spot right 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. 15 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 just before a number is about to appear, such as after 'the rate of' or 'from the age of' when the next word is a figure. Turning it up made the model write number-heavy, repetitive text, so it seems tied to predicting numbers.
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.
-
= 2 x 1/4 or 1/4 is 1/2 of 1/2<|endoftext|>Messier 52 (also known as M 5
-
3.6 years over last 36 years, from the age of 21.4 years in 1970 to the age of 25 in 2006.⏎Researchers T.J. Mathews and
-
9.3% in 1967 to 26.2% in 2014, which was still double the white rate of 12.7%⏎Education. In 1950 only 1
-
fact-family partner. If your student doesn’t recognize the logic of that switch, review how fact families work by using smaller numbers: 5 = 9 – 4 means also that 4 = 9 – 5.⏎Made =
-
women (10.0 per 100,000), with mortality rates twice as high for African Americans as Whites (4.4 versus 2.2 per 100,000).⏎These researchers
-
evergreen broadleaf forest. In this context, we observed the restoration dynamics of plant communities over a period of 15 years (1996 to 2011) with different levels of harvesting intensity, including selective harvesting at
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 to be careful with the user's own and the user's own's's and the user's's's and the user's's's, because the user's's's, and the user's's, because the user's,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
caused<|endoftext|>This preview shows pages 1–2. Sign up to view the full content.⏎This preview has intentionally blurred sections
-
8 Counting and Cardinality Task Cards (CC) 9-13 Numbers and Operations in Base Ten Task Cards (NBT) 14-18 Operations and Algebraic Thinking Task Cards (OA) 19-2
-
no control. A quick check of Jesus’ parables shows the relationships between slaves/servants and wealthy people. (See Matthew 20:1-16, Luke 16:1-13 or Luke 19
-
available⏎Int. J. Environ. Res. Public Health 2012, 9(11), 4170-4186; doi:1
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
- L13:2357 suppresses
- L13:2303 excites
- L13:1372 excites number before an age or week count
- L12:2516 suppresses
- L14:131 excites
- L13:2997 suppresses
- L14:2195 weakly excites
- L0:136 weakly suppresses space after a dash in number ranges
- L14:2078 weakly excites
- L8:7 weakly excites
- L12:698 weakly excites
- L11:2037 weakly suppresses
- L13:1308 weakly excites
- L6:756 weakly suppresses
- L8:16 weakly suppresses
- L14:1081 weakly suppresses
- L5:2642 weakly excites
- L14:926 weakly suppresses
- L9:2326 weakly suppresses
- L12:1269 weakly excites
- L13:573 weakly suppresses
- L6:1754 weakly excites
- L13:587 weakly excites
- L14:2494 weakly suppresses
Feeds later layers
- L16:1107 excites
- L18:987 excites
- L18:1904 excites
- L18:2084 excites
- L16:256 excites number right after 'age of' or 'ages'
- L17:2957 weakly excites
- L16:270 weakly excites
- L16:2048 weakly excites
- L17:712 weakly excites
- L16:2353 weakly suppresses
- L24:1460 weakly excites
- L19:1226 weakly excites
- L24:2931 weakly suppresses
- L20:1170 weakly excites
- L16:2105 weakly excites
- L20:1684 weakly excites
- L25:1830 weakly excites
- L17:577 weakly excites
- L17:676 weakly excites
- L17:2904 weakly suppresses
- L19:293 weakly suppresses
- L18:2152 weakly excites
- L18:629 weakly excites
- L19:2568 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
- uristic
- witnessing
- telesc
- Majesty
- �
- 辅助
- ведение
- thé
- agal
- aur
Suppresses
- ies
- amics
- Bush
- isan
- 骈
- OWN
- langs
- />⏎⏎
- cribes
- Exports