periods inside names and initials
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. 10 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 dot in a person's middle initial, like the period in 'Gerald R. Ford' or 'Harold G. Moore'. When it was turned up, the writing drifted into rambling, unrelated text about topics like houses and family.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
Gerald R. Ford.⏎Transcript from the February 10, 1995 online interview with former President Gerald R. Ford.⏎Grades: PreK–K, 1–2,
-
“We Were Soldiers Once . . . And Young,” by Lt. Gen. Harold G. Moore and war correspondent Joseph Galloway. The 2002 movie
-
(828) 395-1307 fax (828) 286-8208⏎Charles P. Wiggins, Director of Library Services⏎Documentation Styles for Citing Resources⏎(
-
also ill-equipped and trained, had nearly 40,000.⏎Price left a small garrison at Lexington, which was overrun by Major Frank J. White’s First Squadron Prairie Scouts on the early morning of October 1
-
was founded in 1975 by conservative activists John Terry Dolan, Charles Black and Roger Stone, with help from Richard Viguerie and Thomas F. Ellis. The group got its
-
7:30 a.m. and again at 2:00 p.m. in rooms 393-396 of the Ernest N. Morial Convention Center. The session is designed for all hematologists who treat patients
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 can't wait for you to play the game of the same color." What is the probability of getting a certain number of points? The probability of a win or loss game, the same color. If I choose a number of points
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
San Francisco Bay Area. Greater Eureka, one of California's major commercial fishing ports, is the location of the largest deep water port between San Francisco and Coos Bay, a stretch of about . The headquarters of both the Six Rivers National Forest
-
and pdf⏎- Tuesday, June 1, 2021 2:11:00 AM⏎- 1 comment⏎File Name: the death and life of the great lakes .zip⏎- Discussion questions for ‘
-
.codeguru.com/algorithms/factorial.shtml I don't happen to think that the code there
-
.⏎Word Origin Middle English (in legal<|endoftext|>Keywords: Sargent Mountain⏎Historical Items Showing 1 of 1 View All⏎Contributed by: Mount Desert Island Historical Society⏎Location: Northeast Harbor⏎Images taken by itinerant photographers
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
- L9:1618 excites
- L8:2315 excites
- L9:1599 suppresses
- L6:1447 weakly suppresses
- L9:1689 weakly suppresses
- L8:878 weakly excites
- L9:1625 weakly excites
- L2:2328 weakly suppresses
- L3:1641 weakly excites
- L4:3002 weakly excites
- L9:1522 weakly excites
- L8:1675 weakly suppresses
- L0:1329 weakly suppresses
- L0:1224 weakly suppresses
- L2:2259 weakly excites
- L0:178 weakly suppresses
- L3:2496 weakly excites
- L5:2274 weakly suppresses
- L1:2673 weakly suppresses
- L3:2589 weakly excites
- L8:493 weakly suppresses
- L9:832 weakly excites
- L8:1063 weakly excites the word "genus" in taxonomy
- L3:785 weakly suppresses the word 'other' after 'each'
Feeds later layers
- L11:1780 excites
- L12:2254 excites
- L11:1672 weakly excites
- L12:1976 weakly excites
- L15:1629 weakly suppresses
- L21:1607 weakly suppresses
- L21:146 weakly excites
- L17:1702 weakly excites
- L23:3046 weakly excites
- L26:2933 weakly suppresses
- L18:999 weakly excites
- L15:654 weakly suppresses
- L11:2450 weakly excites
- L27:3007 weakly excites
- L24:168 weakly suppresses
- L21:283 weakly suppresses
- L15:1435 weakly suppresses
- L11:704 weakly excites
- L23:3054 weakly excites
- L13:3054 weakly excites
- L12:2664 weakly suppresses
- L13:2873 weakly suppresses
- L26:42 weakly suppresses
- L15:1739 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
- ↑
- 眼光
- ien
- dir
- uild
- 四处
- tract
- eur
- 眼
- 纲
Suppresses
- 合一
- ystery
- lycer
- *)⏎
- setResult
- .scalajs
- schließen
- CallCheck
- "⏎⏎⏎
- 县政府