period after abbreviations and dates
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. 17 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 a full stop that follows an abbreviation or a date, such as "B.C.E.", "U.S.", "A.D." or "Jr.", where the sentence may or may not be over. Turned up, the model got stuck repeating phrases about time and days.
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.
-
.⏎English is Germanic and therefore based on Proto-Indo-European (PIE) language. PIE was spoken around 3500 B.C.E. by a group of nomads called the Aryans living in what is now Southern
-
the seeds of spruce and tamarack, so White-winged Crossbills live and thrive mainly in the conifer forests of the far northern U.S. as well as Canada. They also eat insects, fir seeds, and the occasional
-
by heft. Nonetheless, this particular book is the definitive guide to descents from the Magna Carta Barons of 1215 A.D. for over 200 individuals who emigrated from the British Isles to the
-
14% increase during the 1980s. The population density in 2000 was 41.5 per sq mi. The estimated
-
965, the white high school student traveled to Selma, Alabama, to participate in the march there, led by the Rev. Martin Luther King Jr. He stayed on as a volunteer to help register African-Americans to vote.⏎An invitation
-
War; interrelationships of economic, social, political, and intellectual developments.⏎This class covers the political, economic, social, and cultural history of the U.S. from the end of Reconstruction to the late 20th Century. It will
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 have always been working with the people their family... they are their family... their family... their... their family... their family... their... their... the family... their family... their... their... their...their...their... their
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
thermal Community College⏎P.O. Box 804, Spindale NC 28160⏎(828) 395-1307 fax (828) 286-
-
9, 12). The exam will be comprehensive and will include material from the entire course. To get the most recent copies of the study guides click at left.⏎The field trip to Great Falls (MD) took place on Saturday April
-
is part of the UPMC family.⏎Be safe anytime, anywhere.⏎To find a pediatrician or pediatric specialist, please call 412-692-7337 or search our directory.⏎A resource for our network of
-
Missouri Extension⏎Published: Monday, March 18, 2013⏎David H. Trinklein, 573-882-9631⏎COLUMBIA, Mo. – It
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
- L16:1939 strongly suppresses line break mid-sentence in wrapped text
- L15:286 strongly suppresses
- L14:409 suppresses
- L14:348 excites
- L16:1608 suppresses
- L8:517 excites
- L13:132 suppresses "of" after a noun
- L7:6 excites
- L13:1046 excites 'in' or 'on' ending a phrase
- L13:212 excites
- L15:316 excites
- L14:2677 suppresses
- L6:101 excites
- L10:217 excites
- L16:314 suppresses
- L14:1857 suppresses
- L15:1601 suppresses
- L6:1752 suppresses
- L14:917 suppresses
- L14:453 excites
- L10:272 excites
- L7:8 suppresses
- L12:792 excites
- L14:404 suppresses
Feeds later layers
- L19:1363 excites
- L18:93 suppresses
- L18:90 suppresses
- L19:223 suppresses
- L20:383 excites
- L20:2552 excites
- L18:1270 suppresses
- L19:1122 excites
- L19:219 suppresses
- L19:1935 excites
- L18:186 excites
- L19:413 excites
- L19:257 suppresses
- L19:1506 suppresses
- L18:1731 excites
- L19:1611 excites
- L18:1479 excites
- L18:602 excites
- L18:635 excites
- L21:81 weakly excites
- L20:1639 weakly excites
- L19:1204 weakly excites
- L20:405 weakly excites
- L18:1094 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
- ffset
- .Absolute
- uable
- mbH
- umed
- ategy
- iates
- ировка
- 硌
- ocities
Suppresses
- **
- ⏎
- [assembly
- [
- perfect
- [
- ####⏎
- ...
- [X
- و