:1301 suppresses
- L0:1224 weakly suppresses
- L0:178 weakly suppresses
- L0:516 weakly excites
- L1:179 weakly suppresses
- L1:103 weakly suppresses
- L0:1223 weakly excites
- L0:1259 weakly suppresses
- L1:75 weakly suppresses
- L0:9 weakly suppresses
- L0:2839 weakly excites
- L1:1754 weakly excites
- L0:2110 weakly excites
- L0:17 weakly suppresses
- L0:2768 weakly excites
- L0:1508 weakly excites
- L1:79 weakly excites
- L1:919 weakly suppresses
- L1:1013 weakly excites
- L0:963 weakly suppresses
- L0:3030 weakly excites
- L1:1269 weakly excites
- L0:2986 weakly excites
Feeds later layers
- L13:38 excites
- L15:1096 weakly suppresses
- L21:290 weakly excites
- L20:294 weakly excites
- L11:2578 weakly suppresses
- L10:1985 weakly excites
- L20:3022 weakly suppresses
- L14:3058 weakly excites
- L3:497 weakly excites email addresses at the end of text
- L5:529 weakly excites
- L25:2189 weakly excites
- L4:514 weakly excites
- L5:144 weakly suppresses
- L8:504 weakly excites
- L14:2116 weakly excites
- L18:1635 weakly suppresses
- L15:658 weakly excites
- L8:2420 weakly excites
- L12:2469 weakly excites
- L8:274 weakly excites
- L8:891 weakly excites
- L8:2566 weakly suppresses
- L16:1662 weakly excites
- L22:199 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
- enne
- iotics
- arded
- {}.
- steen
- iotic
- >\⏎
- 尹
- uchen
- ARY
Suppresses
- (Optional
- .Clock
- preorder
- invert
- delim
- Links
- (optional
- 陬
- dvd
- OPTIONAL