'by' after a title or author credit
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. 6 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 word 'by' when it introduces an author or creator, as in 'articles by', 'poems by', 'translated by' or 'printed from'. Turning it up produced repetitive, unrelated text.
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.
-
Nursing Research, 24(2), 195-216. My approach is based, in part, on the trailblazing article by Bettina Bergmann, The Roman House as Memory Theater.”B. Bergmann
-
Representation considers the ways in which ethnic and cultural communities have been represented in NFB documentaries. In it, you will find film clips, archival materials, original articles by NFB experts, interviews with filmmakers and a classroom activity.<|endoftext|>HYDERAB
-
, detail. Artwork and poems by migrants tell their own stories in expressively rich visual and verbal images.⏎Research reports
-
Eye, one of the most popular viewpoints for tourists.⏎Under the British Empire, which stretched out<|endoftext|>(excerpt from “A Study of Freeway Traffic Operation” by Charles J. Keese, Charles Pinnell
-
from the Seven Seas” by Stan Hugill (1st ed p 219).⏎The lyrics: “Shanties from the Seven Seas” by Stan Hugill (1st ed p 219, 2
-
use the same architecture,<|endoftext|>A Hundred Verses from Old Japan (The Hyakunin-isshu), tr. by William N. Porter, , at sacred-texts.com⏎LONG is the mountain
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: this is a test question. The question is: "I have been thinking about lately." So, the question is: "This is a test question." The question is: "I have been thinking about lately." So, the question is
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
business in the world, and energy is its newsletter. Every Tuesday and Thursday, Energy Source will be sent directly to your inbox, bringing you important news, forward-looking analysis and insider intelligence. Register here.⏎According to data from the U.S.
-
am is ideal. False Dragonhead can tolerate both full sun and partial shade.<|endoftext|>Copyright © University of Cambridge. All rights reserved.⏎'The Solid' printed from http://nrich.maths.org/⏎is one of a series of problems designed
-
um are available?⏎Dried chrysanthemum flowers can be found<|endoftext|>Copyright © University of Cambridge. All rights reserved.⏎'Trick or Treat' printed from http://nrich.maths.org/⏎Emily and Tristan from St James' School
-
:<|endoftext|>Copyright © University of Cambridge. All rights reserved.⏎'Aim High' printed from http://nrich.maths.org/⏎Why do this problem?⏎offers the opportunity
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
- L5:1621 suppresses
- L5:1468 weakly excites
- L5:757 weakly excites
- L4:2310 weakly excites
- L4:565 weakly excites
- L3:630 weakly excites
- L1:2382 weakly excites
- L0:2933 weakly suppresses
- L2:192 weakly excites
- L4:1623 weakly excites
- L5:247 weakly suppresses
- L5:2817 weakly excites
- L1:1269 weakly excites
- L1:1464 weakly excites
- L3:1441 weakly suppresses
- L5:1997 weakly excites
- L4:2781 weakly suppresses
- L4:2917 weakly excites
- L4:1839 weakly excites
- L5:157 weakly excites
- L3:2672 weakly suppresses
- L2:2208 weakly excites
- L1:1335 weakly excites
- L5:771 weakly excites
Feeds later layers
- L9:1627 excites
- L9:1562 excites 'by' before a byline or agent
- L14:1796 excites
- L10:223 excites
- L15:2606 excites
- L7:2571 excites
- L10:629 excites
- L11:2427 excites
- L13:2465 excites
- L14:2508 excites
- L12:1286 excites
- L9:1830 excites
- L15:2889 excites
- L16:2208 excites
- L22:399 excites
- L11:361 excites
- L15:2005 excites
- L21:772 excites
- L7:304 excites
- L18:56 excites
- L10:3018 excites
- L9:1961 excites 'of' in titles and place names
- L10:1054 weakly excites 'from' after movement or transfer verbs
- L19:2006 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
- Amen
- isc
- ôme
- aine
- commissioners
- ACM
- Commissioners
- ån
- 年代
- 层
Suppresses
- 彀
- PAD
- 直
- ��
- emap
- >Data
- otional
- �
- 过得
- nqu