blank space 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. 26 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 empty gap just before a number, such as the space in "on July 19-20" or "from 1332 BC", where the number has not started yet. Turned down, the model produced only blank lines, so the neuron seems tied to whitespace.
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.
-
urons are ‘brains of the brain’, so to speak, and the rest of the cells were thought to be there only for support. But neurons account for only 15 percent of the brain, while these so-called ‘support’ cells occupy
-
and medical personnel; and U.S. Navy and Coast Guard ships are deployed to support launch contingencies and astronaut recovery. Additionally, the Navy provides a KC-130 tanker for helicopter air refueling, E-2C aircraft for enhanced
-
ges 8-11)⏎The Seneca Falls Convention was a meeting of supporters of women’s rights in the United States. The meeting took place on July 19-20, 1848, at Seneca Falls
-
, death will eventually take us all.⏎However, some people don't look like they're dead at all. Read on!⏎Tutankhamun lived from 1332 BC to 1323 BC. He is widely
-
politics” to the impact of the “digital revolution,” the course looks at the historical background to some of the key phenomena that shaped the controversy-laden campaign of 2
-
2014|⏎|ISBN: ||9781312744967 (DRM-EPUB)|⏎|
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:
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
873) the Sarmatian ideal is contrasted with Western cultural influences. Rzewuski was the author of the novels Cracow Castle (1847–48; Russian translation, 1875
-
The Central Business District as seen from the top of the Carlton Centre⏎|Municipality||City of Johannesburg|⏎|• Total||1.09 km2 (0.42 sq mi)|⏎|• Density||13,0
-
which was purchased by the joined States in 1803. subsequent to several territorial reorganizations, Minnesota in its current form was admitted as the country’s 32nd own up on May 11, 1858
-
10,000 BCE to 300 BCE), and Haniwa figures, which were used as burial objects during the Kofun period (300 CE to 710 CE). Over time, Japanese
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
- L25:146 strongly excites
- L25:854 strongly suppresses
- L20:787 strongly excites
- L21:339 strongly excites spot just before a year
- L25:232 strongly suppresses
- L22:2275 strongly excites
- L25:95 strongly suppresses
- L25:182 strongly excites
- L22:193 strongly suppresses
- L24:2744 strongly suppresses
- L20:433 strongly suppresses
- L19:1174 strongly excites
- L14:140 strongly suppresses
- L24:197 excites
- L21:395 excites
- L23:262 suppresses
- L21:613 suppresses
- L22:743 suppresses
- L24:13 excites
- L23:90 suppresses
- L23:199 suppresses
- L18:246 excites
- L16:396 suppresses
- L19:430 excites
Feeds later layers
- L27:113 strongly suppresses
- L27:287 strongly excites
- L27:2172 strongly suppresses
- L27:306 strongly excites
- L27:163 strongly suppresses
- L27:448 strongly suppresses
- L27:978 excites
- L27:1456 excites
- L27:1117 suppresses
- L27:126 suppresses
- L27:962 excites
- L27:283 suppresses the spot right before a number
- L27:731 suppresses
- L27:33 suppresses
- L27:428 excites
- L27:862 suppresses
- L27:462 weakly suppresses
- L27:902 weakly suppresses
- L27:97 weakly suppresses
- L27:1534 weakly excites
- L27:2089 weakly excites
- L27:1196 weakly suppresses
- L27:170 weakly suppresses unclear: spaces in mixed-language text
- L27:241 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
- ivism
- XX
- →→→→→→
- Î
- dın
- entifier
- バランス
- の
- zero
Suppresses
- ⏎ ⏎
- :\/\/
- ⏎ ⏎
- ⏎ ⏎
- ⏎ ⏎
- -LAST
- //**⏎
- OPTARG
- rending
- 十三五