unclear: word-initial pieces of longer words
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 mixed: they peak on different tokens. 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 first pieces of longer words that have been split up, such as 'fore' in 'forebears' or 'Fur' in 'Furstenberg'. Turned down, the model got stuck repeating 'perfect' or 'nature'. The pattern is loose and the examples do not share one 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.
-
. Beneath the castle is a network of tunnels, a barracks during the Napoleonic Wars and a military command centre during World War II.<|endoftext|>Our English forebears just partly anglicized a German term back in the sixteenth century
-
, la him/her/you formal||los, las them/you plural|⏎A direct object
-
A Memoir⏎- The Woman I Wanted to Be. Diane Von Furstenberg⏎- Dressed: A Century of Hollywood Costume Design⏎- Secrets
-
(the angle at which a mineral goes black and you rotate the slide) and birefringence (the colour of the mineral using crossed polars).⏎Minerals have different characteristics using the different types of light – plain polarised (P
-
along the Pacific Coast. The climate is mild and the soil is fertile and easily farmed.⏎While medieval European medicine was still mired<|endoftext|>ASIAN WILD HORSE⏎SCIENTIFIC NAME: Equus Ferus⏎HABITAT
-
sacrifice can be a message for each and every one of us.⏎Our Chamber of Commerce estimates the Treaty Oak to be over 600 years old. Most foresters settle on about 300,
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: the fact that a perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect perfect
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
and beyond.”⏎Yersinia pestis is a bacterium responsible for tens of millions of deaths, most occurring in the course of three devastating global pandemics. Catastrophic as this
-
to notice that they are actually different. Mind readers jump to conclusions that are true for them, without checking whether they are true for the other person.⏎5. Catastrophizing: You expect disaster. You notice or hear about a problem and
-
scenes?⏎- What cuts are structural.⏎Example: Indiana Jones And The Last Crusade: the boat race in Venice.⏎After facing rats and burning waters in the catacombs<|endoftext|>Google has once again surprised us with another of their now infamous homepage
-
caused about $84 billion in losses since then.⏎The Boll Weevil Eradication Program—initiated in 1978 by USDA’s Animal and Plant Health Inspection Service with ARS research support and run by eradication foundations
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
- L5:1294 excites
- L2:1518 weakly suppresses
- L18:420 weakly excites
- L18:1023 weakly suppresses the digit 3 inside numbers
- L17:134 weakly suppresses
- L2:1758 weakly excites
- L16:117 weakly suppresses
- L2:1622 weakly excites
- L17:994 weakly suppresses
- L6:4 weakly excites
- L17:378 weakly excites
- L3:1888 weakly suppresses
- L16:14 weakly excites
- L9:35 weakly suppresses
- L17:385 weakly excites
- L4:97 weakly excites
- L4:591 weakly excites
- L20:147 weakly excites
- L16:2808 weakly excites
- L10:237 weakly suppresses
- L1:6 weakly excites
- L22:848 weakly excites
- L14:1721 weakly suppresses
- L1:23 weakly excites
Feeds later layers
- L27:298 weakly excites
- L27:2406 weakly excites
- L27:2448 weakly excites
- L27:1309 weakly excites
- L27:2611 weakly suppresses
- L27:2556 weakly excites
- L27:2039 weakly excites
- L27:7 weakly excites
- L27:758 weakly excites
- L27:1750 weakly suppresses
- L27:2056 weakly excites the 'w' in shipwreck
- L27:1013 weakly suppresses
- L27:2904 weakly excites
- L27:1246 weakly suppresses
- L27:991 weakly excites
- L27:1270 weakly excites
- L27:723 weakly excites
- L27:1792 weakly suppresses
- L27:1726 weakly suppresses
- L27:770 weakly excites
- L27:1848 weakly suppresses
- L27:441 weakly excites
- L27:1505 weakly excites
- L27:1710 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
- /LICENSE
- ้าง
- ื
- entrant
- roach
- uant
- ouns
- \Queue
- �
- 眷
Suppresses
- CppObject
- -parser
- 主义者
- 总资产
- ⁎
- ⇑
- ActivatedRoute
- marvin
- ��
- Gtk