unclear: mixed, repeated phrasing
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. 14 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 many unrelated spots in text, such as a comma in a list, the word 'characterized' after 'disorder', and the start of a repeated phrase. No single pattern stands out, so the meaning is unclear.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
physical ability to perform the movements. (Source: excerpt from NINDS Apraxia Information Page: NINDS)⏎Apraxia is a neurological disorder characterized by loss of the ability to execute... (Source: excerpt from NINDS
-
eye that sees, a hand that obeys, and a soul that feels.” Through art activities, children learn how to “Our schools prepare an eye that sees, a hand that obeys, and a soul that feels
-
of electromagnetic energy. They move through space at the speed of light and can be man-made or occur naturally. They move through space at the speed of light and can be man-made or occur naturally
-
. One out of three people aged 65-74 has some level of hearing loss. After age 75, that ratio goes up to one out of every two people.⏎Researchers don't fully understand why hearing decreases with
-
33.zip⏎It will create a folder as /var/www/html/drupal-7.33, I will rename the folder as /var/www/html/drupal.⏎# mv /var/www/html/drupal-7.33/
-
and use them to print output at certain stages of the program. For example:⏎>>> >>> s1 = 'String in single quotes' >>> s1 'String in single quotes' >>> >>> s2 = "String in double quotes" >>> s
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 number of these things, the number of things, the number of things, and the things, the things, and the things, and the things, the things, and the things, and the things, the things, and the things
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
new sentences. Suddenly, the genetic equivalent of “The rain in Spain stays mainly on the plain” becomes “Mainly the rain in Spain stays mainly on the plain” with the word “mainly” acting as a retrotransposon.
-
this word list.⏎Words formed by adding one letter before or after endoplasm (in bold), or to adelmnops in any order:⏎s - endoplasms⏎All words formed from endoplasm by changing one letter⏎Browse
-
arkshire and King William I granted them Arbuthnott in Mearns.⏎Spelling variations of this family name include: Oliphant, Olifant, Olifard and others.⏎First found in Perthshire (Gaelic
-
difficult because she can't just repeat what the teacher says. She has to swap the words. "I give you the brush" becomes "you give me the brush.") The two main things she hasn't learned at all are tense (was/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
- L13:1208 excites
- L13:1745 excites
- L12:849 excites
- L13:2214 suppresses
- L13:3055 excites
- L11:893 excites
- L13:1749 suppresses
- L13:411 excites
- L13:51 excites
- L12:371 suppresses
- L13:2442 weakly suppresses
- L9:964 weakly suppresses
- L13:1524 weakly suppresses
- L13:2664 weakly excites
- L11:1848 weakly excites
- L9:409 weakly suppresses
- L13:2086 weakly excites
- L13:982 weakly suppresses
- L12:1071 weakly suppresses
- L13:1993 weakly suppresses
- L13:2309 weakly suppresses
- L11:352 weakly suppresses
- L6:1919 weakly excites
- L12:15 weakly suppresses
Feeds later layers
- L17:2570 excites
- L15:1357 excites
- L16:2261 excites
- L16:998 excites
- L16:2075 excites
- L16:500 excites
- L15:2154 suppresses
- L16:2882 suppresses
- L15:1994 excites
- L16:1601 suppresses
- L15:2089 excites
- L17:1976 excites
- L17:2846 excites
- L15:2218 suppresses
- L15:1435 weakly excites
- L17:1421 weakly excites
- L15:1892 weakly excites
- L16:1512 weakly excites
- L15:2966 weakly suppresses
- L16:2806 weakly suppresses
- L16:2717 weakly suppresses
- L23:538 weakly suppresses
- L18:981 weakly suppresses
- L15:1628 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
- zte
- fühl
- instead
- acers
- getTable
- ylan
- coração
- ANGED
- aleb
- qli
Suppresses
- EMBER
- ainer
- ething
- aturity
- dition
- ăng
- طبع
- 植
- 文本
- text