middle digits inside numbers
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 mixed: they peak on different tokens. 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 single digit that sits in the middle of a number, such as the 4 in 1473 or the 3 in 573. It reacts to numbers being split inside, not to whole numbers. Turning it up did not give a clear change in the model's output.
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.
-
description are not made subject or are accorded privileges or advantages<|endoftext|>Tom Standage, digital editor at the Economist, author of A History of the World in 6 Glasses and Writing on the Wall: Social Media - The First 2,0
-
ours — and the galaxies beyond. And we have Copernicus, among others, to thank for that.⏎Copernicus was born in Poland in 1473 and as a young man became fascinated by math, medicine and science,
-
des voyages à Londres, en Angleterre, en Ecosse, au Pays de Galles et en Irlande du Nord. 1400-1600)-language text
-
media⏎University of Missouri Extension⏎Published: Monday, March 18, 2013⏎David H. Trinklein, 573-882-9631⏎COLUMBIA, Mo
-
et d'Irlande. Francis is portrayed in Diane Haeger's novel Courtesan about Diane de Poitiers and Henri II. From 22 June to 7 July 1815, Bonapartists considered Napoleon
-
did the history of yeontan begin? Despite my labors, I could not find a definite<|endoftext|>Paetawa Flax Mill⏎Location: 267 SH1 North of Waikanae⏎Archibald Brown (1
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 am not sure if it's a disease or not, and if it is not a disease, what should I do next? I am very sorry for the loss of my husband, and I am very sorry for my loss of my wife
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
.W.; Thomas, J.A.; Pywell, R.F.. 2012 Identifying time lags in the restoration of grassland butterfly communities: a multi-site assessment. Biological Conservation, 155. 50-
-
Group was established.⏎At this time, Australia had the most gender-segregated workforce in the Organisation for Economic Co-operation and Development (OECD).⏎Affirmative Action (Equal Employment Opportunity for Women) Act passed⏎Following recommendations by the Working Group
-
EDAW. The Act made it unlawful to discriminate against people because of their sex and other factors , but omitted provisions for affirmative action.⏎Instead, Affirmative Action for Women: A Policy Discussion Paper was presented to Parliament, outlining reasons for affirmative
-
TAVERNINI, Daniele NIZZOLI, Giampaolo ROSSETTI, Pierluigi VIAROLI⏎This work is licensed under a Creative Commons Attribution-NonCommercial 4.
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
- L3:285 weakly suppresses
- L2:223 weakly suppresses 'and' or 'or' inside noun lists
- L3:421 weakly suppresses
- L3:247 weakly excites
- L5:2310 weakly excites
- L5:38 weakly suppresses
- L4:1611 weakly suppresses
- L4:339 weakly suppresses
- L0:2132 weakly excites
- L2:2084 weakly excites
- L2:1334 weakly suppresses
- L4:911 weakly suppresses
- L5:1489 weakly excites
- L3:427 weakly suppresses
- L0:2884 weakly excites
- L2:413 weakly suppresses
- L4:2992 weakly excites
- L2:49 weakly suppresses
- L1:1437 weakly excites
- L0:535 weakly excites
- L0:2120 weakly excites
- L4:2586 weakly suppresses
- L3:948 weakly excites
- L3:996 weakly suppresses
Feeds later layers
- L15:2194 weakly suppresses
- L16:2485 weakly suppresses
- L24:2590 weakly excites
- L25:3037 weakly excites
- L8:342 weakly suppresses
- L18:2194 weakly suppresses
- L9:512 weakly suppresses
- L12:380 weakly suppresses
- L14:1721 weakly excites
- L21:1631 weakly excites
- L13:2500 weakly suppresses
- L10:865 weakly excites
- L18:2183 weakly excites
- L10:104 weakly excites
- L15:2107 weakly excites
- L13:2779 weakly suppresses
- L23:405 weakly excites
- L27:598 weakly suppresses
- L15:1285 weakly excites
- L9:1361 weakly excites
- L19:1104 weakly excites
- L17:197 weakly suppresses
- L17:1403 weakly excites
- L17:2141 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
- grilled
- せ
- 岁的
- osition
- lifespan
- ��
- uran
- utable
- ingly
- antwort
Suppresses
- 我自己
- 更好
- acus
- ahoma
- fundament
- 一笔
- anca
- acknow
- .Align
- 一期