unclear: mid-word pieces like 'sup' and 'mad'
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. 25 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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 small chunks of longer words, such as the 'sup' in 'supplied' or 'supersonic' and the 'Mad' in 'Madinah'. It does not seem to care about the topic. Turning it up made the model produce loose, repetitive text with no clear theme.
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.
-
foremost combat weapons.⏎Acknowledging the significance of the availability of good shields for successful warfare, Carolingian rulers took special care to ensure that their soldiers were well-supplied with ample shields.⏎Shields were used by both the cavalrymen and infantry
-
North and Central America to provide population data for use in science and conservation biology and is the longest-running citizen survey in the world.⏎The Matagorda County Mad Island Marsh Christmas Bird Count (CBC) is one of the most successful in the
-
pools or water parks.⏎- Use of Tampons⏎- Problems emptying the bladder completely⏎- Having a urinary catheter⏎- Kidney stones⏎- Suppressed immune system⏎- Immobility for a long period⏎- Holding your
-
rows;⏎indeed, such sorrows were his bosom companions⏎in that frozen island dungeon⏎where Nithad fettered him:⏎so many strong-but-supple sinew-bands⏎binding the better man.⏎That passed away; this
-
the 14,000-square kilometer Harrat Khaybar, located approximately 137 kilometers to the northeast of the city of Al Madinah (Medina). According to scientists, the volcanic field was formed by
-
be the biggest of all dog breeds around the globe. They’re renowned for their majesty and imposing presence. While they are often gentle in nature and even mild-mannered members of many families, they’re still animals. Just like any dog
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 to use the photos photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo photo
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
and not appearing to be surface feeding at all.’⏎- ‘And wigeon and scaup, teal and shovelers, buffleheads and a few mottled ducks.’⏎- ‘The sea ducks - common golden-eyes and
-
his neighbours demonstrated that diversification of seeds and tubers, along with traditional methods of preparing the soil, enhanced crop and grassland yields. Although the region is economically impoverished, he showed that by reviving the diversity of their natural heritage, rather
-
used mortality, and isolation of small populations are major concerns in the conservation of large carnivores (Clark et al., 1996; Singleton & Lehmkuhl, 2001). The loss of habitat and connectivity between
-
stood out from the majority of artists living there. First, she was a single young female in a society that basically frowned upon such status. Second, she was financially independent, and third, she had received a typical "proper" education as beh
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
- L24:1190 excites
- L21:2519 suppresses
- L22:1138 suppresses
- L23:1046 suppresses
- L21:1679 excites
- L16:1408 weakly suppresses
- L19:1373 weakly suppresses
- L23:206 weakly suppresses
- L19:2007 weakly excites
- L2:630 weakly suppresses
- L18:2383 weakly suppresses
- L12:1214 weakly excites
- L5:2146 weakly suppresses
- L21:1047 weakly excites
- L3:2045 weakly suppresses
- L4:1636 weakly excites
- L20:1319 weakly suppresses
- L9:1713 weakly suppresses
- L14:693 weakly excites
- L8:2055 weakly excites
- L22:151 weakly suppresses
- L0:55 weakly excites
- L18:328 weakly suppresses
- L14:2759 weakly suppresses
Feeds later layers
- L27:873 weakly excites
- L26:2600 weakly excites
- L26:2476 weakly suppresses
- L27:1899 weakly suppresses
- L26:2792 weakly suppresses
- L27:2244 weakly suppresses
- L27:2031 weakly excites
- L27:996 weakly excites
- L27:168 weakly excites
- L27:544 weakly excites
- L27:1878 weakly suppresses
- L27:1830 weakly excites
- L26:558 weakly suppresses
- L27:2796 weakly suppresses
- L26:1897 weakly suppresses
- L26:1938 weakly excites
- L26:2319 weakly excites
- L26:2126 weakly excites
- L27:2546 weakly excites
- L26:2321 weakly excites
- L26:2875 weakly excites
- L27:1206 weakly suppresses
- L26:1626 weakly excites the word 'I' as subject
- L26:1812 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
- <Option
- 日消息
- offs
- undert
- weekends
- eft
- paid
- yme
- ERSION
- _axes
Suppresses
- minster
- documentation
- mine
- edly
- ⎣
- converged
- dataType
- omb
- monton
- fony