unclear: weak, scattered tokens
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. 13 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 responds only weakly, and to a scattered set of word pieces with no obvious shared meaning. Its strongest examples are hard to tell apart from ordinary text, so it is best treated as a mystery.
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.
-
The answer to soaring gas prices lies in renewable heating, a technology that is ready to replace gas
-
moral principle)⏎3. We ought to do something to help them. (ought)⏎
-
Copyright © 1997, by Random House, Inc., on Infoplease.<|endoftext|>This was
-
Pretty Balloons (counting/number sentences) (Resource Book Only) eBookKindergarten|Grade 1⏎Please Note: This ebook is a digital download, NOT a physical product. After purchase, you
-
Temp. (Celsius)||-65|⏎|Max. Temp. (Celsius)||165|⏎There are no reviews yet.<|endoftext|>Talk about doing something in the blink of an eye.⏎Google wants to take that phrase to a
-
- In search of their lost ones.<|endoftext|>The Verbal Reasoning section of the Graduate
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 have always been a fan of the book, but I have never been a fan of the book, but I have always liked the book, but I have a history of liking the book, but I have a mix of like and dislike
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
aton & Olivier Dupriez, 2011.⏎"Purchasing Power Parity Exchange Rates for the Global Poor,"⏎American Economic Journal: Applied Economics,⏎American Economic Association, vol. 3(2), pages 137
-
spell”November 23rd, 2010 - 6:33 pm ICT by ANI⏎London, Nov 23 (ANI): A study has suggested that Internet chatrooms and social networking sites are to be
-
aceous amberJune 16th, 2010 - 1:53 pm ICT by ANI⏎London, June 16 (ANI): Palaeontologists have discovered two mammal hairs encased in amber that
-
. Dwapara Yug<|endoftext|>Simple technique appears to be safe and effective, review suggests⏎By Robert Preidt⏎MONDAY, Oct. 15 (HealthDay News) -- A technique called the "mother's kiss" is a
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
- L10:333 excites
- L11:171 excites
- L10:124 suppresses
- L10:1578 excites
- L12:1501 suppresses
- L5:780 suppresses
- L12:3045 suppresses
- L9:716 excites
- L10:3055 suppresses
- L10:169 excites
- L12:1979 excites
- L12:1115 excites
- L12:2174 suppresses mixed, web boilerplate phrases
- L10:812 excites
- L9:2521 suppresses
- L5:986 suppresses
- L11:17 suppresses
- L12:2036 excites
- L9:2085 excites
- L10:1526 excites
- L8:957 excites
- L9:2622 excites
- L11:29 excites
- L10:2973 suppresses
Feeds later layers
- L14:1226 strongly excites
- L14:267 strongly excites
- L17:892 strongly excites
- L16:200 suppresses
- L15:2628 excites
- L14:2941 suppresses
- L15:23 suppresses
- L14:2974 suppresses
- L15:1178 excites
- L14:932 suppresses
- L20:543 excites
- L14:84 excites
- L16:1163 excites
- L14:2556 excites
- L16:883 excites
- L16:248 suppresses
- L19:417 excites
- L17:722 excites
- L22:157 excites
- L17:661 excites names in website boilerplate
- L16:1543 excites
- L25:388 excites
- L20:38 suppresses
- L27:581 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
- .should
- irse
- ENTER
- Unidos
- tro
- roph
- '||
- indexed
- Harlem
- paces
Suppresses
- odal
- https
- apter
- ита
- antom
- ạ
- latex
- angle
- airy
- odb