line break at the end of a line
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. 24 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 responds to the point where a line of text ends and a new line begins. When it is turned up, the model starts writing nothing but blank line breaks, so it is tied to newlines.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
, delivering a smooth user experience. With the ever-growing demand for innovative apps, proficient coding skills have become invaluable for developers looking to make a mark in the industry.⏎The Challenges of Coding for Apps⏎Coding for apps presents its fair share of challenges
-
screen time ) 😌⏎- Go to bed and wake up at the same time each day. Our physical
-
Symmetric encryption AES/block ciphers Block modes (ECB, CTR, OFB) Asymmetric encryption RSA in Java Comparison of algorithms Key sizes Hash functions⏎Secure hash functions in Java⏎As you probably know if you've worked on certain
-
package can be installed like so:⏎cabal install bool-extras⏎If you don’t have cabal-install, you can manually download and install bool
-
⏎- Configure Spark to modify the Spark properties, environmental variables, or logging properties⏎- Monitor Spark using the web UIs, metrics, and external instrumentation⏎-
-
⏎- Subject: Drama - World⏎- Type: Essay⏎- Paper: #87743533⏎Excerpt from Essay :⏎Perceptions of Che Guevera⏎PERCEPTION
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:fullyfullyfullyfullyfullyfullyfullyfullyfullyfullyfullyfullyfullyfullyfullyfullyfullyfullyfullyfullyfullyfullyfullyfully
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
The investigation was carried out using AWS 5.22–95 filler wire of size 1.2 mm diameter and the base metal used was IS:2062 structural steel plate of 20 mm thickness. An
-
done.⏎Answer: True. There are hundreds of thousands of dust mites in our homes, from
-
70%.⏎Answer: B. Humidity levels should be kept below 60 percent, ideally between
-
bad mood.⏎- He’s very cross today, so don’t bother him.⏎synonyms: cranky, disagreeable, grouchy⏎Word Builder: cross +⏎crosswalk: a part of the road that is marked off for people to cross
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
- L21:173 strongly excites
- L15:123 suppresses
- L18:48 suppresses
- L20:8 suppresses
- L14:51 excites colon that introduces a statement
- L17:38 suppresses
- L19:187 suppresses
- L13:10 excites
- L9:7 excites
- L10:84 suppresses
- L17:829 suppresses
- L15:44 suppresses
- L16:170 suppresses
- L20:863 excites
- L18:5 suppresses
- L12:5 excites
- L10:33 suppresses
- L14:56 suppresses document ends and list line breaks
- L22:262 suppresses comma inside big numbers
- L17:62 suppresses
- L15:81 suppresses line breaks inside running text
- L23:35 excites
- L14:348 suppresses
- L1:1 excites
Feeds later layers
- L27:487 strongly excites
- L26:42 strongly suppresses
- L26:241 strongly excites
- L25:40 strongly suppresses
- L25:91 strongly excites
- L27:1801 strongly excites
- L25:248 strongly suppresses
- L26:230 excites
- L25:297 excites
- L25:533 excites first letter m or oct fragments
- L27:143 suppresses
- L27:598 suppresses
- L26:753 suppresses
- L27:1496 excites
- L26:1018 excites
- L26:156 weakly suppresses
- L25:144 weakly excites
- L27:116 weakly suppresses
- L25:71 weakly suppresses
- L26:167 weakly excites
- L26:685 weakly suppresses
- L27:200 weakly suppresses
- L27:1687 weakly excites
- L27:433 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
- ⏎
- ⏎⏎
- ⏎
- ⏎ ⏎
- ⏎ ⏎
- ⏎⏎⏎
- ⏎ ⏎
- ⏎ ⏎
- ⏎⏎⏎⏎⏎⏎
Suppresses
- ly
- ful
- LY
- ся
- gether
- shire
- theless
- alyze
- hip
- ">'+⏎