unclear: weak, scattered activations
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. 2 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 barely fires, and the places it fires on look unrelated: a word inside a hyphenated compound, 'Schedule' in drug and legal texts, and 'Action' in an organisation name. Nothing clear ties these together, and turning it up only made the model repeat words like 'company'.
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.
-
and eventually communism crumbled.⏎Walesa called for that persistence from his listeners. The message was only slightly different from his speech four years earlier at a College-sponsored event attended by, among others, local mayors and elected officials.⏎Things have
-
It grows from egg to adult in three stages, with its appearance changing during the molting process for each stage, noted Amy Korman, an extension educator based in Northampton County.⏎“Unless you are paying daily attention to changes in this insect
-
cell phone.⏎Jen Hensely, who works for the company installing them, explained: “People are on their phones all the time.<|endoftext|>National Conservative Political Action Committee⏎The National Conservative Political Action Committee (NCPAC; pronounced "nick
-
sympathetic dystrophy (RSD), fibromyalgia, and chronic headaches to determine whether ketamine infusions would effectively reduce their pain. Ketamine, a Schedule III drug, has previously been used successfully to treat chronic pain in adults. Schedule
-
, wetland plants and trees that grew near the ancient town.<|endoftext|>For decades, the rightwing gadfly Ed Blum, his group Students for Fair Admissions, and his shadowy rightwing funders have been trying to end affirmative action
-
”⏎When the Public Health (Tobacco Control) Regulations Jamaica 2013 were being drafted, the issue of e-cigarettes was noted. The First Schedule referred to the banning of smoking in public places and carried specific reference to electronic cigarettes
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 use a special kind of kind of kind of kind of kind of kind of kind of kind of kind of kind of kind of kind of kind of kind of kind of kind of kind of kind of kind of kind of kind of kind
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
РАБОТА ПО АНГЛИЙСКОМУ ЯЗЫКУ⏎1. Соотнесите значения слов и выражений, данных в левой колонке, с их английскими эквив
-
the three descriptions.⏎A. Ask a volunteer to leave the room. When s/he returns, s/he will be asked to see if s/he can guess an<|endoftext|>Allergies and
-
or is not clear. Often, they will propose revised sentences, a different subject organization, reordering of paragraphs—all to help the author effectively develop and convey his/her subject. They also challenge the author to make his/her research interesting and relevant.
-
six grade-appropriate math problems: one each of addition, subtraction, multiplication, division, fractions, and percents. I stressed "grade-appropriate" because single digit addition, for example,
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
- L1:1057 suppresses
- L0:2942 weakly excites
- L1:222 weakly excites
- L0:1224 weakly suppresses
- L0:178 weakly suppresses
- L1:1797 weakly excites
- L1:1849 weakly suppresses
- L1:3041 weakly excites
- L0:1083 weakly suppresses
- L1:583 weakly suppresses
- L1:2354 weakly excites
- L1:1229 weakly excites
- L0:1061 weakly excites
- L1:1099 weakly excites
- L0:1634 weakly suppresses
- L1:3009 weakly suppresses
- L1:1464 weakly excites
- L1:1522 weakly suppresses
- L0:1013 weakly suppresses
- L0:2374 weakly excites
- L1:2463 weakly excites
- L0:2766 weakly excites
- L0:1301 weakly suppresses
- L1:536 weakly suppresses
Feeds later layers
- L3:1525 excites
- L9:446 suppresses
- L7:887 excites
- L17:664 excites
- L10:2432 weakly excites
- L9:1462 weakly excites
- L6:97 weakly excites "as well" after "as"
- L9:2913 weakly excites
- L4:267 weakly excites
- L12:2509 weakly suppresses
- L20:1597 weakly excites
- L16:903 weakly excites
- L8:1398 weakly excites
- L8:277 weakly excites
- L3:1262 weakly excites
- L8:464 weakly excites
- L6:464 weakly excites
- L10:1042 weakly excites
- L9:232 weakly suppresses
- L12:2050 weakly excites
- L10:664 weakly excites
- L3:1351 weakly excites
- L7:281 weakly excites
- L21:752 weakly suppresses
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
- pawn
- anto
- avig
- lene
- rance
- """⏎⏎⏎
- velte
- enty
- ocos
- 柰
Suppresses
- 定点
- amd
- Ơ
- 不上
- ETCH
- одном
- Churchill
- zano
- mó
- Champ