unclear: 'Le' and 'Gu' name fragments, 'custom' when pushed up
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 middle of a word that is split in two, such as the 'Le' in 'Leblanc' or the 'Gu' in 'Guiana'. When it was turned up, the model kept writing the word 'custom' over and over.
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.
-
, the first of Europe's Automated Transfer Vehicles, was launched from the ESA spaceport in French Guiana early March 9. It linked up with the space station almost a month
-
, 15th May 1823 as ‘Holbein’⏎Archibald Alexander Leslie Melville, 13th Earl of Leven and 12th Earl of Melville⏎by whom sold, Robinson
-
portrait collection following an international public appeal between 1917 and 1920 which was advertised in both the press and ration books.⏎Diane Lees, director general at the IWM, said: "Everybody in the world
-
. Against his advice, the pair had married before the Express left Edmonton House. He wondered what the Governor would think.⏎Nancy 'Matooskie' Leblanc had her own reasons for reflection on that hasty marriage. It reminded
-
, too, have wildly shifted over the years. Jacques Chirac, at least according to Schofield, was ‘a dyed-in-the-wool Guallist’, and an ideological successor to a young
-
between disparate research communities will create new opportunities to advance conservation,” said co-author Dr Allison Leidner of NASA’s Earth Science Division. “For example, it will help
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: how can a first-year student insaps a first-year insaps insaps insaps insaps insaps insaps insaps insaps insaps insaps insaps insaps insaps insaps insaps insaps insaps insaps ins
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
What type of Treatment is best?⏎Every teen with social issues may have a different combination of needs. We meet our clients where they are, assessing their unique situation and tailoring treatment to be the most efficient and effective while encouraging them throughout.<|endoftext|>
-
content is as effective as possible.⏎Additionally, AI can help you find your target audience. This can allow you to create specifically designed content to meet your target audience’s needs. Finally, AI can help you
-
the barbarous or the savage stage, which displayed a regard for the happiness and well-being of others.⏎It is often difficult to tell how much of Indian manners and morals may have been derived from white people; but there are still some tribes which
-
By knowing what works best for me, I can develop learning strategies for myself that play to my own strengths and preferences.⏎Perhaps I can
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
- L22:1396 excites
- L23:2284 excites
- L23:1933 weakly suppresses
- L24:2029 weakly suppresses
- L17:2691 weakly suppresses
- L25:1400 weakly excites
- L15:2160 weakly excites
- L11:3020 weakly excites
- L21:97 weakly suppresses
- L9:445 weakly suppresses
- L17:2119 weakly suppresses
- L25:943 weakly suppresses
- L22:1645 weakly excites
- L3:2566 weakly suppresses
- L19:2185 weakly excites
- L3:2484 weakly suppresses
- L21:250 weakly suppresses
- L20:2765 weakly suppresses
- L12:2471 weakly suppresses
- L21:1627 weakly suppresses
- L22:678 weakly suppresses
- L9:33 weakly suppresses
- L5:428 weakly excites
- L24:2263 weakly suppresses
Feeds later layers
- L27:794 weakly suppresses
- L27:2024 weakly suppresses
- L27:1863 weakly excites
- L27:2339 weakly excites
- L27:1753 weakly excites
- L27:1868 weakly excites
- L27:1310 weakly excites
- L27:1498 weakly excites
- L27:393 weakly excites
- L27:1915 weakly excites
- L27:1629 weakly suppresses
- L27:2698 weakly suppresses
- L27:1763 weakly excites
- L27:728 weakly suppresses
- L27:814 weakly suppresses
- L27:1712 weakly suppresses
- L27:2241 weakly suppresses
- L27:2798 weakly excites
- L27:2107 weakly excites
- L27:1983 weakly suppresses
- L27:237 weakly suppresses
- L27:2815 weakly suppresses
- L27:1123 weakly excites
- L27:1811 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
- Custom
- custom
- custom
- (custom
- Custom
- customized
- .custom
- _custom
- 定制
- .Custom
Suppresses
- atorium
- tings
- .mainloop
- �
- iture
- etchup
- mitted
- ющ
- rix
- 饔