word-start 'l' after a space
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 when a word begins with the letter l after a space, as in 'a lopsided trade' or 'a lurch to the right'. The steering outputs were unrelated to this, so the pattern is only clear from the examples.
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.
-
agitated to keep the United States out of the war against Hitler, the Irish-Americans who bankrolled the IRA’s “armed struggle”, or the Cuban-Americans who lobbied to keep the US shackled to a pointless embargo of the island,
-
routes that account for the bulk of trade between the nations. A lack of quality roads around ports, insufficient warehouses, high tariffs and visa restrictions have contributed to a lopsided and lackluster trade relationship. To help ease trade barriers, negot
-
did not tangibly improve. This and the widespread anti-immigration sentiment stoked, but by no means instigated by Trump, fits this same pattern of a lurch to the right in the public mood.<|endoftext|>Food Safety: Serving⏎You
-
for action.⏎One hundred extras charged up the hill; another hundred came storming down the same hill to do mock battle. In another location Roman centurions lashed and shouted at two hundred slaves who labored to move a huge stone monument
-
the rest is used to<|endoftext|>The turn of the twentieth century is often referred to as the nadir of race relations in the United States. Jim Crow laws, lynchings, and other forms of racial violence plagued the southern region of the country
-
advanced against Egypt by land and by sea. Ramses’ land army checked the enemy’s advance in southernmost Palestine, and the hostile ships were trapped after being lured into the waterways of the delta. Egypt averted conquest by the nor
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: instead of using the abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract abstract
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
(Trachypithecus geei),⏎- Asiatic Water Buffalo (Bubalus bubalis),⏎- Guar (Bos gaurus), Rhino (Rhinoceros unicorns),⏎- Pigmy Hog (Sus salvanius).⏎
-
this method to find out exactly what goes wrong in the nerve cells in movement disorders such as ataxias," said Prof. Dr. Stefan Herlitze (RUB Department for Biology and Biotechnology).⏎The researchers report is in the Journal of
-
’s nocturnes, Bach’s preludes and fugues, and the works of Clementi, Mozart and Beethoven
-
in 1657. In the 18th century the Lorraine family kept collecting and the Grand Duke Peter Leopold of Habsburg-Lorraine founded in Torrigiani Palace a veritable Museum of Physics. The adjoining
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
- L17:134 suppresses
- L5:1294 excites
- L18:420 excites
- L0:662 suppresses
- L21:252 excites
- L18:198 suppresses
- L2:1518 suppresses
- L10:273 suppresses
- L7:2731 excites
- L22:961 suppresses
- L20:257 weakly suppresses
- L2:146 weakly excites unclear: 'lesson' and 'lique' word pieces
- L18:331 weakly suppresses
- L16:83 weakly suppresses
- L5:1712 weakly excites
- L19:2084 weakly excites
- L19:593 weakly excites
- L7:421 weakly suppresses
- L23:1760 weakly suppresses
- L2:1724 weakly suppresses
- L8:599 weakly suppresses
- L20:1230 weakly excites
- L18:365 weakly suppresses
- L21:2568 weakly suppresses
Feeds later layers
- L27:2675 weakly excites
- L27:1583 weakly excites
- L27:1323 weakly suppresses
- L27:2313 weakly excites unclear: 'g' starts of words and hyphenated 'g' pieces
- L27:2069 weakly excites
- L27:1792 weakly excites
- L27:503 weakly suppresses
- L27:2010 weakly excites
- L27:767 weakly excites
- L27:2640 weakly suppresses
- L27:636 weakly excites
- L27:1482 weakly suppresses
- L27:1805 weakly suppresses
- L27:1301 weakly suppresses
- L27:3070 weakly excites
- L27:2763 weakly excites
- L27:2917 weakly suppresses
- L27:1608 weakly excites
- L27:665 weakly suppresses
- L27:2760 weakly excites 'the' after 'between' in age ranges
- L27:2682 weakly suppresses
- L27:1385 weakly excites
- L27:147 weakly excites
- L27:1784 weakly suppresses
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
- yard
- ipo
- ây
- 객
- unittest
- 练
- #import
- 缦
- 简易
- *)[
Suppresses
- บาท
- vaz
- CVE
- 璋
- جماع
- etti
- anged
- crown
- title
- itre