sentence openers after a break
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. 10 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 first word of a new sentence or list item, such as 'Okay', the 's' in 'Here's how', or a dash that starts a bullet point. Turning it up did not produce a clear theme in the sample text.
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 220.1<|endoftext|>Re export or Re-exportation is a form of international trade in which a country exports goods previously imported from another country.⏎Okay, does this definition sound way too elaborate? Then for ease of understanding and usage
-
can also be beneficial for children. Here’s how it works: With your child, choose a few categories for saving and budget
-
9,000 people, asked why people with mental health and substance abuse disorders did not seek treatment. About 15% of the respondents<|endoftext|>Okay, I am by no means knowledgable in this field, probably more dangerous
-
this spectacle, drawing on evidence gathered by NASA’s Galaxy Evolution Explorer and the Pan-STARRS1 telescope located in Hawaii. Here’s how NASA describes what you’re seeing in the clip above:⏎Some of the stellar
-
24 times. This is particularly crucial in distillation, where even heating of the mash ensures more precise temperature control, preventing the loss of the final product.⏎Antibacterial Properties of Copper:⏎Copper naturally exhibits antibacterial properties, positively
-
while boosting in bile production.⏎- Scientists believe that curcumin may also be able to prevent liver diseases.⏎Spin
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: why do do we feel so confident that we are not the only ones who can do it? because the power is so high that it is so common that I can't even imagine how I am so happy that I feel the same feeling of
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Space Exploration Technologies (SpaceX)⏎Due to the recent retirement of the space shuttle, NASA⏎now
-
, astronomers say the newly-discovered brown dwarf is one of a pair in a double system 75 light-years from Earth. Astronomers describe the brown dwarfs, identified as CFBDSIR 1458+10
-
the cooler climate and higher pH levels on top of Round Mountain in Saluda provide just the
-
iodiversity and Conservation (2004) 13: 2511. doi:10.1023/B:BIOC.0000048451.07820
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
- L9:1556 suppresses
- L9:1573 excites
- L9:23 suppresses
- L9:134 suppresses
- L9:302 excites
- L9:614 suppresses
- L9:2023 excites
- L6:1400 excites
- L9:234 excites line break inside hard-wrapped text
- L8:2023 weakly excites
- L8:609 weakly suppresses
- L9:866 weakly suppresses
- L9:263 weakly excites
- L6:102 weakly suppresses
- L9:1706 weakly suppresses
- L9:403 weakly suppresses
- L2:444 weakly suppresses
- L9:1489 weakly excites
- L2:443 weakly excites
- L0:42 weakly excites
- L7:838 weakly suppresses
- L5:3023 weakly excites
- L8:2902 weakly excites
- L8:1104 weakly suppresses
Feeds later layers
- L12:2633 strongly excites
- L11:1708 strongly excites
- L18:1518 strongly suppresses
- L15:605 excites
- L13:2506 suppresses
- L15:703 excites
- L18:1357 excites
- L17:247 excites
- L11:2225 suppresses
- L19:1466 suppresses
- L12:702 suppresses end of text token after prose
- L15:2164 excites
- L16:1891 suppresses
- L12:2865 suppresses end of document token after a finished passage
- L12:2666 excites
- L19:1303 suppresses
- L18:2100 suppresses
- L12:1627 suppresses
- L17:244 excites
- L15:1648 weakly excites
- L20:269 weakly excites
- L19:556 weakly suppresses
- L17:61 weakly suppresses
- L27:1377 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
- 第一条
- ographs
- etched
- 她说
- торр
- 条
- legg
- stell
- ograph
- ophile
Suppresses
- Jun
- kü
- clr
- Consider
- Wiki
- .GetName
- {
- Great
- $(
- "${