'of' inside fixed noun phrases
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. 8 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 word 'of' inside set phrases, most often medical symptoms like 'shortness of breath' or 'loss of appetite', and also 'sleight of hand'. Turning it up did not produce a clear change in the model's writing.
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.
-
vessels cause a decrease in the symptoms of congestive heart failure, including shortness of breath, fatigue, and swelling, and can help improve tolerance for exercise.⏎When
-
- Trouble concentrating⏎- Joint pain and stiffness⏎- Muscle cramps⏎- Impaired memory⏎- Can’t find the right words⏎- Shortness of Breath⏎- Metallic taste⏎- Watery eyes⏎- Sensitivity to light
-
way, Gao and his colleagues demonstrated that the tool could reveal signs of pulmonary fibrosis in COVID long-haulers, thus helping to account for shortness of breath, coughing and other lung troubles. A diagnosis, he suggests, that
-
of the most crusading anti-spiritualists of the 1920s. Because of his familiarity with the illusions of stage magic and sleight of hand, Houdini was particularly adept at spotting the trickery that the so
-
dramatic swings . . . the Chesapeake Bay had one of the hottest years,” he added.⏎Mike Armstrong, a fisheries biologist and assistant director of the Massachusetts Division of Marine Fisheries, underscored the significance of the striped bass fishery. “It
-
150 beats/min in patients not taking atrioventricular (AV) nodal blockers. It can lead to symptoms of palpitations, shortness of breath, fatigue, or lightheadedness, as well as an increased
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: "I think, my dad, I would like to have a son who is a teacher in the name of Nishant Kumar" - the first name of the person who is in the state of... So, the first name is a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
glue the printed pages to cardboard (like old cereal boxes) and then cut out the cards and boards.⏎The Rules: To play, you need to print out the deck of picture cards and at least one board for each player. Each player
-
kidney disease. Nevertheless, the rate of kidney trouble was about 20 percent higher among people who had received high-dose NSAIDs than among those who had taken none. The absolute risk of acute kidney injury was 3 in 1,
-
20 temperature sensor that didn’t require a battery or constant source of wired electricity. 3D printing isn't new either, and already, researchers can print out human organs using biological "ink".⏎The UW team are, however, the first
-
since "anno mundi" is Latin, it is technically more correct to write A.M. 1656, just as it is more correct to write A.D. 1991 than to write 1991
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
- L6:2785 excites
- L6:10 excites
- L4:1686 excites
- L6:368 excites
- L6:388 excites
- L4:4 suppresses
- L6:357 suppresses
- L6:1410 weakly excites
- L7:26 weakly suppresses
- L5:248 weakly suppresses
- L5:20 weakly suppresses
- L6:1767 weakly suppresses
- L7:2323 weakly suppresses
- L5:1209 weakly excites
- L5:1745 weakly suppresses
- L6:765 weakly excites
- L6:589 weakly excites
- L2:2950 weakly suppresses
- L6:1751 weakly excites
- L5:1231 weakly suppresses
- L3:315 weakly excites 'of' after a noun phrase
- L4:2138 weakly suppresses
- L7:2611 weakly excites
- L4:1735 weakly suppresses
Feeds later layers
- L9:270 suppresses
- L10:67 excites
- L9:139 suppresses
- L13:2897 suppresses
- L9:152 excites
- L9:29 excites opening parenthesis before an acronym
- L13:1335 excites 'of' inside fixed noun phrases
- L14:1430 weakly suppresses
- L9:459 weakly excites
- L15:395 weakly excites
- L18:77 weakly excites
- L9:742 weakly excites "of" after an agency or department name
- L15:42 weakly excites
- L9:222 weakly excites
- L13:1108 weakly suppresses
- L13:641 weakly excites
- L18:1128 weakly excites
- L17:2905 weakly excites 'from' after 'differ' or 'vary'
- L13:743 weakly suppresses
- L13:210 weakly excites
- L23:146 weakly excites
- L14:605 weakly excites
- L18:58 weakly excites
- L21:1050 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
- ��
- ALAR
- ths
- sf
- ��
- аниц
- cripts
- ully
- ฟา
- faith
Suppresses
- irc
- しながら
- 另有
- .IContainer
- zeichnet
- 的一些
- 不要
- 在线咨询
- -President
- 最多的