digits inside years and decades
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. 19 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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
It lights up on the middle digits of years and decades, like the '0' in 2001 or the '8' in the 1980s. Its strongest negative values come from digits in long ID numbers such as book codes. Turning it up or down gave text that did not relate to dates, so steering does not show its role.
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.
-
faculty member at Carnegie-Mellon University in Pittsburgh, Pennsylvania. He spent fifty years of his life, from 1949 through 2001, as a social scientist and professor of administration, psychology, and computer
-
, yes, US victory in the Cold War all played a part, because the US space program of the 1960s through the 1980s was more about perfecting technologies for defeating the Soviet Union than about pursuing
-
70s bought those semi-enclosed pedestal-style pay phones which, for most of us (Superman included) found lacking.⏎By the 1990s there were nearly three million pay phones in America. But now just a
-
, streams, groundwater).⏎Final Exam (May 17: 10:30-12:30)⏎The final exam will be on materials presented during the final quarter of
-
results⏎Authors: Walker, Robert J. (John), 1801-1869⏎Date: May 27, 1857⏎
-
for parents, knowing<|endoftext|>The N wagons were utilised for construction and maintenance of rail lines from the 1860's to the 1900's.⏎They were replaced by the _QN_ and _NN_ bog
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 be with you in a way that you are with me. But this is not the only way to have a positive and effective work for you, you can have all the other ways of working in a different way. But the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
/27/15⏎Isothermal Community College⏎P.O. Box 804, Spindale NC 28160⏎(828) 395-1307 fax (8
-
||12.0 MB|⏎|Publisher: ||Apple Academic Press|⏎|Date published: || 2016|⏎|ISBN: ||9781771882903 (DRM-PDF)
-
Size: ||4.4 MB|⏎|Publisher: ||CRC Press|⏎|Date published: || 2015|⏎|ISBN: ||9781498738552 (DRM-PDF)
-
Frank”. The book is published by Red Fox, an imprint of Random House. (see www.kidsatrandomhouse.co.uk) ISBN 978-0-0
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
- L11:1629 suppresses
- L17:2360 suppresses
- L17:2884 suppresses
- L17:2721 excites
- L17:926 weakly suppresses
- L14:2673 weakly suppresses
- L1:1245 weakly excites
- L17:1264 weakly excites
- L17:820 weakly suppresses
- L16:151 weakly excites
- L17:3006 weakly excites
- L12:697 weakly suppresses
- L18:900 weakly suppresses
- L8:2329 weakly suppresses
- L10:622 weakly excites
- L13:3039 weakly excites
- L3:2122 weakly suppresses
- L7:1058 weakly suppresses
- L11:1314 weakly suppresses
- L18:2519 weakly suppresses
- L18:2150 weakly excites
- L10:2892 weakly excites
- L6:1149 weakly suppresses
- L1:1446 weakly excites
Feeds later layers
- L20:672 weakly suppresses
- L21:906 weakly suppresses
- L20:1415 weakly suppresses slot before the second number in a range
- L22:823 weakly suppresses
- L20:1884 weakly suppresses
- L20:2482 weakly excites
- L20:1998 weakly excites
- L21:1746 weakly excites
- L24:1666 weakly suppresses
- L21:1198 weakly suppresses
- L22:610 weakly excites
- L22:252 weakly suppresses
- L20:1473 weakly suppresses
- L20:1709 weakly suppresses
- L20:924 weakly excites
- L22:151 weakly suppresses
- L25:512 weakly suppresses
- L21:488 weakly excites
- L24:467 weakly suppresses
- L23:2142 weakly excites
- L21:2784 weakly excites
- L23:1955 weakly excites
- L20:2576 weakly excites
- L23:799 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
- деся
- ply
- 几步
- sign
- committed
- ancer
- ittance
- 簌
- 玙
- ularity
Suppresses
- iminal
- 视听
- reur
- inate
- ligt
- 府
- theories
- imus
- representations
- ல