blank slot before a number
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. 4 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 spot right after a label like 'Number of pages:' or 'Word Count:' where a number should come next. It is a 'a figure goes here' signal. Turning it up did not produce a clear theme in the written 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.
-
from which they had been absent<|endoftext|>THE RAATZ ROUNDUP⏎Submitted for use on April 22, 2022⏎Word Count: 432⏎By State Sen. Jeff Raatz (R-Richmond
-
:59 PM MDT · Views: 1,130⏎HUMAN EMBRYONIC STEM CELL RE
-
<|endoftext|>The Art of Story Writing⏎by Nathaniel C. Fowler⏎Publisher: Sulley and Kleinteich 1913⏎Number of pages: 282⏎The writing of stories of every class and of any length,
-
vi, 80 p. :|⏎|Number of Pages||99|⏎|2||WCC exchange -- 2|⏎nodata File Size: 5MB.⏎Yet for many in the Roman Catholic church in Latin America, not
-
anna Climate).⏎The average temperature for the year in Mount Surprise is 73.0°F (22.8°C). The warmest month
-
Smith Colleges 2010⏎Number of pages: 137⏎These notes represent an attempt to develop a new computer graphics course
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: the user is a bit of a problem. The user is a bit of a problem. The user is a bit of a problem. The user is a bit of a problem. The user is a bit of a problem. The user is
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
% of the⏎Denver is the state's largest city and was, in 2002, the 26th largest US city. Its estimated 2002 population was 560,415, but
-
ostellar cloud freely and the temperature remains constant. It is then obtained for Jeans' Mass, MJ . The fact that MJ, the minimum mass able to collapse, increases when the collapse proceeds produces the
-
years when they were under the protection of Lord Thirumal these group of peoples are popularly known by the name of Yadavars. The fact that they are named as Yadavars is due to the reason that Lord Krishna belonged
-
!<|endoftext|>It may still be difficult for some people to accept that an activity such as gambling, which is not a drug, can be addictive. But the idea that addiction is confined to drug addiction rests on an outdated concept of what addiction is.
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
- L2:371 excites
- L3:178 weakly excites
- L2:520 weakly excites
- L3:534 weakly excites
- L1:205 weakly suppresses
- L3:2281 weakly excites
- L2:782 weakly suppresses
- L0:1142 weakly suppresses
- L3:50 weakly excites
- L3:3 weakly suppresses
- L1:394 weakly suppresses
- L2:940 weakly suppresses
- L1:2480 weakly suppresses
- L0:1911 weakly excites
- L2:449 weakly excites
- L2:434 weakly suppresses
- L2:2798 weakly suppresses
- L2:1249 weakly excites
- L1:755 weakly suppresses
- L1:8 weakly excites
- L3:748 weakly suppresses
- L1:1193 weakly excites
- L0:932 weakly excites
- L0:997 weakly suppresses
Feeds later layers
- L10:1829 excites
- L9:291 excites
- L7:806 weakly suppresses
- L5:134 weakly excites
- L5:2027 weakly excites
- L13:1355 weakly suppresses
- L5:1544 weakly excites
- L11:681 weakly excites
- L21:488 weakly suppresses
- L10:205 weakly suppresses
- L5:1479 weakly suppresses
- L7:347 weakly excites unclear: mid-sentence function words
- L11:2321 weakly suppresses
- L15:175 weakly excites
- L16:2741 weakly suppresses
- L10:206 weakly excites colon or equals sign before a definition
- L19:206 weakly excites
- L10:463 weakly excites
- L24:2340 weakly suppresses
- L12:1008 weakly excites
- L9:1907 weakly excites
- L18:759 weakly suppresses
- L17:1244 weakly suppresses
- L11:502 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
- 多样
- azines
- どの
- めて
- どう
- いくら
- 本书
- 是什么呢
- 位
- nowhere
Suppresses
- ')]
- '></
- '>$
- '>";⏎
- '>
- ')</
- amped
- lider
- ző
- '>"